Technical Report by CelsiusPro AG, March 2021
Technical Report – Component 2
1265762 – Feasibility Study of Excess Rainfall and Drought
Insurance for Pacific Island States
World Bank Disaster Risk Finance and Insurance Program
March 2021
in partnership with
1
Technical Report by CelsiusPro AG, March 2021
Contents
Executive Summary .......................................................................................................................... 3
1. Introduction ............................................................................................................................ 5
2. Background & Recap of Component 1 .................................................................................... 6
2.1 Project Stakeholders: WB and PCRIC ................................................................................... 6
2.2 The Pacific Island Countries ................................................................................................. 6
2.3 Component 1 Findings ....................................................................................................... 10
3. Deliverables – Excess Rainfall ................................................................................................ 11
3.1 Methodology and Insurance Product Options................................................................... 11
3.2 GPM IMERG data expansion using ERA reanalysis data .................................................... 20
3.3 Product Analysis – RMI ...................................................................................................... 24
3.4 Product Analysis – Solomon Islands................................................................................... 28
3.5 Product Analysis – Samoa ..................................................................................................32
3.6 Product Analysis – Fiji ........................................................................................................ 36
3.7 Product Analysis – Tonga ...................................................................................................39
3.8 Product Analysis – Cook Islands ......................................................................................... 43
3.9 Product Analysis – Vanuatu ............................................................................................... 46
3.10 Overall Feasibility Assessment ........................................................................................... 50
3.11 Next Steps Towards Product Development ....................................................................... 52
4. Deliverables – Drought .......................................................................................................... 54
4.1 Methodology and Insurance Product Options................................................................... 54
4.2 Drought Structures used in the Analysis............................................................................ 56
4.3 Specific Approach to Drought for RMI, Tonga and Samoa. ............................................... 56
4.4 Product Analysis – RMI South ............................................................................................ 58
4.5 Product Analysis - RMI North .............................................................................................64
4.6 Product Analysis - Samoa ................................................................................................... 69
4.7 Product Analysis - Tonga ....................................................................................................74
4.8 Overall Feasibility Assessment ........................................................................................... 79
4.9 Next Steps Towards Product Development ....................................................................... 80
Annex ............................................................................................................................................. 81
Annex 1 - Data Expansion Method Report ................................................................................ 81
2
Technical Report by CelsiusPro AG, March 2021
Executive Summary
The Technical Report for Component 2 of the “Feasibility Study of Excess Rainfall and Drought
Insurance for Pacific Island States” project assesses several weather index product options for
seven Pacific Island Countries (PIC’s). The project's overall objective is to provide an assessment
on the technical feasibility of structuring parametric insurance solutions for flood and drought
events for the PIC’s participating in the Pacific Catastrophe Risk Assessment and Financing
Initiative (PCRAFI). This report focuses on determining a suitable product structure, based on
findings from Component 1 of the project, on addressing excess rainfall and drought risks
effectively and cost-efficiently.
To refine the feasibility study’s scope, Component 2 is targeted at seven countries within
PCRAFI. The primary focus group (Group A) includes the Marshall Islands, the Solomon Islands,
Fiji, Samoa, Tonga, the Cook Islands, and Vanuatu, based on an initial evaluation of the
governments’ past engagement with PCRIC as well as the prevalence of severe flood and
drought events affecting each country. The feasibility of introducing an excess rainfall insurance
product is conducted for all seven countries, while the drought index insurance options focus on
the Marshall Islands, Samoa, and Tonga.
A sub-national product structure is evaluated, allowing for localized excess rainfall events
within a PIC’s region to be detected. For most of the Group A countries, the sub-national
administrative boundaries are used, with the product’s payout triggers being based on regional
precipitation return periods (RP). The overall coverage limit is divided between the country’s
regions based on their population. Payouts are then aggregated on a national level and provide
the government with access to immediate liquidity in the aftermath of a rainfall-induced flood.
The drought product options are assessed on a national level, except the Marshall Islands, which
are split into the two climatologically distinct regions, North and South.
Building on the findings from Component 1, the project team appraises the use of GPM IMERG
as well as an extended dataset with bias-corrected ERA-5 for the excess rainfall product. A
rainfall aggregation method using eight consecutive 3-hourly periods to get a rolling 24-hour
accumulation is applied. The unweighted mean value of each PIC's region's GPM IMERG grid
cells is then used as the basis for the product structure. To review precipitation peaks prior to
GPM IMERG’s commencement in mid-2000 and increase the robustness of RPs exceeding 80
years, an extended dataset using ERA-5 reanalysis data is conducted. Following an extensive
review of the resulting product pricing using both the extended dataset and RPs compared to
GPM IMERG only, the decision is taken to focus the latter's product analysis. This approach also
aligns the feasibility study with practices in the (re)insurance market.
While suitable excess rainfall product options for most Group A countries are identified, the
alignment to the PICs flood risks remains largely uncertain. The product analysis per country
confirms that a regional product structure can pick up localized and dispersed excess rainfall
events that would fall below the thresholds of product focused on the national level. Therefore,
the project team recommends proceeding with regional product options per country that yield
annual payout frequencies within the vicinity of a 5-year RP and affordable indicative premium
rates between 8-12% per year. A caveat on the basis of risk assessment remains the scant
availability of historical loss event records for rainfall-induced floods. To accurately assess and
align the weather index to the PIC’s needs, a closer review of flood damages should be
conducted in most Group A countries prior to approaching (re)insurance markets.
3
Technical Report by CelsiusPro AG, March 2021
Different drought product options are found to trigger payouts in line with historical events
but require further consultations with the PICs to determine the alignment with their liquidity
needs as droughts emerge. Selected drought products, based on a lack of cumulative rainfall
and combining different coverages across different risk periods, are found to capture historical
drought events and deemed feasible risk transfer solutions. However, a better understanding of
local seasonality and each country’s drought risk management is required to fine-tune the most
suitable product structures prior to market consultations.
4
Technical Report by CelsiusPro AG, March 2021
1. Introduction
1. The CelsiusPro (CP) and Risk Frontiers’ (RF) joint project team combines a strong understanding
of catastrophe risk and designing financial products. The project team is supporting the World
Bank and the Pacific Catastrophe Risk Insurance Company’s (PCRIC) efforts to expand the range
of parametric insurance products offered to the Pacific Island Countries (PICs) as part of the
Pacific Catastrophe Risk Assessment and Financing Initiative (PCRAFI). The project, “Feasibility
Study of Excess Rainfall and Drought Insurance for Pacific Island States”, comprises two distinct
but interconnected components:
Component 1: Analysis of Excess Rainfall and Drought Hazards
Component 2: Feasibility of Parametric Insurance Product.
2. This report builds on the work conducted for and the findings derived from the “Technical
Report – Component 1”, submitted to the World Bank in July 2020. For Component 1, CP and RF
produced an analytical hazard assessment for excess rainfall and drought, assessing the
feasibility of designing a parametric insurance product with a low basis risk. The project team
found that introducing parametric insurance solutions with the current precipitation data
provided by GPM IMERG is feasible. Although the limited calibration for certain countries, due
to a lack of historical loss events, restricts the ability to structure highly customized products in
these cases.
3. The objective of Component 2 is to apply the previously conducted hazard assessment to
structure suitable sovereign level insurance products for a group of selected PICs. The excess
rainfall and drought coverages aim to provide rapid payouts to support the governments’
emergency response efforts. As part of this report, the project team provides the World Bank
and PCRIC with several parametric product options and the tools required to compare them.
Finally, an expert judgment on each product’s basis risk is articulated, followed by an evaluation
of the feasibility of including the products within PCRIC’s insurance offering.
4. The work on Component 2 commences with the project team’s support of the World Bank and
PCRIC’s engagement with the PICs. Clarifying the governments’ risk transfer priorities lays the
foundation for the product structure and subsequent options. To achieve additional insights in
historical excess rainfall events, following a request from the World Bank and PCRIC, the project
team combines GPM IMERG data reaching back to mid-2000 with bias-corrected ERA-5 data
reaching back to 1979. This approach allows for the number of reported years to be doubled
and to subsequently more robustly extrapolate the return period levels to approximately 150
years.
5. The excess rainfall pricing tool outlines different product options, visualizes the precipitation
levels between 1979 – 2020, and highlights the “as-if” performance of the product during past
excess rainfall events. An indicative premium is calculated by applying a burn cost pricing
methodology on the GPM IMERG data between 2001-2020. The subsequent commentary on
the product's basis risk per PIC is based on a comparison with the historical loss event catalog
compiled during Component 1. The report concludes with detailed comments on the feasibility
of introducing excess rainfall and drought risk transfer solutions for the PICs, including an
overview of the next steps required to move forward with the development of the products.
5
Technical Report by CelsiusPro AG, March 2021
2. Background & Recap of Component 1
2.1 Project Stakeholders: WB and PCRIC
6. Between the completion of Component 1 in July 2020 and initiating Component 2, a number of
interactions with the PIC governments needed to be held. Within this feasibility study, PCRIC
acts as the interface between the project team and the governments. In close collaboration
with the World Bank, which manages the overall feasibility study and the wider objective of the
project, PCRIC examines the willingness among PCRAFI member states to purchase a parametric
excess rainfall and/or drought insurance product. Furthermore, PCRIC and the WB will
potentially initiate a dialog to evaluate potential donor support for premium subsidies.
7. Prior to the kickoff call for Component 2 on October 8, 2020, PCRIC shared the initial feedback
received from Group A countries providing a first indication of their product preferences (see
Section 2.2). The call included CP, RF, the World Bank as well as PCRIC and laid out the approach
towards defining the product options and an outlook of the steps required to bring the product
to market in the future. A clear understanding of the product’s priorities was gained during the
kickoff call and a decision to structure the product along the lines of sub-national regions was
taken.
8. On November 10, 2020, an interim call was held, providing CP and RF with the opportunity to
update the WB and PCRIC on the progress achieved to date. A first excess rainfall product,
including an overview of the related excel based pricing tool, was presented to the participants.
In addition, the project team’s findings from a drought analysis were shared and a number of
open questions were discussed. It was agreed to using the unweighted, mean precipitation data
(metric 1) for the product structuring and to weight the coverage/sum insured distribution
across regions based on the population.
9. Following the information received during the interim call, the World Bank and PCRIC requested
further analysis on the possibility of expanding GPM IMERG’s 20-year precipitation time series
with bias-corrected ERA-5 data, which reaches back to 1979. In December 2020, the World Bank
mandated the project team with the data extension work. An addition to the existing contract
between the World Bank and CP was finalized thereafter.
2.2 The Pacific Island Countries
10. Climatology of the PICs
The region encompassing the 14 PCRAFI members covers a vast area of the Western Equatorial
Pacific Ocean and, as such, leads to a predominantly maritime climate influenced by ocean-
atmosphere interactions. While most of the land area is small, these Pacific Island States'
territorial waters traverse the tropical zones in both hemispheres, extending from
approximately 20 degrees latitude North to 25 degrees latitude South (about 5,000 km).
Similarly, the east-west extent is great, extending from approximately 130 degrees longitude
East to 140 degrees longitude West (about 9,000 km). The climatology of individual PICs is
included in the Technical Report – Component 1, Section 3.1.
6
Technical Report by CelsiusPro AG, March 2021
11. Sub-annual variability
During the Austral summer period (November to April), the Inter-Tropical Convergence Zone
(ITCZ) shifts southward, bringing the tropical monsoonal trough into the Southern Hemisphere.
Convective low-pressure systems are often embedded in the trough, which is the principal
cause of excess rainfall and flooding in this region. In the Austral winter, the trough shifts
northward, bringing drier and more stable weather to the southern parts of this region. The
seasonality of this zonal (north-south) shift is strongest in the southern and northern parts of
this region; seasonality is reduced towards the equator.
12. Year-on-year variability
The major cause of year-on-year climate variability in the region is El Niño–Southern Oscillation
(ENSO). ENSO can be seen as a continuum whereby ‘ENSO neutral’ represent normal or average
state conditions of the Equatorial Pacific Ocean, with the two extremes being La Niña and El
Niño. El Niño is a reversal or break-down in the Equatorial Pacific circulation and usually brings
drier and more stable conditions to the Southwest and Western Equatorial Pacific region. La
Niña, however, represents a ‘spin up’ of the tropical Pacific circulation and is associated with
wetter, stormier conditions on the western side of the Equatorial Pacific. While heavy rain and
flooding can occur during all states of ENSO, periods of heavy rain and cyclogenesis are more
likely to occur in the Southwest Pacific during La Niña phases. Heightened cyclone activity in the
North Pacific can also occur during periods of El Niño (e.g. Marshall Islands). ENSO typically
operates on timescales of two to seven years. The significant level of climate variability
impacting this area is an important consideration when deriving return periods of extreme
rainfall, as it introduces non-stationarity to the time series. This means that any single year may
be significantly under- or over-risk depending on the phase and strength of climate oscillations
(see Paragraph 53 for more details). It is a risk taker's decision to to adjust the pricing annually
depending on the actual risk situation.
13. Grouping of PCRAFI member countries
To target the feasibility study's scope and increase the efficiency of the project, a division of the
14 PCRAFI member countries was conducted during Component 1. Table 1 provides an overview
of the segregation into three groups. The initial focus during the hazard assessment and product
design phase of the project lies squarely with the seven Group A countries. The selection is
based on an initial evaluation of the governments’ past engagement with PCRIC on the tropical
cyclone (TC) and earthquake/tsunami insurance, a basic assessment of their focus on fiscal
resilience to natural disasters as well as the prevalence of severe flood and drought events
affecting each country. The numerical position or group allocation is not intended as an
assessment of the respective country’s relevance or importance within PCRAFI. The allocation is
solely used within the context of the feasibility study to support a structured analysis across the
vast Pacific region.
7
Technical Report by CelsiusPro AG, March 2021
Table 1: PCRAFI Member Countries are split into three groups.
Group A - Primary Focus Group B – Countries with Group C – Other PCRAFI
Group Potential Members
Marshall Islands - RMI (1)* Tuvalu (8) Micronesia (11)
Solomon Islands (2) Palau (9) Kiribati (12)
Fiji (3) Niue (10) Nauru (13)
Samoa (4)* Papua New Guinea (14)
Tonga (5)* Micronesia (11)
Cook Islands (6)
Vanuatu (7)
*PICs in focus for the drought insurance product
Figure 1: Map of the 14 PCRAFI Members
14. PCRIC’s dialog with the Group A governments confirms a strong interest in an excess rainfall
product. Precipitation induced flooding in the region is considered one of the primary disaster
risk management concerns among officials, with all seven countries recognizing the need to
increase their financial resilience to such events. Having access to immediate liquidity in the
aftermath of the severe event would support the governments’ emergency response efforts.
Among Group A countries, Samoa (4) and Tonga (5) show robust interest for an excess rainfall
product and highlighted their interest in drought coverage. The Republic of Marshall Islands
(RMI) raise a stronger preference towards a drought insurance product, driven by the limited
freshwater supply across the dispersed atolls comprising the country. Therefore, the project
team’s efforts on structuring a drought product is focused on RMI, Samoa, and Tonga.
8
Technical Report by CelsiusPro AG, March 2021
15. To provide the product structure with additional flexibility, both the excess rainfall and drought
products are structured along the lines of sub-national administrative areas, splitting the seven
Group A countries into 42 regions. Next to noting each Group A country’s various regions, Table
2 also includes an overview of the number of grid cells and population size. The weighting based
on population is proposed as the standard distribution of the coverage/sum insured across
different PICs, see Section 3.1. As described in Section 2.3, the analysis conducted as part of the
Technical Report – Component 1 requires climatic zones for RMI and Cook Islands, which differ
from the administrative boundaries of the countries’ regions, to be used for the product
structure.
16. In addition to the regional split per country, national-level data is provided for comparative
purposes within the excess rainfall pricing tools, as noted in Section 3.1. This time series
combines the IMERG cells per PIC depicted in Table 2.
Table 2: Overview of regions per Group A country
Country Region Population Weighting Number Weighting per
GPW* based on of IMERG number of grid
population cells cells
Marshall South 51’768 98% 192 81%
Islands
North 1’092 2% 44 19%
Solomon Guadalcanal 226’682 30% 69 12%
Islands
Malaita 147’987 20% 93 16%
Honiara 111’831 15% 2 1%
Western 91’570 12% 140 23%
Makira Ulawa 51’252 7% 53 9%
Choiseul 33’849 5% 56 9%
Isabel 32’502 4% 77 13%
Central 30’361 3% 32 5%
Temotu 23’137 3% 49 8%
Rennell and B. 3’777 1% 24 4%
Fiji Central 415’001 44% 69 15%
Western 357’389 38% 116 26%
Northern 132’915 14% 136 31%
Eastern 35’988 3% 118 27%
Rotuma 1’319 1% 5 1%
Samoa Tuamasaga 128’847 38% 13 15%
A'ana 40’239 12% 5 6%
Atua 39’478 12% 11 13%
Palauli 28’291 8% 12 14%
Fa'asaleleaga 20’843 6% 7 8%
9
Technical Report by CelsiusPro AG, March 2021
Satupa'itea 17’843 5% 7 8%
Gaga'ifomauga 15’834 4% 9 10%
Va'a-o-Fonoti 14’278 4% 4 5%
Gaga'emauga 13’993 4% 7 8%
Vaisigano 12’477 4% 10 11%
Aiga-i-le-Tai 9’092 3% 2 2%
Tonga Tongatapu 85’165 77% 15 20%
Vava'u 14’485 13% 18 24%
Ha'apai 5’399 5% 37 48%
Eau 4’879 4% 3 4%
Niuas 833 1% 3 4%
Cook Islands Rarotonga 14’153 66% 4 11%
South 6’482 30% 13 35%
North 813 4% 20 54%
Vanuatu Shefa 114’363 39% 42 16%
Sanma 57’727 19% 61 23%
Malampa 40’432 14% 53 20%
Tafea 35’652 12% 39 15%
Penama 34’981 12% 31 12%
Torba 11’128 4% 37 14%
* Source: Gridded Population of the World, version 4, 2020 (~1-km resolution)
2.3 Component 1 Findings
17. To gain an understanding of the overall “Feasibility Study of Excess Rainfall and Drought
Insurance for Pacific Island States” project, this report should be read alongside the Technical
Report – Component 1, submitted in July 2020. While both components are distinct, this report
strongly relies on insights derived during Component 1. The most important findings of the
Technical Report – Component 1 can be summarized as follows.
18. NASA GPM IMERG precipitation data found to be of sufficient spatial granularity to observe
localized excess rainfall events. The use of the largest rolling 24-hour daily value, the aggregate
of eight consecutive 3-hourly periods, captures heavy short-duration precipitation more
precisely than using standard daily data aggregated between 00:01 and 23:59. Therefore, this
precipitation aggregation acts as the basis for the excess rainfall product, despite the deviation
from existing market practices in the (re)insurance market, which rely on standard daily
precipiation data. Due to the prolonged duration of droughts, the standard daily 24-hour GPM
IMERG data provides sufficient granularity to detect dry spells and act as the basis of the
drought insurance coverage.
10
Technical Report by CelsiusPro AG, March 2021
19. Comparing historical loss events and rainfall data observations for Group A countries hint at a
relatively close alignment. Excess rainfall volumes gathered by satellites are relatively well
matched with recorded loss events, especially in Solomon Islands, Fiji, and Samoa. This finding
holds particularly true for flood events that are not associated with TCs. For Marshall Islands,
Cook Islands, Tonga, loss events are strongly focused on past TCs, which are not always
detected as extreme events in the precipitation data. For Vanuatu, the historical flood events
are not captured well within the GPM IMERG data. The limited accuracy of publicly available
historical loss events, particularly pronounced for droughts, and the infrequent inclusion of
relevant economic damage data hamper these records' reliability.
20. As part of the Extreme Value Analysis, the conducted homogeneity assessment ensures that
each PIC can be regarded as a single climatological unit. This holds true for 10 out of the 14
countries reviewed, with RMI, the Solomon Islands, the Cook Islands, and Kiribati requiring
further sub-divisions given their geographical extents. From a meteorological perspective, these
four countries are subject to different excess rainfall risks across their territory.
21. The hazard assessment concludes that introducing a parametric insurance product targeting
medium to severe rainfall events is technically feasible. However, due to limited historical loss
event data in some countries, the product design and trigger definition will need to rely on
statistics and comparable Pacific Island Countries. This challenge is particularly pronounced
when attempting to correlate droughts observed in the data with historical events.
3. Deliverables – Excess Rainfall
3.1 Methodology and Insurance Product Options
22. Excess Rainfall Product Structure
The project team recommends proceeding with the product structure outlined in Table 3. The
product pays a fixed amount, also known as the tick, for every mm of excess rainfall above a
specified trigger/attachment point. The payout structure is linear, providing the same payout
amount for each mm of excess rainfall until the specified exhaustion/exit point in mm is
reached. The payout per event is determined by the mm of excess rainfall above each region’s
trigger threshold. CP and RF propose to determine both the attachment and exhaustion points
based on each region’s statistical return periods (RP) values.
23. The payout is capped at each region's policy limit/sum insured. Depending on the operational
processes in place, the payout speed of 10 days indicated in Table 3 can be reduced. Following
an exceptionally severe event, a region’s policy limit can be exhausted after only one extreme
event. A PIC could receive multiple payouts due to excess rainfall events per region within an
annual policy.
11
Technical Report by CelsiusPro AG, March 2021
Table 3: Product Structure Details
24. Event Definition Metric
Following the extensive assessment of 18 different precipitation metric in Component 1,
Component 2 focuses on two of them, the unweighted mean value of all GPM IMERG grid cells
(metric 1) and the population-weighted mean of all grid cells (metric 9). The latter is prioritized
compared to the similar method of using the population-weighted mean of the top 5% grid cells
(metric 10) due to its understandable approach. A third option of weighting the grid cells by
either the existing infrastructure or buildings is considered. However, this option is not pursued
in more detail due to the limited availability of granular and uniform infrastructure or building
data. CP and RF ultimately incorporate Metric 1, the unweighted mean value of the GPM IMERG
grid cells in question, as the basis for the product structure. The population weight per PIC
region is subsequently factored in via the coverage/sum insured.
25. Alternative Data Sources
Alternative data sources to GPM IMERG are ERA-5, CHIRPS and Numerical Weather Prediction
(NWP) datasets. In Component 1, CP and RF mention that NWP model forecast are not suitable
for parametric products due to significantly lower reliability compared to satellite data.
Although the ERA-5 dataset comprises a dataset with a long historical record, the spatial and
temporal resolution as well as the latency (monthly), make the dataset unreliable to act as the
sole dataset for a parametric product. The drawbacks in spatial and temporal resolution
increase the risk of payouts being mismatched with actual losses, increasing a product’s basis
risk. ERA-5’s long historical record can be utilized to gain additional insights on a country’s
precipitation patterns as well as increase the robustness of an Extreme Value Analysis, as
described in Section 3.2.
26. On the other hand, CHIRPS data has very high spatial resolution and adds value to an analysis
over longer timescales (1 months, 3months and 6 months) due to the nature of its setup to
monitor seasonal droughts and trend analysis. However, it is outperformed by GPM IMERG1
with regards to the accuracy of metrics at low timescales. Furthermore, temporal resolution (6
hourly), as well as latency (third week of following month) of CHIRPS data, make it less
favorable for a parametric product compared to GPM IMERG. Conceptually, bias-corrected
CHRIPS data could be used to gain further insights on past precipitation patterns indicating
drought emergence.
1
Guoqiang Tang, Martyn P. Clark, Simon Michael Papalexiou, Ziqiang Ma, Yang Hong: Have satellite precipitation
products improved over last two decades? A comprehensive comparison of GPM IMERG with nine satellite and
reanalysis dataset, Remote Sensing of Environment, 2020
12
Technical Report by CelsiusPro AG, March 2021
27. Spatial Resolution
In line with the decision noted in Section 2.2, the GPM IMERG grid cells (roughly 10km x 10km)
are aggregated on a sub-national level according to the political boundaries of the different
regions or climatic zones within each PIC. Consequently, the excess rainfall product structure
allows for the coverage to be divided across a PIC’s regions, providing additional flexibility to
focus the product on priority areas within the country. In addition to the regional split, a
national-level aggregation of grid cells is conducted separately for each PIC.
28. Daily Rainfall Aggregation
The project team applies eight consecutive 3-hourly periods to get a rolling 24-hour
accumulation, whereby the start and end-times are shifted by three hours for each iteration
(see Figure 2). This approach results in eight different daily aggregations, ending between 02:00
and 24:00 UTC of “Day t” in Figure 2, with the highest 24-hour accumulation for each calendar
day being included within the GPM IMERG time series. Applying a less static definition of what
constitutes a daily 24-hour period was confirmed in Component 1 to capture heavy short-
duration precipitation more precisely for the PICs
Figure 2: Example of a rolling 24-hour precipitation aggregation method.
29. The approach of applying a rolling 24-hour period technically allows for intense rainfall taking
place within one 3-hourly period to be reflected in two successive calendar days’ 24-hour
accumulations. This holds especially true when recorded precipitation is low in the preceding
and succeeding seven 3-hourly periods. To address the potential double counting of peak
precipitation, a temporal window of seven days before and after each rolling 24-hour
accumulation is introduced within the product design. The event independence of seven days is
based on findings from the feasibility study’s Component 1. In practice this results in a rolling
24-hour period only being recognized as an excess rainfall event if no higher 24-hour
accumulation is identified in the seven days preceding and the seven days following that date.
30. Return Period Methodology
Based on the GPM IMERG and extended dataset, an Extreme Value Analysis (EVA) is conducted
to determine the Average Recurrence Interval (Return Period) of events in the tail of a
distribution of a population of data. The aim is to derive return periods, with associated
uncertainty bounds, of extreme precipitation to define the attachment and exhaustion points of
a proposed excess rainfall insurance product. As part of the data expansion using ERA reanalysis
data, extrapolations of up to the 150-year return period (RP) interval for each of the seven
Group A countries’ regions are produced.
13
Technical Report by CelsiusPro AG, March 2021
31. The considerably higher concentration of severe precipitation between 1979-2000, especially
pronounced for the highest severity events, strongly increases many of the RP thresholds
derived from the extended dataset (1979-2020). Applying the thresholds derived from the
extended dataset to the GPM IMERG of approximately 20 years, results in a mismatched pricing
structure (see Paragraph 38). Therefore, GPM IMERG RPs of up to 100 years are used in the
product pricing, which considers the GPM IMERG precipitation data between 2001-2020.
Additional details on the EVA methodology and assumptions are based on the Technical Report
– Component 1.
32. Correlation Analysis
During Component 1, RF assessed both the cross-correlation of the grid cells within a country
and the cross-correlation of extreme rain events between the PICs. On the latter, the timing of
the largest three events between 2000 – 2020 for each country was compared to investigate
the likelihood of extreme excess rainfall events occurring simultaneously between countries.
While a few events have impacted both countries, not more than two countries were impacted
by one extreme rainfall event. The results suggest that the occurrence of extreme rainfall
events between countries are largely uncorrelated.
33. As part of the groundwork related to the extreme value analysis in Component 1, RF conducted
a homogeneity assessment. A cross-covariance analysis was undertaken using a 7-day window,
deemed sufficiently large to capture the slowest weather systems' progression within the
territory. For every PIC, each cell’s 20-year daily data series was cross-correlated with every
other cell in the group (i.e. every cell pair) within a 7-day window. The cross-correlation function
measures the similarity between a time series and lagged versions of another time series as a
function of the lag. The assessment found significant cross-correlation between the grid cells of
most PICs (see Table 4). Thus, they can be regarded as experiencing the same (homogenous)
rain climate and consequently also affected by the same excess rainfall events. The three non-
homogenous Group A PICs were split into smaller component parts, allowing the analysis to be
re-run. The splits were guided by the GPM cells' location that was uncorrelated, by plotting
these pairs, spatial clusters of homogenous 7-day rain climates became apparent. This approach
allowed for RMI and the Cook Islands to be split into climatologically homogenous groups by
dividing the islands into a north/south area (around 11.2 and -14 deg latitude, respectively),
while the Solomon Islands are split into east/west groupings (around 164 deg longitude).
34. Consequently, if the single grid cells are correlated, then it can be derived that, for example,
Samoa’s 11 regions would also show a significant cross-correlation among each other. Within
the regional split selected for Component 2, RMI’s regions of North and South are independent,
given the vast geographical spread of the atolls/islands. For the Solomon Islands, 9 out of the 10
regions can be considered significantly cross-correlated, with the eastern region of Temotu
being independent. Finally, the Cook Island’s most populous island of Rarotonga and the
remaining South region are correlated while the region of North is independent. While the
strong correlation between the grid cells’ GPM IMERG time series, i.e. precipitation patterns
between 2001-2020, does not infer that exceptionally severe excess rainfall events must also be
correlated between the regions, the project team expects any correlation analysis of extreme
events to yield similar results.
14
Technical Report by CelsiusPro AG, March 2021
Table 4: Cross-Correlation Results - Component 1
Country Country Name Climatological No. GPM No. cross- All cells cross-
No. Split cells correlations correlated
performed within 7 days?
1 Marshall Islands Required 236 55,460 No
2 Solomon Islands Required 593 351,056 No
3 Fiji None 421 176,820 Yes
4 Samoa None 51 2,550 Yes
5 Tonga None 76 5,700 Yes
6 Cook Islands Required 41 1,640 No
7 Vanuatu None 262 68,382 Yes
1 Marshall Islands North 44 1’892 Yes
1 Marshall Islands South 192 36’672 Yes
2 Solomon Islands East 49 2’352 Yes
2 Solomon Islands West 544 295’392 Yes
6 Cook Islands North 20 380 Yes
6 Cook Islands South 21 420 Yes
35. Weather Index vs Nat Cat Product Methodology
Despite using the same terminology, the conceptual structure of a weather index product, as
applied to the excess rainfall solutions presented herein, differs from parametric natural
catastrophe (nat cat) coverage based on a probabilistic model (see Table 5). The difference
between weather index RP10-20 and RP10-50 products, both with the same coverage limit in
USD, is the lower probability of exhausting the latter product option. Accordingly, a weather
index RP10-50 structure will always represent a less expensive premium option. On the other
hand, a parametric nat cat product with an RP10-50 structure will always represent a more
comprehensive coverage and, therefore, a more expensive product option than an RP10-20
structure.
Table 5: Theoretical Comparison of a Weather Index and Parametric Nat Cat Structure (for an
excess rainfall RP10-20 and RP10-50 structure)
Product Weather Index Parametric Nat Cat
Component
Coverage Limit Identical, e.g. USD 5mn, for both an RP10- Dependent on the modelled
20 and an RP10-50 structure expected loss (EL). For example:
- RP10-20 = USD 5mn xs 2mn
- RP 10-50 = USD 13mn xs 2mn
Tick Value The coverage limit in USD is divided by the Not relevant
coverage length in mm of rainfall. So with a
RP 10 set at 200mm, RP 20 at 300mm, and
RP 50 at 450mm, the tick value would be:
15
Technical Report by CelsiusPro AG, March 2021
- RP10-20 = USD 50k per mm (5mn/100)
- RP10-50 = USD 20k per mm (5mn/250)
Attachment Point / A recorded rainfall event with a 10% A recorded rainfall event with a
Payout Trigger annual probability of occurrence 10% annual probability of
occurrence
Payout Amount for Different for the two coverage options. Identical for the two coverage
an RP 20 event Based on the mm of rainfall exceeding the options. Based on the modeled
attachment point multiplied with the tick loss for a RP 20 event. Resulting in
value in USD. E.g. a payout of, for example, USD
- RP10-20 = USD 5mn (100mm x 50k) 5mn for both RP10-20 and RP10-
- RP10-50 = USD 2mn (100mm x 20k) 50.
36. Excess Rainfall Pricing Tool
As part of this project, a separate indicative premium pricing tool is provided to the World Bank
and PCRIC for each Group A PIC. The 42-year time series for the extended dataset, combining
the ERA-5 bias-corrected and GPM IMERG data, is included in each pricing tool. The rainfall
distribution between 1979 and 2020 is visualized in the “Precipitation Graphs” tab. As described
in Paragraph 28 and 29, the rolling 24-hour accumulation is included as the daily precipitation
value for each country’s region and on a national level, reflected in the “Payout Calculations”
tab. Payouts are calculated for each product option, factoring in the RP thresholds and tick
values for the different regions and the national level. The calculations factor in the seven-day
temporal window preceding and following each daily value. A payout only occurs if the daily
value is the highest across these 15 days. If two events have the exact same value, then only the
event occurring on the earlier date results in a payout. Payout graphs and an overview of the
correlation between the regional payouts are found in the respective tabs.
37. The as-if payouts are then used for the burn cost (BC) calculation included in “Payouts per
year”, deriving the technical premium for each region. Due to existing market practices for
weather index products and CP’s recommendation to only use GPM IMERG, the BC calculation
only factors in the most recent 20 years of as-if payouts (2001-2020). The remaining payouts
remain included for illustrative purposes only. The pricing tool provides the World Bank / PICRIC
with the expected indicative net premium to the reinsurer, after factoring in an unused limit
cost of 1.5%, a volatility loading of 10%, and an assumed 5% margin loading. Depending on the
distribution/broking channel used for the (re)insurance placement, further loadings of between
10-20% can be expected. It is important to note that the pricing tool’s premium calculations
only represent an estimation, the reinsurers' own pricing approach will largely influence the
product’s final price.
38. Following an extensive assessment of the RP thresholds based on both the extended dataset
and GPM IMERG only, the project team has decided to base the pricing tool on the latter.
Utilizing the RPs derived from the extended dataset for a 20-year BC calculation results in a
mismatched pricing outcome, with the triggers not being aligned with the dataset. The RP
thresholds, as shown in "Return Period Comparison” tab, can vary significantly across regions
and national-levels.
16
Technical Report by CelsiusPro AG, March 2021
39. Applying the extended dataset’s RPs as the product thresholds for the GPM IMERG data
between 2001-2020 results in certain regions only triggering minor as-if payouts and some not
resulting in any payout across the full 20 years. Using RP thresholds that are fully aligned to the
underlying dataset provides considerably more evenly distributed premium estimations across
the regions and different PICs.
40. Product Options
The project team proposes to mirror PCRIC’s current TC coverage by setting a fixed attachment
point at a 10-year RP. It is important to note that the attachment and exhaustion points on a
national level do not correspond with the regional level thresholds. The national-level RP is
ultimately lower if a regional level 10-year RP is applied and varies depending on the number of
sub-national regions included in the product structure. Consequently, as a general rule, it is not
advised to reduce the attachment point below a 10-year RP, since the high payout frequency
will reduce the ability to reinsure the product at an affordable price. For those countries with a
high payout frequency, opting for an attachment point at a regional 20-year RP is advisable to
ensure the product remains cost-competitive.
41. The product structure analysis per PIC within this report is focused on the following four
coverage options for both the regional and national structures:
Table 6: Excess Rainfall Product Options Overview
Product Options Attachment Point Exhaustion Point
Coverage 1 10-year Return Period 20-year Return Period
Coverage 2 10-year Return Period 50-year Return Period
Coverage 3 10-year Return Period 100-year Return Period
Coverage 4 20-year Return Period 50-year Return Period
42. The first three products all identify the same excess rainfall events as payouts due to the
uniform attachment point. The differentiation between Coverages 1-3 lies in the successively
lower price points, with the coverage limit distributed across a longer coverage length (in mm
rainfall) and the resulting lower tick value (in USD). Selecting Coverage 2 over Coverage 1 will
result in the same payout pattern but a different payout amount per event (see Figure 3). By
providing three product options with the same attachment point, a suitable indicative premium
for each PIC can be identified at the initial stage. Coverage 4 represents the most cost-effective
option and is included to assess the viability of attaching at a higher threshold for countries that
experience a higher frequency of payouts with Coverage 1-3. The fine-tuning of the structure,
depending on the confirmed available premium per country, can take place at a later stage in
the product development.
17
Technical Report by CelsiusPro AG, March 2021
Figure 3: Illustration of the Payout Functions per Coverage Option
43. It must be noted that precipitation induced floods are primarily seen as a frequency, not a
severity risk. Damages from this peril tend to be driven by aggregates from recurring medium
severity events rather than from rare high severity rainfall. Therefore, excess rainfall coverages
usually look to attach at lower RP thresholds such as RP 5-10 while having exhaustion points of
between RP 20-50. Such structures are generally seen as providing an adequate balance
between coverage and price consideration.
44. The above described and applied methodology results are summarized in the overview of
indicative premiums on for the regional product option in Table 7, and on a national level in
Table 8. The indicative premiums are displayed per country and per different coverage options
from Coverage 1 (RP 10-20) to Coverage 4 (RP 20-50). The indicative premiums are listed in
notional values in USD and as a percentage of the coverage limit, both based on a coverage limit
of USD 5’000’000. It is important to note that the indicative premium is linear to the limit,
doubling when the coverage is increased to USD 10’000’000. Therefore, the indicative premium
in percentage remains the same and in turn allows coverage limits to be estimated based on
premium inputs.
45. Within one country, the indicative premiums are highest for Coverage 1 and decrease towards
Coverage 2, 3. Indicative premiums for Coverage 4, with its higher attachment point, are usually
but not always lower than Coverage 3. On the regional level the indicative premiums range from
the lowest value of 3.1% for Coverage 4 for Vanuatu to a maximum value of 16.8% for Samoa’s
Coverage 1. On a national level the indicative premiums range from the lowest value of 1.6% for
Coverage 4 for Vanuatu to a maximum value of 19.4% for Coverage 1 for both Solomon and
Samoa. The indicative premiums per country will be elaborated and discussed in more detail in
Sections 3.3 - 3.9.
46. For Vanuatu, Coverage 3 is indicated as n/a, since no representative value for
attachment/exhaustion point for RP100 and 150 with IMERG GPM only data could be modeled
due to certain significant outlier events at the tail end of the curve.
18
Technical Report by CelsiusPro AG, March 2021
Table 7: Pricing Overview - Regional Product Options (Indicative Premiums based on a coverage
limit of USD 5’000’000)
Regional Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Vanuatu Premium (USD) USD 586’166 USD 329’417 n/a USD 155’864
Vanuatu Premium (%) 11.7% 6.6% n/a 3.1%
Tonga Premium (USD) USD 407’787 USD 362’467 USD 286’400 USD 334’209
Tonga Premium (%) 8.2% 7.2% 5.7% 6.7%
Solomon Premium (USD) USD 672’783 USD 524’662 USD 401’645 USD 425’097
Solomon Premium (%) 13.5% 10.5% 8.0% 8.5%
Samoa Premium (USD) USD 838’849 USD 481’723 USD 359’552 USD 224’156
Samoa Premium (%) 16.8% 9.6% 7.2% 4.5%
RMI Premium (USD) USD 440’331 USD 321’123 USD 254’706 USD 235’359
RMI Premium (%) 8.8% 6.4% 5.1% 4.7%
Fiji Premium (USD) USD 605’760 USD 432’661 USD 326’812 USD 201’448
Fiji Premium (%) 12.1% 8.7% 6.5% 4.0%
Cook Islands Premium (USD) USD 403’440 USD 326’904 USD 277’097 USD 275’624
Cook Islands Premium (%) 8.1% 6.5% 5.5% 5.5%
Table 8: Pricing Overview – National Level Product Options (Indicative Premiums based on a
coverage limit of USD 5’000’000)
National Level Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Vanuatu Premium (USD) USD 316’322 USD 158’830 n/a USD 78’750
Vanuatu Premium (%) 6.3% 3.2% n/a 1.6%
Tonga Premium (USD) USD 260’459 USD 217’737 USD 181’162 USD 166’600
Tonga Premium (%) 5.2% 4.4% 3.6% 3.3%
Solomon Premium (USD) USD 969’152 USD 712’279 USD 511’159 USD 532’158
Solomon Premium (%) 19.4% 14.2% 10.2% 10.6%
Samoa Premium (USD) USD 969’152 USD 624’005 USD 459’399 USD 367’408
Samoa Premium (%) 19.4% 12.5% 9.2% 7.3%
RMI Premium (USD) USD 514’401 USD 318’702 USD 254’715 USD 176’334
RMI Premium (%) 10.3% 6.4% 5.1% 3.5%
Fiji Premium (USD) USD 753’301 USD 698’799 USD 532’113 USD 490’300
Fiji Premium (%) 15.1% 14.0% 10.6% 9.8%
Cook Islands Premium (USD) USD 753’301 USD 649’436 USD 549’637 USD 544’281
Cook Islands Premium (%) 15.1% 13.0% 11.0% 10.9%
19
Technical Report by CelsiusPro AG, March 2021
3.2 GPM IMERG data expansion using ERA reanalysis data
47. As previously highlighted in the Technical Report – Component 1, GPM IMERG is made available
from mid-2000 onwards. This vantage point permits a robust assessment of each PIC’s recent
excess rainfall events, empirically determining the precipitation volume for incidents with a
return period (RP) of up to 20 years. Excess rainfall-induced flooding is considered a natural
disaster with a relatively frequent occurrence, allowing a 21-year time series to provide strong
indications of a PIC’s near-future hazard risk. However, there remains a limitation when
attempting to provide potential clients with a product addressing excess rainfall events with a
lower frequency and higher severity. Following international norms (ISO 19901-1), a 21-year
track record can be extrapolated to a maximum return period of approximately 80 years. To
assess the benchmark 100-year return period event or longer with additional confidence, access
to a longer historical time series is required.
48. Following the request by the World Bank, CP and RF address the need for a longer view of
historic rainfall in the Pacific by combining the GPM IMERG with the ERA-5 dataset, which
features datapoints reaching back to 1979.2 The extended time series of 42 years doubles the
number of observed excess rainfall events, allowing for return periods to be robustly
extrapolated to approximately 150 years.
49. To this end, RF assesses the suitability of two reanalysis products for the extension of the
rainfall data over PICs; ERA-5 Land and ERA-5. ERA-5 Land covers only land areas and is at a
comparable 0.1-degree (~10 km) resolution to GPM IMERG, whereas ERA-5 covers both land
and water but is gridded at a 0.25-degree (~25-km) resolution. The higher resolution ERA-Land
is found to not capture all of the PIC’s islands, with some of the small landmasses not being
picked up in the data. Therefore, the analysis is undertaken with ERA-5 data for the area
covering the cells for the seven Group A PICs.
50. ERA-5 precipitation data is downloaded at 1-hourly resolution for the period January 1979 to
December 20203 across all PIC grid cells. The data is then accumulated to 3-hourly intervals and
aligned with time periods used by GPM (0200, 0500, 0800 etc.). The maximum value from each
24-hour accumulation period on a rolling 3-hourly window is extracted to arrive at time-aligned
maximum daily values. The ERA-5 data is downscaled and bias-corrected to the GPM data, on a
cell-by-cell basis, for the overlapping period 2000 – 2020, using a quantile mapping approach.
This bias-correction method is commonly used when adjusting climate models to observational
data, with a detailed methodology outlined in Annex 1 - Data Expansion Method Report. Finally,
given the focus on sub-national regions within each PIC, the downscaled and bias-corrected
ERA-5 data is then aggregated according to the regions' administrative or climatic boundaries.
The resulting bias-corrected ERA-5 data provides a homogenized rainfall time series from 1979
2
The ERA-5 reanalysis extends back further than 1979, although 1979 is regarded as the beginning of the assimilation
of satellite-derived observations into the reanalysis. Therefore, the time series 1979 to present can be regarded as
homogenous in terms of reanalysis methodology, whereas the period prior to 1979 includes considerably less global
observations, introducing potential heterogeneity into extracted time series.
3
ERA-5 is only available for download at either hourly or monthly resolutions. Data is obtained up to December 2020,
although the most recent data has not been quality-controlled and is termed ERA-5 Interim. On inspection, erroneous
data values are found in this period, and for this reason only data up to 31 August 2020 is used.
20
Technical Report by CelsiusPro AG, March 2021
– 2020 (42 years) with the extended dataset used to fit and extrapolate an appropriate extreme
value distribution to peaks.
51. Findings – Extended Dataset
The data extension analysis leverages a widely-used climate reanalysis dataset, ERA5, to extend
the rainfall time series for each PIC from 2000 – 2020 (GPM period) to 1979 – 2020, more than
doubling the data record. The aim of this work is twofold; first, to examine whether using 41 full
years of rain data (1979 – 2019) instead of 19 (2001 – 2019) could reduce the confidence
intervals of long return period rain volumes for PICs and their regions, and secondly, whether
the pre-GPM period can provide further insight into the probabilities of extreme rainfall across
the Pacific region.
52. Main Findings
The analysis demonstrates that using 41 annual maxima peaks instead of 19 does indeed reduce
the confidence intervals attached to long return periods in almost all of the 42 provinces across
the 7 PICs analyzed, as expected. It also demonstrates that the period 1979 – 2000 is likely to
have been wetter – and therefore more flood-prone – than the period captured by GPM (2001 –
2020). RF has undertaken a detailed validation of the bias-correction methodology focusing on
extremes, and indeed sensitivity tested several different bias-correction methods, to ensure the
finding of a wetter period prior to 2000 is not simply a function of a non-homogenous
ERA5/GPM blended time series. This finding is widely supported in the literature and is also
apparent in the non-corrected ERA5 data.
53. Pacific Climate Variability
Rainfall variability across the Pacific Islands is strongly modulated on interannual timescales by
the El Niño Southern Oscillation (ENSO)4, and on multidecadal timescales by the Interdecadal
Pacific Oscillation (IPO)5,6. These Pacific-wide climate drivers strongly influence the location of
the South Pacific Convergence Zone (SPCZ), which is a region of increased convective activity
stretching in a southeast direction from near Papua New Guinea toward the Cook Islands
(Figure 4). During La Niña years the axis of the SPCZ moves southwest bringing increased
precipitation to islands such as Fiji and Vanuatu while reducing precipitation over islands to the
northeast, such as Samoa and French Polynesia—this situation is reversed during El Niño years5.
4McPhaden, M. J., Zebiak, S. E. and Glantz, M. H.: ENSO as an Integrating Concept in Earth Science, Science,
314(5806), 1740–1745, doi:10.1126/science.1132588, 2006.
5Folland, C. K., Renwick, J. A., Salinger, M. J. and Mullan, A. B.: Relative influences of the Interdecadal Pacific
Oscillation and ENSO on the South Pacific Convergence Zone, Geophys. Res. Lett., 29(13), 1643,
doi:10.1029/2001GL014201, 2002.
6McGree, S., Schreider, S. and Kuleshov, Y.: Trends and Variability in Droughts in the Pacific Islands and Northeast
Australia, J. Climate, 29(23), 8377–8397, doi:10.1175/JCLI-D-16-0332.1, 2016.
21
Technical Report by CelsiusPro AG, March 2021
Figure 4 Illustration of the South Pacific Convergence Zone (Australian Bureau of
Meteorology and CSIRO, 2011)
54. The IPO is similar to ENSO, but operates on approximately 15 – 30 year timescales. Figure 5
shows IPO variability since 18907. During 1976 to 1998 the IPO was in its’ positive (El Niño like)
state resulting in the mean position of the SPCZ being displaced further northeast compared to
the negative (La Niña like) phases (1958 to 1975 and 1999 to present)8,9.
Figure 5 IPO index from 1890 to 2010 (Henley et al 2015)
7Henley, B. J., Gergis, J., Karoly, D. J., Power, S. and Kennedy, J.: A tripole index for the interdecadal Pacific oscillation,
Clim. Dynam., doi:10.1007/s00382-015-2525-1, 2015.
8Folland, C. K., Renwick, J. A., Salinger, M. J. and Mullan, A. B.: Relative influences of the Interdecadal Pacific
Oscillation and ENSO on the South Pacific Convergence Zone, Geophys. Res. Lett., 29(13), 1643,
doi:10.1029/2001GL014201, 2002.
9Salinger, M. J., Renwick, J. A.Mullan, AB: Interdecadal Pacific oscillation and south Pacific climate, Int. J. Climatol.,
21(14), 1705–1721, doi:10.1002/joc.691, 2001.
22
Technical Report by CelsiusPro AG, March 2021
55. Variations in the location of the SPCZ influence rainfall over the South Pacific Islands including
extreme rainfall associated with tropical cyclone activity10. Because the mean position of the
SPCZ varies in conjunction with the IPO on multidecadal timescales, rainfall totals from many
islands differ substantially between different phases of the IPO11. Sampling from only the past
20 years, when the IPO has been mostly in its’ negative (La Niña like) state provides only a
limited view of the actual range of natural climate variability.
56. Impacts of Pacific Climate Variability on Return Period Estimates
As a result of this extended sampling across multidecadal climate variability, the return period
curves for most provinces shift upward reflecting a higher risk environment compared to the
GPM-only period. This has an impact on the product pricing.
57. Summary of Outputs from Data Extension Work
The outputs from this data extension work include:
A report detailing the methodology used and findings.
Blended, homogenous ERA5/GPM time series (24-hr accumulations of 3-hrly base data)
for each of the 42 provinces from 1 Jan 1979 – 18 Dec 2020, and the same again at the
national level for each of the 7 PICs
Return period plots showing GPM-only and ERA5/GPM blended estimates for each of
the provinces and at the national level for each PIC
A set of tables within the report that detail the best-fit distribution and parameters of
both the GPM-only and ERA5/GPM blended return period estimates, for each of the 42
provinces and at the national level for each PIC
The distribution and parameter information can be used by the World Bank to recreate
all of the return period curves at a later date. In addition, the homogenized 42-year time
series for each province/country may provide a valuable data source for further
analyses of flood risk across the Pacific.
These outputs have been partially included in deliverables (report and pricing tool).
10Vincent, E. M., Lengaigne, M., Menkes, C. E., Jourdain, N. C., Marchesiello, P. and Madec, G.: Interannual variability
of the South Pacific Convergence Zone and implications for tropical cyclone genesis, Clim. Dynam., 36(9), 1881–1896,
doi:10.1007/s00382-009-0716-3, 2011.
11McGree, S., Schreider, S. and Kuleshov, Y.: Trends and Variability in Droughts in the Pacific Islands and Northeast
Australia, J. Climate, 29(23), 8377–8397, doi:10.1175/JCLI-D-16-0332.1, 2016.
23
Technical Report by CelsiusPro AG, March 2021
3.3 Product Analysis – RMI
58. As-If Analysis – Product Options
The Technical Report – Component 1 finds little agreement between RMI’s recorded historical
loss events and the excess rainfall events observed in the GPM IMERG data. The two flood, two
TC and one tropical depression (TD) events flagged within the event catalog are not aligned with
the GPM IMERG peaks recorded in 2003 and 2006 (see Figure 7). Figure 6 provides an overview
of the observable precipitation figures, instead of the economic damages, for the five events
flagged in the historical event catalog.
GPM IMERG Marshall Islands
Precipitation in mm
80
TC Bavi Flood Event
70
60
50
40
30
20
TD Dolphin TC Penny
10 Flood Event
0
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
TC losses non TC losses
Figure 6: Historical Loss Event Catalog – RMI, based on the Technical Report – Component 1.
National precipitation levels are included due to a lack of economic damages information.
59. The pricing tool identifies four as-if payouts, including a marginal payout in 2015, for the
regional Coverages 1-3, overwhelmingly driven by two severe rainfall events in 2003 and 2006.
As visible in Figure 7, the excess rainfall event in 2003 is reached close to the RP50 threshold,
with the exhaustion point for Coverage 2 almost being reached that year. The 2006 event also
exceeds the RP20 threshold since Coverage 4 is also triggered, but the as-if payouts remain
further apart due to the difference in tick values between the coverage options. Including the
fractional payout in 2015, the payout frequency for all regional coverages indicates the product
has a return period of 6.67 years.
24
Technical Report by CelsiusPro AG, March 2021
Figure 7: Total as-if payouts for all regions between 2001 to 2020 – RMI
60. The payout frequency curves for Coverage 1 and 4 are highlighted in Figure 8, plotting the as-if
payouts for the two options by their size. The curves highlight the substantial difference in the
two largest payouts between the two coverages.
Figure 8: Payout frequency curve – RMI, for regional Coverage 1 - RP 10-20 and Coverage 4 - RP
20-50.
61. Regional Comparison
Figure 9 demonstrates the dominating effect of RMI South’s 98% share of the coverage
limit/sum insured, with almost the entire population residing within this region. Even though
RMI North data includes an RP50 event in 2004, the payout of USD 100,000 remains marginal.
25
Technical Report by CelsiusPro AG, March 2021
Figure 9: Comparison of as-if payouts across RMI's regions - Coverage 1. Based on an annual
coverage limit of USD 5’000’000.
62. Basis Risk Considerations
Based on the current information available, the proposed excess rainfall product would contain
a substantial basis risk for RMI. There is a complete mismatch between the observations
derived from the GPM IMERG data and the historical loss event catalog. No alternative product
structure would come close to triggering a payout for TC Bavi and the flood event in 2015 (see
Figure 6). To fully understand the mismatch and basis risk implication of the product, the actual
impact of the catalog’s historical loss events and the excess rainfall events in 2003, 2006, and
2004 would need to be gained. If the actual losses are truly centered around the 2003 and 2006
events, having impacted the key population cluster in RMI South, then the basis risk concerns
could be alleviated.
63. Feasibility Assessment and Product Recommendation
The RMI’s population distribution, and consequently also the standard allocation of the
coverage limit/sum insured, is strongly skewed towards the RMI South region, which accounts
for 98% of the PIC’s inhabitants. Average precipitation levels are also considerably higher in RMI
South, even though RMI North does record frequent excess rainfall events with comparable
precipitation levels, resulting in relatively similar RP threshold structures.
64. As noted in Section 2.3, RMI is split according to two differentiating climatic zones, splitting the
vast area covered by the atolls and islands into a southern and northern region. Consequently,
applying a national-level structure is technically unfeasible and strongly advised against by the
project team. The national-level RP averages out GPM IMERG cells which should not be
combined, a key factor behind the low thresholds and high payout levels noted in Table 9.
26
Technical Report by CelsiusPro AG, March 2021
65. Looking at the sprawling composition of RMI’s islands/atolls and the population concentration
around Majuro and other southern islands/atolls, the country is probably best served with a
coverage specifically targeting RMI South. Among the different regional coverage options,
Coverage 1 should be targeted. The low as-if payout frequency requires an attachment point at
RP 10 and the higher price for this coverage is legitimized by Coverage 1 providing a higher
payout for rainfall severities up to RP 20. Coverage 2 is the only other option that could be
considered, given relatively high rainfall variability and the slightly more cost-efficient indicative
premium costs. Nevertheless, proceeding with such a product requires an in-depth prior
assessment of the actual flood induced damages affecting the key population centers. With the
current level of information on historical loss events, an excess rainfall product should not be
pursued for RMI.
Table 9: Product Options Comparison – RMI (Indicative Premiums based on a coverage limit of
USD 5’000’000)
Regional Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
RMI Premium (USD) USD 440’331 USD 321’123 USD 254’706 USD 235’359
RMI Premium (%) 8.8% 6.4% 5.1% 4.7%
National Level Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
RMI Premium (USD) USD 514’401 USD 318’702 USD 254’715 USD 176’334
RMI Premium (%) 10.3% 6.4% 5.1% 3.5%
Note: RMI’s indicative premiums are below the historical average payout rates over 20y since
there have not been payouts in the last 10y. Reinsurers would likely increase premiums to reach
historical burn cost (Example Coverage 1 has an average payout of 495k and a calculated
indicative expected premium of 440k).
27
Technical Report by CelsiusPro AG, March 2021
3.4 Product Analysis – Solomon Islands
66. As-If Analysis – Product Options
The Solomon Islands have an extensive record of historical loss events, with 18 out of 22 events
including an actual economic loss record. Most of the incurred losses are noted as being
relatively minor, with only seven events exceeding trended economic losses of above USD 2
million. While most events are associated with TCs, the country has faced some substantial,
heavy rainfall losses, most notably in 2014. As visualized in Figure 10, the loss event catalog
highlights two substantial TC losses stemming from TC Oma in 2018 and TC Harold in 2020.
Figure 10: Historical Loss Event Catalog – Solomon Islands based on the Technical Report –
Component 1
67. The as-if payouts provide a mixed picture of the product’s alignment with the loss event
catalog. The major flood event in 2014 is reflected as a key excess rainfall event within the data.
The other major as-if payout visible in Figure 11 is associated with the rains accompanying TC
Freda in 2012, which is not flagged as one of the significant loss events in the historical catalog.
The payouts noted in 2015 are in line with TC Pam and TC Raquel, which affected the Solomon
Islands during that year. The missing payout for the 2010 flood is explained by the event being
driven by storm surge, while the losses from the recent major TCs are expected to be largely
driven by wind speed damages. However, the smaller as-if payouts between 2001 to 2011,
illustrated in Figure 11, are not aligned with the minor events recorded in the event catalog for
the respective year or fall within years without an official loss record.
68. The difference between the coverage options is evident when contrasting the payout events in
2012 and 2014. The precipitation values for the event in 2014 are well above the RP50
threshold for many provinces, exhausting annual coverage limits and leading to comparable
payouts across the options. While the rainfall-induced by TC Freda also affected multiple
regions, the daily values remain below the RP 20 threshold for most, with the higher tick value
of Coverage 1 resulting in a considerably higher total payout.
28
Technical Report by CelsiusPro AG, March 2021
69. The high payout frequency observed in Figure 11 is driven by smaller regions such as Choisuel,
Isabel, Temotu as well as Rennell and Bellona triggering payouts from relatively severe but
independent events from other provinces. Aside from Temotu, which is climatologically
independent, this finding is surprising. Coverage 1 would payout in 10 out of 20 year, indicating
a very frequent return period of 2 years. However, the annual payout only exceeds USD
500’000, or 10% of the coverage limit, in four years. Even Coverage 4, which records seven
years with payouts, results in a return period of 2.9 years.
Figure 11: Total as-if payouts for all regions between 2001 to 2020 – Solomon Islands
70. The payout frequency curves for Coverage 1 and 4 noted in Figure 12, highlight two relatively
closely aligned curves with similar slopes, with Coverage 4 paying out consistently lower
amounts. The smoother curve shapes found for the Solomon Islands are also related to the
higher payout frequency.
Figure 12: Payout frequency curve – Solomon Islands, for regional Coverage 1 - RP 10-20 and
Coverage 2 - RP 20-50.
29
Technical Report by CelsiusPro AG, March 2021
71. Regional Comparison
While the most populous region of Guadalcanal does account for the highest total payout
across the 20- year period, the third and fourth most populated regions of Honiara and Western
seem more strongly affected by excess rainfall than Malaita. Figure 13 demonstrates the
widespread of payouts across all ten regions.
Figure 13: Comparison of as-if payouts across Solomon Islands’ regions - Coverage 1. Based
on an annual coverage limit of USD 5’000’000
72. Basis Risk Considerations
The basis risk of the product structure for the Solomon Islands depends strongly on the
assumption that the TCs affecting the Solomon Islands in 2018 and 2020 are confirmed to have
primarily caused wind speed damages and didn’t result in precipitation induced floods.
Furthermore, the high payout triggered in 2012 during TC Freda warrants a closer review. While
lower category storms can indeed bring more rainfall with them, the substantial excess rainfall
recorded would have to be observable in actual losses. The product does seem to be aligned
with the severe heavy rainfall losses observed in 2014. But the frequent mismatches in other
years, despite a relatively extensive historical loss catalog, imply a higher basis risk for the
Solomon Islands.
73. Feasibility Assessment and Product Recommendation.
Table 10 also highlights the high costs which can be expected for a national-level coverage, with
payouts only triggered in 2012 and 2014. A national-level product captures the 2014 flood event
well, paying out the full sum insured, compared to the regional Coverage options that never
exhausts all ten region’s sub-limits. Nevertheless, the national-level product overlooks the
various regional severe excess rainfall events observable in the data. Given the product’s
objective of addressing medium severity losses to the country, a national-level product does not
fulfill this aspect. National-level options should only be considered if the Solomon Islands aims
to target severe rainfall events affecting large swaths of the islands.
30
Technical Report by CelsiusPro AG, March 2021
74. The payout frequency and high local severity of excess rainfall events, results in higher
indicative premium rates for all coverage options attaching at RP10. A product with the
potential to trigger at this frequency results in higher annual premiums despite the overall limit
not being fully utilized due the coverage being split across the ten regions.
75. To reduce the payout frequency of the product and the annual costs incurred, CP recommends
utilizing Coverage 4 on a regional level. By attaching at RP20, the Solomon Islands would avoid
some of the localized which are not reflected in the historical loss event catalog, for example, in
2001 and 2011. Therefore, it can be concluded that Coverage 4 reduces basis risk while still
capturing localized higher severity, which has the potential to cause considerable damage on
the islands/regions in question. Proceeding with such an excess rainfall product for the Solomon
Islands is deemed feasible, despite the remaining basis risk, once the loss severity of TC Freda in
2012 has undergone a closer assessment.
Table 10: Product Options Comparison – Solomon Islands (Indicative Premiums based on a
coverage limit of USD 5’000’000)
Regional Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Solomon Premium (USD) USD 672’783 USD 524’662 USD 401’645 USD 425’097
Solomon Premium (%) 13.5% 10.5% 8.0% 8.5%
National Level Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Solomon Premium (USD) USD 969’152 USD 712’279 USD 511’159 USD 532’158
Solomon Premium (%) 19.4% 14.2% 10.2% 10.6%
31
Technical Report by CelsiusPro AG, March 2021
3.5 Product Analysis – Samoa
76. As-If Analysis – Three Product Options
In the Technical Report – Component 1, Samoa has a relatively strong correlation between
recorded historical loss events and extreme rainfall events observed in the GPM IMERG data.
Next to TC Evan in 2012 and a flood event in 2003, TC Amos is also noted in both the historical
losses and precipitation data. The third flood event in 2001, highlighted in Figure 14, is not
observed in the data.
Figure 14: Historical Loss Event Catalog – Samoa, based on the Technical Report – Component 1
77. An analysis of the as-if payouts comes to a similar conclusion, with the events in 2012, 2003,
and 2016 featured as the major as-if payouts. However, as visible in Figure 15, all three product
structures would have also triggered payouts in 2004, 2006, and 2018, which are not featured
within the historical loss events.
78. The stable divergence in payouts levels across Coverages 1-4 indicates that most loss events all
are centered around the RP20 threshold across most of Samoa’s regions. The excess rainfall
events in 2003 and 2005 fall below the RP20 threshold, since no payout is triggered for
Coverage 4. The rainfall resulting from TC Amos in 2016 triggers payouts in 10 out of 11 regions,
illustrating that Coverage 1 performs better when faced with a lower severity but widespread
excess rainfall event. In total, coverages with an RP10 attachment point payout in 7 of the 20
years, compared to 5 for Coverage 4. This results in a product return period of 2.9 and 4 years,
respectively.
32
Technical Report by CelsiusPro AG, March 2021
Figure 15: Total as-if payouts for all regions between 2001 to 2020 – Samoa
79. The strongly diverging payout frequency curves illustrated in Figure 16, point to Coverage 4
being an unsuitable option for Samoa. The curve slope for Coverage 1 implies a good
distribution of the as-if payouts between the 10-year attachment and the 20-year exhaustion
points.
Figure 16: Payout frequency curve – Samoa, for regional Coverage 1 - RP 10-20 and Coverage 4 -
RP 20-50.
33
Technical Report by CelsiusPro AG, March 2021
80. Regional Comparison
The impact of the population-weighted coverage is demonstrated in Figure 17, where the most
populous region of Tuamasaga dominates the as-if payout related to all three major events. The
two following regions of A’ana and Atua also reflect the second and third highest aggregated
payouts across the observed period. Based on the assumption that Samoa's actual losses are
largely driven by flooding in populated areas, the adopted regional weighting functions well for
the country. This holds especially true since the two major as-if payouts in 2003 and 2012 match
with historical losses.
Figure 17: Comparison of as-if payouts across Samoa's regions - Coverage 1. Based on an annual
coverage limit of USD 5’000’000
81. Basis Risk Considerations
The basis risk assessment of the proposed excess rainfall product for Samoa is restricted by only
two historical loss events recording economic losses (see Figure 14). The product does seem to
perform well in capturing the devastation caused by TC Evan and the flood event in 2003. The
performance of the product depends on the actual damages inflicted by TC Amos in 2016, which
would need to be recorded as one of most devastating TCs impacting Samoa since 2001. Subject
to this finding, the product seems to result in a tolerable levels of basis risk, with most historical
loss events being identified within the data, considering the mismatched events as-if payouts in
2004, 2006 and 2018 remain relatively small.
82. Feasibility Assessment and Product Recommendation
The national-level product options do well in triggering large payouts for TC Evan and TC Amos.
Combined with the effect of the volatility loading in the pricing methodology, this leads to high
indicative premium costs, as highlighted in Table 11. The national-level product is not advised
for further pursual due to the high costs combined with these product options not capturing the
major excess rainfall induced flood event in 2003,.
34
Technical Report by CelsiusPro AG, March 2021
83. Among the regional product options, Coverage 1 performs excellently in capturing the major
historical loss events. The high payouts associated with this structure result in annual premium
costs assumed to be above the acceptable levels. To bring the cost back to attainable levels, the
coverage length in mm needs to be increased, lowering the probability of the limit being fully
exhausted. CP recommends considering Coverage 2 on a regional level for Samoa, especially
since certain regions' maximum rainfall since 2001 reached this threshold.
Table 11: Product Options Comparison – Samoa (Indicative Premiums based on a coverage limit
of USD 5’000’000)
Regional Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Samoa Premium (USD) USD 838’849 USD 481’723 USD 359’552 USD 224’156
Samoa Premium (%) 16.8% 9.6% 7.2% 4.5%
National Level Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Samoa Premium (USD) USD 969’152 USD 624’005 USD 459’399 USD 367’408
Samoa Premium (%) 19.4% 12.5% 9.2% 7.3%
35
Technical Report by CelsiusPro AG, March 2021
3.6 Product Analysis – Fiji
84. As-If Analysis – Three Product Options
Fiji includes the largest number of records among the PICs featured in the historical loss event
catalog created during Component 1. The TC and flood events visible in Figure 18 represent the
eight largest events according to the country's actual losses. A further 21 events are flagged in
the catalog for Fiji, many of them with minor losses in the low single-digit millions of USD. While
TC losses outweigh the floods caused by heavy rainfall, driven by TC Winston in 2016, Fiji
includes a balanced number of both types of events.
Trended economic loss Fiji
(USD million
800
TC Winston
700
600
500
400
300 TC Evan TD 17F
200
Flood Event Flood Event
100
TC Gene TC Harold
0
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
TC losses non TC losses
Figure 18: Historical Loss Event Catalog – Fiji, based on the Technical Report – Component 1
85. The as-if payouts for the regional product are strongly focused on the year 2009 and 2012 (see
Figure 19). Out of the eight large historical events documented, the product would have
triggered a payment for the 2007 and 2009 flood event as well as TC Evan and TD 17F in 2012.
While TC Evan, TC Harold and the most recent TC Yasa in December 2020 are observable within
the GPM IMERG data with above-average excess rainfall values, their precipitation levels remain
well below the 10-year RP in Fiji’s five regions. The flood event recorded in the catalog for
March 2020 is barely detectable within the GPM IMERG data. Furthermore, the payouts in 2003
and 2010 fall into years without a loss record in the catalog.
86. The payout distribution across the different coverage options is relatively consistent. Out of the
eight years with recorded payouts for Coverages 1-3, half can be considered relatively minor,
with payouts below USD 100,000. The product’s return period can therefore be pinned at either
2.5 years or 5 years, when excluding the minor payouts. While Coverage 4 eliminates some of
these small payments, it still results in six years with payouts, two of them minor, resulting in
product return periods of between 3.3 years and 5 years.
36
Technical Report by CelsiusPro AG, March 2021
Figure 19: Total as-if payouts for all regions between 2001 to 2020 – Fiji
87. Figure 20 illustrates that despite a relatively high payout frequency, the payout curve for
Coverage 4 flattens out very quickly, making the product option unsuitable for Fiji’s needs. The
curve shape for Coverage 1 visualizes the high divergence between the two largest payouts, the
two medium sized events and the remaining smaller payouts.
Figure 20: Payout frequency curve – Fiji, for regional Coverage 1 - RP 10-20 and Coverage 2 - RP
20-50.
88. Regional Comparison
Unsurprisingly, given their population weighting of 44% and 38%, the regions of Central and
Western account for almost all of Fiji’s as-if payouts (see Figure 21). Nonetheless, the other
three regions trigger at least one payout each since 2001, independently from Central and
Western in 2003, 2007, 2008, and 2014. This finding explains the high frequency of minor
payouts observed within the regional products. The payout distribution across the years could
be considerably altered if the weighting across the regions was adjusted.
37
Technical Report by CelsiusPro AG, March 2021
Figure 21: Comparison of as-if payouts across Fiji's regions - Coverage 1. Based on an annual
coverage limit of USD 5’000’000.
89. Basis Risk Considerations
Despite the relatively extensive historical loss event catalog for Fiji, the as-if analysis triggers a
few payments for excess rainfall events not included in the current records. The product’s
payouts do align with certain key events such as TD 17F and the 2009 flood event. The
disproportionately low payout related to the 2007 flood event is mainly due to the weighting of
the regions. The three recent TCs, namely Evan, Winston and Harold, might well be found to
primarily have caused wind speed damages, and as such not an indication of the basis risk of the
excess rainfall product. The two larger flood events recorded in 2005 and 2020 which wouldn’t
have triggered a payout, warrant a closer review since this finding would imply a higher basis
risk for the proposed product due to no payouts being triggered.
90. Feasibility Assessment and Product Recommendation
Given the high number of recorded historical loss events, a national-level product that only
triggers during higher severity and geographically widespread events, such as those in 2012 and
2009, is not suitable for Fiji’s needs. Furthermore, Table 12 point to the relatively high cost of
national-level products.
91. In Fiji’s case, the high payout frequency seems somewhat warranted given the extensive
historical loss catalog. There are some indications, such as the frequent minor payouts and the
2007 flood event, that the product’s alignment with actual losses could be improved by
adjusting the weighting of the regions, decreasing the dominance of Central and Western. A
rebalanced Coverage 2 on a regional level could be a feasible solution for the country. However,
concerns remain regarding mismatch with major loss events in 2005, 2016 and 2020. Prior to
proceeding with the excess rainfall product, these TCs and floods should be reviewed in detail,
to understand their loss drivers.
38
Technical Report by CelsiusPro AG, March 2021
Table 12: Product Options Comparison – Fiji (Indicative Premiums based on a coverage limit of
USD 5’000’000)
Regional Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Fiji Premium (USD) USD 605’760 USD 432’661 USD 326’812 USD 201’448
Fiji Premium (%) 12.1% 8.7% 6.5% 4.0%
National Level Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Fiji Premium (USD) USD 753’301 USD 698’799 USD 532’113 USD 490’300
Fiji Premium (%) 15.1% 14.0% 10.6% 9.8%
3.7 Product Analysis – Tonga
92. As-If Analysis – Three Product Options
Tonga’s recorded historical loss events are exclusively focused on named TCs, with TC Waka in
2001 and TC Harold in 2020 recording considerable damages across the islands. Next to the
storms with recorded losses as per Figure 22, a further seven TCs are noted of having taken
place between 2000 to 2020.
Figure 22: Historical Loss Event Catalog – Tonga, based on the Technical Report – Component 1
93. The lack of heavy rainfall events within the available event catalog limits the ability to match
excess rainfall events reported in the GPM IMERG data. Of the 11 TCs recorded for Tonga, the
product structure only records a payout associated with TC Waka in 2001. All the other
payments which would have been triggered are assumed to be related to heavy rainfall or lower
windspeed storms/tropical depressions, which remain unnamed.
39
Technical Report by CelsiusPro AG, March 2021
94. As visible in Figure 23, the 2001 and 2009 payout events account for a vast majority of the total,
despite Coverages 1-3 triggering eight times since 2001. The product’s return period would thus
be at 2.5 years, increasing to 4 years if payouts below USD 100,000 are not considered.
Coverage 4 also records five payouts bringing its return period to 4 years. While the 2001
payout related to TC Waka does trigger a payout in the directly affected region of Vava’u, a
majority of the payouts from that year stems from an excess rainfall event in Tongatapu not
associated with the storm.
Figure 23: Total as-if payouts for all regions between 2001 to 2020 – Tonga
95. Figure 24 features a wide disparity between the two largest payouts triggered by Coverage 1
and the remaining small payouts. The payout curve for Coverage 4 is very steep due to the large
difference between the first and the second largest as-if payout recorded.
Figure 24: Payout frequency curve – Tonga, for regional Coverage 1 - RP 10-20 and Coverage 2 -
RP 20-50.
40
Technical Report by CelsiusPro AG, March 2021
96. Regional Comparison
The dominance of Tongatapu within the product structure is visualized in Figure 25 with two
events affecting the region accounting for a vast majority of the product payouts. The 77%
share of the sum insured results in such a dominance, despite the other regions also
experiencing severe rainfall events and associated payouts across the 20-year period reviewed.
Figure 25: Comparison of as-if payouts across Tonga's regions - Coverage 1. Based on an annual
coverage limit of USD 5’000’000.
97. Basis Risk Considerations
Assessing the basis risk of Tonga’s excess rainfall product is challenging due to the historical loss
event record focusing entirely on TCs. With only a single match to one of the 11 storms
recorded, this could indicate that catastrophe damages in Tonga are primarily driven by TC-
induced wind speed and high category storms, with heavy rainfall events and lower category
storm with larger precipitation volumes only being a secondary concern. If confirmed, this
would leave it misaligned with the country’s payout expectation, resulting in a high basis risk.
98. However, should a much-needed investigation of rainfall induced floods highlight actual losses
incurred in the years triggering as-if payouts, then the basis risk could be verified in detail. With
the current information, the product does not seem to be well matched with Tonga’s historical
losses. Finally, the high weighting of Tongatapu might be negative for the product’s basis risk if
it is found that the more frequent events in Vava’u and Ha’apai cause higher overall costs to
Tonga than rarer events directly affecting the capital region in Tongatapu.
99. Feasibility Assessment and Product Recommendation
For Tonga, a national-level product only fully pays out once and is related to the excess rainfall
event in Tongatapu in 2001, not TC Waka. A national-level product does result in an improved
ability to capturing TC induced losses and doesn’t payout frequently enough to warrant closer
investigation for Tonga.
100. CP recommends proceeding with Coverage 1, given the cost-efficient price point. The strong
divergence in payout amount should be addressed by rebalancing the weighting across the
region, lowering the share allocated to Tongatapua. A rebalanced Coverage 1 product could do
well in capturing damaging excess rainfall events impacting the different regions of the country.
41
Technical Report by CelsiusPro AG, March 2021
101. However, given the focus of the historical loss events on high wind speed TCs, the peak
precipitation events in the different regions should be assessed in detail by Tonga. Further steps
towards implementing the product should only be taken, once confirmed that such events
actually inflicted relevant damages.
Table 13: Product Options Comparison – Tonga (Indicative Premiums based on a coverage limit
of USD 5’000’000)
Regional Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Tonga Premium (USD) USD 407’787 USD 362’467 USD 286’400 USD 334’209
Tonga Premium (%) 8.2% 7.2% 5.7% 6.7%
National Level Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Tonga Premium (USD) USD 260’459 USD 217’737 USD 181’162 USD 166’600
Tonga Premium (%) 5.2% 4.4% 3.6% 3.3%
42
Technical Report by CelsiusPro AG, March 2021
3.8 Product Analysis – Cook Islands
102. As-If Analysis – Three Product Options
The Technical Report – Component 1, notes four TC events affecting the Cook Islands between
2001 to 2020. No economic loss is recorded for these storms, with the corresponding national-
level precipitation level in mm included in Figure 26. While these events do record above-
average rainfall, they remain well below thresholds required to qualify as an excess rainfall
event.
GPM IMERG Cook Islands
Precipitation in mm
80
70 TC Trina
60
50 TC Percy
TC Sarai
40
TC Pat
30
20
10
0
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
TC losses non TC losses
Figure 26: Historical Loss Event Catalog – Cook Islands, based on the Technical Report –
Component 1. National precipitation levels are included due to a lack of economic damages
information.
103. As highlighted in Table 4, the Cook Islands are split into two climatologically independent
regions in the north and south. To partially account for the administrative boundaries of the PIC,
the most populous island of Rarotonga is split off the from South region for Component 2. The
regional product structure results in a number of as-if payouts, visible in Figure 27.
Unfortunately, none of them coincide with the four TCs on record, with the as-if payout in 2010
taking place approximately 10 days after TC Pat hit Aitutaki Island in the North region.
104. For Coverages 1-3, six payout events are detectable across the 20-year period, the equivalent of
a 3.3-year return period for the product. Coverage 4 picks up only three payouts during the
same period, raising the return period to 6.6 years. The payout pattern in Figure 27 highlights
that the 2003 event exhausted the South’s limit up to almost an RP50 exhaustion point, paying
out similarly across the coverages. In 2004, an excess rainfall event affecting Rarotonga reached
close to the RP20 threshold, leading to considerably higher payouts for Coverage 1. The unusual
payout structure related to the 2008 event, stems from a severe excess rainfall event in
Ratotonga, reaching just above the RP50 threshold and thus exhausting the limits for Coverage
1,2 and 4 but not Coverage 3.
43
Technical Report by CelsiusPro AG, March 2021
Figure 27: Total as-if payouts for all regions between 2001 to 2020 – Cook Islands
105. The Cook Islands’ payout frequency curve, illustrated in Figure 28, highlights the higher payout
frequency and the two equally large payouts for Coverage 1. Unusually, the graph also notes an
equal largest payout for Coverage 1 and 4, representing the USD 3’300’000 sub-limit for
Rarotonga.
Figure 28: Payout frequency curve – Cook Islands, for regional Coverage 1 - RP 10-20 and
Coverage 4 - RP 20-50.
106. Regional Comparison
Figure 29 highlights that, despite incurring three relevant excess rainfall events over the past 20
years, the North carries little weight (4%) in the overall structure of the product. The payout
distribution across the years could be more equally distributed if the North region would be
allocated a higher share of the overall limit. The Cook Islands’ product performance remains
strongly dependent on excess rainfall events in and around the most populous island of
Rarotonga and the wider South region.
44
Technical Report by CelsiusPro AG, March 2021
Figure 29: Comparison of as-if payouts across Cook Islands’ regions - Coverage 1. Based on an
annual coverage limit of USD 5’000’000
107. Basis Risk Considerations
To conclusively comment on the basis risk of the product for Cook Islands, the historical loss
events would need to be assessed in more detail. With only four TC events, all without related
economic losses, drawing a conclusion remains challenging. The most viable way of conducting
a preliminary assessment of the product’s basis risk would be to cross-check potential flood-
induced damages on Rarotonga for the rainfall events in 2004 and 2008. There is no measure to
improve the product’s basis risk with the currently available information, neither by changing
the weighting or by changing the attachment or exhaustion points.
108. Feasibility Assessment and Product Recommendation
The national-level product for the Cook Islands contains an unusual spread of precipitation RP
thresholds across the RP10 – RP100 levels while also being lower compared to the regional
product. The wide geographic spread and climatological divide is a key reason for this finding,
hinting that the national level RPs are not representative. Consequently, a national level
product for the Cook Islands is strongly advised against.
109. For the Cook Islands, Coverage 1 is priced at a reasonable rate and appears to be the most
suitable product option, performing especially well for the 2004 event in Rarotonga. However,
further discussions on proceeding with an excess rainfall product for the Cook Islands are
subject to further investigations related to rainfall-induced flood damage. Until further
clarifications on the product’s basis risk can be gained, CP does not view this product as feasible
for the Cook Islands.
45
Technical Report by CelsiusPro AG, March 2021
Table 14: Product Options Comparison – Cook Islands (Indicative Premiums based on a coverage
limit of USD 5’000’000)
Regional Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Cook Islands Premium (USD) USD 403’440 USD 326’904 USD 277’097 USD 275’624
Cook Islands Premium (%) 8.1% 6.5% 5.5% 5.5%
National Level Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Cook Islands Premium (USD) USD 753’301 USD 649’436 USD 549’637 USD 544’281
Cook Islands Premium (%) 15.1% 13.0% 11.0% 10.9%
3.9 Product Analysis – Vanuatu
110. As-If Analysis – Three Product Options
The historical loss event catalog counts 12 incidents between 2000 and 2020, all of which are
related to named TC storm, with the exception of two heavy rainfall induced flood events in
2002 and 2009. Aside from TC Pam in 2015, none of the entries list substantial economic losses.
Figure 30: Historical Loss Event Catalog – Vanuatu, based on the Technical Report – Component
1
111. As illustrated in Figure 31, seven relatively evenly distributed as-if payouts are observable for
Vanuatu. 2004 is listed as the year with the largest as-if payout for Vanuatu, associated with TC
Ivy included in the event catalog. Additionally, the proposed product structure also triggers a
payout coinciding with TC Lusi’s in 2014 and TC Pam in 2015. The other four excess rainfall
payouts featured in Figure 31 fall within years without any recorded historical losses. The two
flood events recorded in Figure 30 in 2002 and 2009 do not coming close to the 10-year RP
threshold.
46
Technical Report by CelsiusPro AG, March 2021
112. The annual payout frequency implies a product return period of 2.9 years, which rises
considerably to 6.7 years when for Coverage 4. As can be seen in Figure 31, many of the
recorded excess rainfall events are centered around a RP 20 threshold, with Coverage 1 paying
out substantially more than Coverage 2. The relatively equal distribution of payouts in the
different years, as well the finding that payouts remain well below the aggregated coverage
limit of USD 5’000’000, can be partially explained with the more equal shares of the population
across the regions. Four out of the six regions receive between 12-19% of the sum insured,
which prevents a severe event in one region from triggering a payout close to the annual
aggregated coverage limit.
Figure 31: Total as-if payouts for all regions between 2001 to 2020 – Vanuatu
113. Vanuatu’s payout frequency curve for Coverage 1 highlights that most payouts are associated
with rainfall severities between RP10-20, resulting in well distributed payout. The Coverage 4
payout frequency curve is a consequence of only three events reaching above the RP20
threshold.
Figure 32: Payout frequency curve – Vanuatu, for regional Coverage 1 - RP 10-20 and Coverage 4
- RP 20-50.
47
Technical Report by CelsiusPro AG, March 2021
114. Regional Comparison
Figure 33 demonstrates the effect of distributing the sum insured more equally across different
regions. While the most populous region of Shefa still accounts for the largest total as-if payout,
excess rainfall events in other regions are equally able to triggers notable payouts. The rainfall
associated with TC Ivy in 2004 results in payouts across four regions, spread across three days.
Aside from this event, excess rainfall events in Vanuatu seem to only impact one or two regions.
Figure 33: Comparison of as-if payouts across Vanuatu's regions - Coverage 1. Based on an
annual coverage limit of USD 5’000’000
115. Basis Risk Considerations
The basis risk of Vanuatu’s product structure partially depends on the actual flood loss
experienced in the aftermath of TC Ivy. The substantial excess rainfall detected in February 2004
would have to be matched with large actual losses on the ground. A further crucial component
of the product’s basis risk in Vanuatu is the assumption that TC Pam losses are not substantially
higher than those caused by TC Ivy and TC Lusi. Furthermore, the two flood events 2002 and
2009, while being flagged as loss events, should be reconfirmed to not have resulted in notable
economic losses. In general, the product does not align very well with the named TCs and flood
events on record in Vanuatu. To gain more confidence regarding the basis risk of the product
more information on the historical floods in 2002 and 2009 and the impact of excess rainfall
events in Shefa during 2010 and 2015 needs to be gained.
116. Feasibility Assessment and Product Recommendation
The national-level product for Vanuatu results in a lower payout frequency and never comes
close to a full payout since 2001. As a consequence, the national-level product provides limited
advantages compared to a regional product and should not be considered going forward.
48
Technical Report by CelsiusPro AG, March 2021
117. The relatively high unused limit in Vanuatu raises some questions on whether the exhaustion
points are set too high. Nonetheless, CP does not propose lowering the product structure to,
e.g. an RP7-15 coverage, since the payout frequency and price will increase. The indicative
premium charged for Coverage 1 is already on the higher end. Since most as-if payouts range
between RP10-20, Coverage 1 still remains the recommended product option for Vanuatu.
However, further investigation are required to gain a better understanding on the historical
non-TC flood events and the damage impact of recorded excess rainfall events prior to defining
the final structure of the product.
Table 15: Product Options Comparison – Vanuatu (Indicative Premiums based on a coverage
limit of USD 5’000’000)
Regional Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Vanuatu Premium (USD) USD 586’166 USD 329’417 n/a USD 155’864
Vanuatu Premium (%) 11.7% 6.6% n/a 3.1%
National Level Product
Coverage 1 - Coverage 2 - Coverage 3 - Coverage 4 -
10/20 10/50 10/100 20/50
Vanuatu Premium (USD) USD 316’322 USD 158’830 n/a USD 78’750
Vanuatu Premium (%) 6.3% 3.2% n/a 1.6%
49
Technical Report by CelsiusPro AG, March 2021
3.10 Overall Feasibility Assessment
118. Precipitation Data
Building on the findings of the Technical Report – Component 1, a first key finding of the
feasibility study is that the improved resolution of the GPM IMERG data at ~10km x 10km (0.1
degree) grid cell size can act as the underlying data source of a parametric excess rainfall
product. The use of 3-hourly data points to obtain a rolling 24-hour aggregate offers solid data
of rainfall intensities. By factoring in a seven-day temporal window before and after each rolling
24-hour accumulation, the double counting of 3-hourly data within the product structure is
prevented. Combined with the improved resolution the GPM IMERG data does well in detecting
the overland (pluvial, flash) flooding across the PICs. Excess rainfall events primarily trigger
overland flooding due to the small basin size across the islands, resulting in a short time-lag
between rainfall and the occurrence of a flood.
119. The product analysis conducted for Component 2 also reconfirms the relatively weak link
between named TC storms and excess rainfall events flagged by GPM IMERG. Especially higher
category storms primarily cause damages due to high wind speeds. The relationship between
lower category storms and tropical depressions with the precipitation volumes recorded
warrants a closer analysis.
120. Basis Risk
The basis risk inherent to a parametric product is recognized as a key concern of such risk
transfer solutions. With the proposed remote sensing data and product structure, large steps
can be taken towards reducing the basis risk of an excess rainfall product. The project team is
confident that the feasibility study can convincingly highlight the product’s ability to detect
excess rainfall-induced floods in most PICs and act as a suitable risk transfer instrument for the
governments’ emergency relief and recovery needs.
121. The current analysis also highlights certain limitations of the product, with historical flood
events seemingly not being observed in the GPM IMERG data. Nonetheless, in most countries,
the flagged heavy rainfall flood events are relatively well aligned with the product payouts,
indicating an acceptable level of basis risk.
122. However, on a country-by-country basis, the assessment of the level of basis risk often reaches
cautious and negative conclusions. A major reason for this remains the predominant focus on
higher category TCs within the loss history of the PICs. Since the product’s focus does not lie on
addressing such high severity storms, the mismatches do not necessarily imply a high basis risk.
To reach a conclusion on this point, gaining further understanding on the relationship between
storm severity, precipitation volume and flood damages in the PICs would be required. A
frequent conclusion within the product analysis per country suggests investigating the damage
impact of recorded excess rainfall events within the particular region of the country affected.
The peak precipitation event detected in the data should be discussed in detail with the
respective countries. A lack of historical loss events in this regard leaves a basis risk assessment
challenging to conduct.
50
Technical Report by CelsiusPro AG, March 2021
123. The key appeal of the regional product structure, from a basis risk perspective, is its alignment
to the localized impacts of excess rainfall. Compared to the national-level data per PIC, this
assumption can be confirmed. The regional product structure is better at picking up heavy
rainfall induced floods recorded in the event catalog than a national-level product, implying a
lower basis risk when aiming to address medium severity and/or localized flood events.
Consequently, the project team recommends not to proceed with a produts based on national-
level data.
124. Finally, the feasibility study recommends using the population weighting as the standard
distribution of the sum insured for most of the PICs. While this approach acts as a good proxy
for the potential costs incurred by a country in the aftermath of a flood, it can strongly skew the
distribution of the product to certain regions in countries. For those countries where one region
accounts for a vast majority of the population, the project team recommends fine tuning the
regional weighting, which has the potential to reduce the basis risk. Alternatively, for these
countries, the product structure could exclusively focus on the region where the vast majority
of the population resides, targeting the excess rainfall product more precisely.
125. Product Structure – Payouts
Weather index products are a widely accepted method of transferring precipitation related risk
to international reinsurance markets. Such products are able to rapidly provide liquidity in the
aftermath of a flood, a key requirement for the PICs. The basic product functionality is
comprehensible outside of expert circles, supporting the wider distribution and endorsement of
the product across the region. By using a fixed payout amount/tick size for each mm of rainfall
above the trigger threshold and a linear progression, the payout function can be easily
understood by the client and other interested parties.
126. By basing the product on sub-national regions, localized impacts of excess rainfall events are
accounted for. Compared to structuring a product on national level, the proposed regional
structure results in more frequent payouts. A product structure attaching at a 10-year RP,
generates relatively frequent payouts across the seven Group A countries. Focusing on the
number of years between 2001-2020 which contain an as-if payout, the product’s return period
ranging between approximately 2-7 years, with an average at approximately 4 years.
127. The payout frequency is a key consideration within each PIC’s product recommendation. Most
national-level products only payout for the largest events, usually only twice across the 20-year
period. For most PICs, attaching at a 10-year RP results in an acceptable payout frequency. Only
for the Solomon Islands the project team’s recommendation targets a higher attachment point
of RP20, as offered by Coverage 4. The risk of localized payouts not resulting in substantial
annual payouts is only partially found in the case of Vanuatu. All other countries highlight at
least one annual as-if payout coming close 70% of the overall coverage limit.
128. Product Structure – Affordability
The additional product granularity and tangibility of a regional excess rainfall product contains
the drawback higher premium costs due to more frequent payouts. Keeping the affordability of
the product in mind, the project teams suggest targeting coverage options with indicative
premium rates of between 8-12% per year (net to reinsurer). Such rates imply a prudent
balance between in payout frequency/severity and premium costs.
51
Technical Report by CelsiusPro AG, March 2021
129. While Coverage 1 remains the recommendation for many PIC’s, performing best with medium
severity events, for countries such as Samoa or Fiji, Coverage 2 should be considered to reduce
the cost of the coverage. By increasing the exhaustion point, the probability of exhausting the
coverage limit is decreased and so is the associated indicative premium.
130. Reinsurance Market
A key aspect of any insurance product’s feasibility remains its acceptance by risk capacity
providers. In the case of PCRIC, the focus lies on the relationship with international reinsurers,
which need to be comfortable with any parametric excess rainfall product. The project team is
confident that reinsurance markets will accept GPM IMERG data as the data basis of the
product, including the use of 3-hourly data points to obtain a rolling 24-hour aggregate,
contrary to the market standard of using 24-hour daily values.
131. Overall Recommendation
The project team recommends to pursue a regional product, not structure based on national-
level data. Within the current regional structure, coverage options and weighting by population,
the product seem relatively well aligned with needs and historical loss events in most Group A
PIC. However, for most countries the project team recommends discussing historical rainfall
induced floods and the as-if payouts with the countries in more detail prior to a concluding
judgement of the product’s feasibility. Among the seven Group A countries, proceeding with
discussions on a risk transfer solution seems the most feasible for Samoa, the Solomon Islands,
Fiji, followed by Tonga and Vanuatu and at a later stage by RMI and the Cook Islands.
3.11 Next Steps Towards Product Development
132. Next to deriving valuable insights on the viability of introducing an excess rainfall product for
the PICs and contemplating possible product options, this report aims to provide the
groundwork to develop a marketable risk transfer instrument. A range of steps would be
required to drive forward the introduction of such a product as part of PCRIC’s offerings. The
non-exhaustive list of next steps below follows the approximate chronological order they could
be handled in:
133. Demand assessment: An extensive dialog with the Group A PICs is imperative to move forward
with the development of an excess rainfall product. Understanding their budget availability,
product option preferences, and sub-national coverage needs will allow for the product to be
refined further. The as-if payouts should be contrasted with their actual or perceived loss
experiences during the events in question. Following the demand assessment, a clear picture of
each PIC’s preliminary buy-in should be available.
134. Premium subsidies: Due to the extensive period of up to one year required to allocate donor-
funded premium subsidies, the WB and PCRIC should initiate the process as soon as the first
batch of PICs confirm their interest in the product.
52
Technical Report by CelsiusPro AG, March 2021
135. Dialog with reinsurers: The product structure can then be discussed with a selected group of
reinsurers, either by involving a parametric specialist such as CP or a reinsurance broker. This
initial dialog will allow for feedback on the product details to be received, concerns to be raised
in advance as well as some preliminary quotes and conditions to be provided by the reinsurers.
If required, the product structure and pricing assumptions can be adjusted thereafter.
136. Policy wording: In collaboration with either parametric specialist such as CP or a reinsurance
broker, a policy wording should be drafted for both the PCRIC policy issued to the PICs and the
outward reinsurance placement. At this stage, formal market consultations can be initiated and
PCRIC’s required reinsurance capacity calculated.
137. Finalize operational set-up: Introduce a lean organizational set-up with ensures the policies can
be efficiently issued. Introducing product monitoring capabilities, including but not limited to
selecting a calculation agent, will be essential to guarantee correct and swift post-event
payouts.
138. Adjust product: After the first policy year, an in-depth analysis of the product and its
functionality can be conducted. This should also include a dialog with the PIC countries to
collect information on their emergency relief expenses and other loss experiences associated
with excess rainfall events during the year, whether they triggered a payout or not. If required,
the product can be adjusted to reduce the basis risk and/or address payout or pricing concerns.
53
Technical Report by CelsiusPro AG, March 2021
4. Deliverables – Drought
4.1 Methodology and Insurance Product Options
139. Based on discussions with the World Bank project team, CelsiusPro and Risk Frontiers take a
targeted approach towards the drought product prototype (see Section 2.2). Given RMI’s
explicit interest in addressing the country’s drought risk, the focus lies on establishing a tailored
structure for these islands first. Furthermore, RMI is covered by the USA’s drought monitoring
institutions, which are seen as a possible underlying data source for a parametric product. This
study also assesses the feasibility of structuring a drought insurance coverage for Samoa and
Tonga, based on the product concept developed for RMI.
140. NOAA and the USDM
The National Oceanic and Atmospheric Administration (NOAA) is an American scientific agency
within the United States Department of Commerce that focuses on the conditions of the
oceans, major waterways, and the atmosphere. It is responsible for compiling the Pacific data
for the United States Drought Monitor (USDM) which recently began covering a number of US
Affiliated Pacific Islands (USAPI). NOAA’s work on drought severity acts as an ideal starting point
for CelsiusPro to assess alternative drought data sources. The USDM covers three PCRAFI
member countries; Marshall Islands, Federated States of Micronesia, and Palau.
141. The USDM is published on a weekly basis, see Figure 34, with the US mainland data reaching
back 20 years to 2001. The index ranks drought events on five intensity levels, ranging from D0
(abnormally dry) to D4 (exceptional drought), and captures a range of droughts (meteorological,
agricultural etc.). The categorization is based on five key indicators as well as local conditions
and impact reports from more than 450 expert observers. The area and share of the population
affected by the different drought severities is noted in the USDM. The five indicators include the
following data sources:
Palmer Drought Severity Index
CPC Soil Moisture Model
USGS Weekly Streamflow
Standardized Precipitation Index
Objective Drought Indicator Blends
54
Technical Report by CelsiusPro AG, March 2021
Figure 34: U.S. Drought Monitor Overview - October 27, 2020
142. Since the various indicators often do not coincide, the USDM’s classification tends to be
determined by the majority trend across the five, in addition to local observations. The inputs
by experts on the local observations incorporate a level of subjectivity into the USDM,
challenging the ability to use this classification as the basis for a parametric product. While the
USDM can technically be converted from a categorical to a continuous dataset and aggregated
over time, this functionality is not yet available for RMI. Moreover, historical data on RMI is only
available since 2018, limiting the ability to analyse the USDM’s historical observations of
drought conditions in the country. Combined with the subjectivity factored into the
classifications themselves, CelsiusPro concludes that the USDM cannot currently act as the basis
for a parametric drought product in the country. A position that could be reconsidered once a
longer USDM historical record on RMI exists and local observers' influence on RMI’s drought
categorizations can be assessed over time.
143. GPM IMERG
As highlighted in the Technical Report – Component 1, GPM IMERG’s precipitation can detect
meteorological droughts and frequently acts as the underlying data source for parametric risk
transfer solutions. The dataset’s transparency and available historical record allows for
historical drought to be analysed remotely and efficiently. The RMI is divided into two distinct
climatic regions, splitting the northern and southern islands (see Section 2.2). An initial drought
product and pricing structure is tested with GPM IMERG data for the aggregated cells of the
Southern islands and Northern islands. This is followed by testing these structures’ ability to
detect historic drought conditions in Samoa and Tonga’s using GPM IMERG gridded data
aggregated and averaged on a national level for both countries.
55
Technical Report by CelsiusPro AG, March 2021
4.2 Drought Structures used in the Analysis
144. Structures analysed primarily consist of Cumulative Dry Seasons, Dry Season Kick-ins and
Multiple Cumulative Dry Season which were analysed and tested for appropriateness and for
historical payouts to align with experienced droughts. Below are general structure definitions
for the covers used in the various structures analysed:
Dry Season Covers: This cover is tailored to run the course of the dry season of the region it
covers. By targeting the most critical time of the year related to rainfall and water supply
infrastructure, the coverage can detect drought conditions when they are the most acute.
Kick-in Covers: Aims to reduce the cost of the product by separately factoring in wet and
dry season conditions. The resulting dual-trigger structure defines a threshold value for the
wet season, indicating below-average aggregated precipitation over the months. A payout is
only provided if the insured suffers from lower rainfall during the following dry season as
well, with a separate trigger point defined for this period of the year.
Kick-in Cover with Contingency Payout: Aims to reduce the cost of the product by
separately factoring in wet and dry season conditions. The resulting dual-trigger structure
defines a threshold value for the wet season, indicating below-average aggregated
precipitation over the months. If there is rainfall below the trigger point the wet
season then a contingency payout will commence so the PIC can start implementing their
emergency drought plans, this is capped at 20% of the total cover to minimise premiums. A
payout is only provided if the PIC suffers from lower rainfall during the following dry season
as well, with a separate trigger point defined for this period of the year.
Multi Dry Season Covers: Individual covers that run simultaneously or independently that
can provide running payouts throughout the term of the structure as a whole.
145. All structures above include a strike value / threshold in mm of cumulative rainfall over the risk
period. Should the recorded cumulative rainfall upon expiry of the risk period be below the pre-
agreed threshold, then the product’s tick value in USD per mm is multiplied by the shortfall in
rainfall in mm below the product threshold. Payouts are then transferred to the insured within
10-20 days after the expiry of the respective risk period. Products with multiple risk periods can
trigger a payout at the expiry of each separate period.
4.3 Specific Approach to Drought for RMI, Tonga and Samoa.
146. The analysis focuses on products used to help compensate RMI, Tonga and Samoa against the
financial impact of drought, should it occur. All structures have exponential payout qualities i.e.,
as the drought gets worse the larger the payout.
147. The five specific products that have been developed for each location are as follows:
Product 1 - Dry Dry Season Multi Cover: This cover runs to the specified local dry season,
each cover is independent of the other which ensures that if the season turns mid-way then
a payout on that cover for that risk period will occur. Then if the dry season continues in
intensity the second payout occurs. This cover insures against a lack of rainfall early in the
dry season, should it occur with, and then again late in the dry season.
56
Technical Report by CelsiusPro AG, March 2021
Once the first structure expires the data will be calculated to determine if an early payout
will be issued. Once the second structure expires a second calculation will occur to
determine if a second later payment will be issued. The idea behind this structure is to
provide insurance as the dry season progresses or intensifies.
The project team recommends weighting the first risk period with 20% and the second with
80% of the sum insured.
Product 2 - Dry Seasonal Year Multi Cover: This cover runs to the specified local “Seasonal
Year”, each cover is independent of the other. The Seasonal Year is defined as one that
starts at the beginning of the wet season and finishes at the end of the dry season. For
example, if the wet season starts in May and the dry season finishes in April the seasonal
year would be from May until the following April. This cover has the first structure running
the length of the specified locales wet season and payout will occur if the insured receives
an unusually dry wet season, at the end of the wet season it will be determined whether an
early payout will occur. The second structure runs during the dry season of the specified
location, payout will occur if the insured experiences an unusually dry dry season.
The project team recommends weighting the first risk period with 20% and the second with
80% of the sum insured.
Product 3 - Dry Rolling Multi Seasonal Year Cover: This cover runs to the specified local
“Seasonal Year”, each cover is independent of the other. It consists of four structures that
all start at the beginning of the wet season, each structure is progressively longer than the
previous to insure against a progressively worsening drought. At the end of each risk period
a calculation is made as to the amount of precipitation and payout will occur if the
threshold is met. This enables compensation to be paid should drought intensify throughout
each individual risk period.
The project team recommends weighting the first and second risk periods with 10% each,
the third risk period with 20% and the fourth with 60% of the sum insured.
Product 4 - Dry Dry Season with Contingency Kick-in Cover: This cover runs to the specified
local dry season. It has two Dry Season structures with the second structure being activated
if the first is triggered. Also, if structure one is triggered a contingency payout is made. The
size of the contingency payout is determined by the severity of the drought in the first risk
period. If structure two is triggered a second payout is made, this covers a progressively
worsening drought as the payout increases as less rain falls.
The project team recommends allocating 20% of the sum insured to the contingency
payment and 80% to the second risk period.
Product 5 - Dry Seasonal Year with Contingency Kick-in Cover: This cover runs to the
specified local “Seasonal Year”. It has two structures. If structure one, which runs
concurrently with the wet season, triggers, meaning there is an unusually dry wet season,
then a contingency payout is made to mobilise relief efforts. If triggered the second
structure is kicked in. If structure two is triggered a second payout is made, this covers a
progressively worsening drought through the wet and dry season.
The project team recommends allocating 20% of the sum insured to the contingency
payment and 80% to the second risk period.
57
Technical Report by CelsiusPro AG, March 2021
148. The product analysis per country and as-if payouts illustrated in percentage are based on an
sum insured of USD 100’000. However, the resulting rate on line (RoL) can be applied to any
sum insured/policy limit required to derive the indicative premium.
149. The index per product and country is created by calculating the cumulative rainfall (daily GPM
IMERG data) for each respective risk period. The median cumulative rainfall and the standard
deviation for each risk period between 2001-2020 is then derived. Between 1-2 standard
deviations are deducted from the median cumulative rainfall to determine the trigger threshold
(strike) for each risk period included in the product structure. The tick value for each mm of
rainfall shortfall is calculated by determining the difference between the lowest cumulative
rainfall observed between 2001-2020 and the trigger threshold, which is then divided by the
sum insured allocated to the risk period. An illustrative example of the calculation steps is:
1. Assuming a median cumulative rainfall of 500mm, standard deviation of 150mm,
lowest cumulative rainfall of 250mm, and sum insured of USD 80’000.
2. Trigger threshold / strike is set at 350mm (500-150)
3. Tick value is USD 800 per mm of rainfall shortfall (80’000/(350-250)).
4. A cumulative rainfall of 300mm during the risk period would result in a payout of
USD 40’000 (800x50)
150. The varying trigger thresholds set for the different countries and risk periods are fine tuned to
capture the historical droughts on record. The different product structure represent different
facets of drought periods, with the choice of standard deviation deducted from the median
cumulative rainfall made individually for every country and each product’s risk period.
4.4 Product Analysis – RMI South
151. Within the historical GPM IMERG data of both RMI climatic zones, three major drought events
are visible in 2006-2007, 2013 and 2015-2016. This is further confirmed within the RMI Met
office and their Director Reginald White. In 2013, the sustained lack of rainfall resulted in the
RMI receiving approximately USD 5.5 million in drought relief from the US government12. A
further drought was experienced in 2015-2016. A post-disaster needs assessment estimates the
damages at USD 4.95 million. These losses were mainly attributed to production disruption due
to the drought, followed by higher costs of production due to limited freshwater availability.13
152. RMI South experiences a wet season and a dry season the wet season runs from May to
November and the dry season runs from December through to April. For the basis of this
analysis a “Seasonal Year” is one that runs from the beginning of the experienced wet season
until the end of the experienced dry season, for RMI South this is the 1st May to the following
30th April (e.g 01/05 2004 to 30/04 2005).
12
https://thediplomat.com/2016/02/drought-in-the-marshall-islands/
13
https://www.ilo.org/wcmsp5/groups/public/---ed_emp/documents/publication/wcms_553635.pdf
58
Technical Report by CelsiusPro AG, March 2021
153. Several as-if payouts are observable analyzing the GPM IMERG data. Observed losses are
accounted for in all covers. For RMI South, Product 1 - Dry Dry Season Multi Cover shows a full
payout in the 2015-2016 dry season which coincides with the locales’ most severe drought in
recent years (Figure 35).
154. Product 2 - Dry Seasonal Year Multi Cover shows partial payouts occurring during the 2008-
2009, 2015-2016 and 2017-2018 seasonal years (Figure 36). This cover coincides with the
drought experienced in 2013-2014 and 2015-2016. Product 3 - Dry Rolling Multi Seasonal Year
Cover shows payouts occurred during the 2008-2009 and 2017-2018 (Figure 37). These do not
coincide with any historic drought experienced in RMI South.
155. The payouts observed for Product 4 - Dry Dry Season with Contingency Kick-in Cover are
recorded during the 2013-2014 experienced dry season and the 2015-2016 experienced dry
season, the 2013-2014 payout is attributed to a dry experienced early dry season and the
payouts for the 2015-2016 experienced dry seasons are attributed to having a dry early
experienced dry season and a dry late experienced dry season (Figure 38). This cover coincides
with the recent experienced drought years.
156. Within Product 5 - Dry Seasonal Year with Contingency Kick-in Cover where a dry wet season
triggers a contingency payout and the second structure during the dry season partial losses are
observed during the 2008-2009 seasonal year and the 2017-2018 seasonal year, these losses
are attributed to contingency payouts for a dry wet season (Figure 39). This cover does not
coincide with any historic drought experienced within the data range for RMI South.
Product 1 - Dry Dry Season Multi Cover RMI South
2000 100%
1800 90%
1600 80%
Percentage Payout
1400 70%
Rainfall mm
1200 60%
1000 50%
800 40%
600 30%
400 20%
200 10%
0 0%
Year (Index Ending)
Payout Structure 1 01/12-12/02 Payout Structure 2 13/02-30/04
Cumulative Payout 01/12-30/04 Rainfall Structure 1 01/12-12/02
Rainfall Structure 2 13/02-30/04 Cumulative Rainfall 01/12-30/04
Figure 35: Dry Season Multi Structure RMI South Showing Rainfall and corresponding Payouts
59
Technical Report by CelsiusPro AG, March 2021
Product 2 - Dry Seasonal Year Multi Cover RMI South
4500 90%
4000 80%
Percentage Payout
3500 70%
Rainfall mm
3000 60%
2500 50%
2000 40%
1500 30%
1000 20%
500 10%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-30/11 Payout Structure 2 01/12-30/04
Cumulative Payout 01/05-30/04 Rainfall Structure 1 01/05-30/11
Rainfall Structure 2 01/12-30/04
Figure 36: As-if payouts Multi Structure Dry Wet Season/Dry Dry Season with payouts Product
Running from 01/05 to 30/11 for the wet season and 01/12 to 30/04 (index ending year) for dry
season - RMI South
Product 3 - Dry Rolling Multi Seasonal Year Cover RMI South
6000 100%
Percentage Payout
5000 80%
Rainfall mm
4000
60%
3000
40%
2000
1000 20%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-30/11 Payout Structure 2 01/05-31/01
Payout Structure 3 01/05-14/03 Payout Structure 4 01/05-30/04
Cumulative Payout 01/05-30/04 Rainfall Structure 1 01/05-30/11
Rainfall Structure 2 01/05-31/01 Rainfall Structure 3 01/05-14/03
Rainfall Structure 4 01/05-30/04
Figure 37: Multi Cover running throughout to protect progressively worse year, showing rainfall
and payouts.
60
Technical Report by CelsiusPro AG, March 2021
Product 4 - Dry Dry Season with Cont. Kick-in RMI South
2’000 100%
1’800 90%
1’600 80%
Percentage Payout
1’400 70%
Rainfall mm
1’200 60%
1’000 50%
800 40%
600 30%
400 20%
200 10%
0 0%
Year (Index Ending)
Payout Structure 1 01/12-12/02 Payout Structure 2 13/02-30/04
Cumulative Payout 01/12-30/04 Rainfall Structure 1 01/12-12/02
Rainfall Structure 2 13/02-30/04 Cumulative Rainfall 01/12-30/04
Figure 38: Dry Season Kick-in Contingency cover with payout in risk period 1, Payouts observed in
the 2013-2014 experienced dry season and the 2015-2016 experienced dry season – RMI South
Product 5 - Dry Seasonal Year with Cont. Kick-in RMI South
6’000 25%
5’000 20%
Percentage Payout
4’000
Rainfall mm
15%
3’000
10%
2’000
1’000 5%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-30/11 Payout Structure 2 01/12-30/04
Cumulative Payout 01/05-30/04 Rainfall Structure 1 01/05-30/11
Rainfall Structure 2 01/12-30/04 Cumulative Rainfall 01/05-30/04
Figure 39: As-if payouts Kick-in with Contingency Payout risk period 1 Wet Season triggers dry
season structure. Rainfall represented but the lines - RMI South
61
Technical Report by CelsiusPro AG, March 2021
Table 16: RMI South Pricing per Product Multi Structures
Type of Risk Risk Risk Risk Strike 1 Strike 2 Strike 3 Strike 4 ROL
Cover Period period period period
1 2 3 4
Product 01/12- 13/02- N/A N/A 438mm 344mm N/A N/A 16.4%
1 - Dry 12/02 30/04
Dry
Season
Multi
Cover
Product 01/05- 01/12- N/A N/A 2624mm 634mm N/A N/A 15.7%
2 - Dry 30/11 30/04
Seasonal
Year
Multi
Cover
Product 01/05- 01/05- 01/05- 01/05- 2624mm 2937mm 3184mm 3470mm 6.8%
3 - Dry 30/11 01/01 14/03 30/04
Rolling
Multi
Seasonal
Year
Cover
Table 17: RMI South Pricing per Product Kick-in Structures
Type of Cover Risk Risk Trigger Strike Exit ROL
Period period
1 2
Product 4 - Dry Dry 01/12- 13/02- 350mm (Up to 20% 500mm 166mm 11%
Season with 12/02 30/04 cover payout)
Contingency Kick-in
Cover
Product 5 - Dry 01/05- 01/12- 2624mm (Up to 955mm 470mm 5.3%
Seasonal Year with 30/11 30/04 20% cover payout)
Contingency Kick-in
Cover
157. Basis Risk
The experienced droughts in RMI South during the years we have access to data to, 2001 (20
years) were 2006-2007, 2013 and 2015-2016. Using specific bespoke drought products for the
RMI South received remuneration in the most severe year of recent drought 2015-2016. As
2015-2016 has been highlighted as the year of most risk it is recommended that this be used as
a basis for drought protection.
62
Technical Report by CelsiusPro AG, March 2021
158. Product 1 - Dry Dry Season Multi Cover has payouts in drought years (Partial Payout 2008-2009,
2013-2014 and Payout full 2015-2016). For this cover the insured would receive an early
payment for all three payout years are structure 1 is triggered in all cases. For the full payout
year (2015-2016) the insured would have received an early payout at the end of the first
structure as well as the remaining sum at the end of the second structure.
159. For the Product 2 - Dry Seasonal Year Multi Cover that runs the first structure during the wet
season and the second structure during the dry season there would have been an early partial
payment for the 2008-2009 and 2017-2018 seasonal years due to having a dry wet season. For
the experienced dry dry season in structure 2 of this cover the insured would have received a
payout for structure 2 at the expiry of the cover term, this coincides with the drought
experienced during the dry season of RMI South of 2015-2016.
160. The Product 3 - Dry Rolling Multi Seasonal Year Cover that accumulates rainfall as the seasonal
year progresses observed full payout in the 2008-2009 seasonal year.
161. Product 4 - Dry Dry Season with Contingency Kick-in Cover (Partial payout 2013-2014, Full
payout 2015-2016). In this overall structure as the First structure triggers the second structure
there would have been an early partial payout in the observed payout years. A full payout
would have been received (2015-2016) the insured during the 2015-2016 season, with the
majority of the payment being received at the end of the second structure.
162. The Product 5 - Dry Seasonal Year with Contingency Kick-in Cover showed partial payouts in
2008-2009 and 2017-2018 seasonal year.
163. Recommendation
Potential product options for RMI South include Product 1 - Dry Dry Season Multi, Product 2 -
Dry Seasonal Year Multi and Product – 4 Dry Dry Season Kickin. The project team recommends
focusing on Product 1 - Dry Dry Season Multi Cover for RMI South, due to each structure within
the cover being independent and having observed payouts in drought years.
164. The ROL of 16.4% (see Table 16) can be reduced by extending the cover length to
unprecedented rainfall levels, this will reduce the ROL by 3-7% but it is ultimately up to the risk
taker and how they price unprecedented risk. Note that the pricing calculator has a function to
change the strikes in order to arrive at suitable cover that meets the budget.
63
Technical Report by CelsiusPro AG, March 2021
4.5 Product Analysis - RMI North
165. RMI North has the same seasons as RMI South and experienced the same years of drought,
namely 2006-2007, 2013 and 2015-2016. RMI North experiences a wet season and a dry season
the wet season runs from May to November and the Dry Season runs from December through
to April. For the basis of this analysis a “Seasonal year” is on that runs from the beginning of the
experienced wet season until the end of the experienced dry season, for RMI North this is the
1st May to the following 30th April (e.g 01/05 2004 to 30/04 2005).
166. Several as-if payouts are observable analyzing the GPM IMERG data. Observed losses are
accounted for in all covers. For RMI North the Product 1 - Dry Dry Season Multi Cover shows a
full payout in the 2015-2016 dry season which coincides with the locales’ most severe drought
in recent years (Figure 40). This cover also shows partial payment during the 2007-2008 and
2008-2009 dry seasons.
167. The Product 2 - Dry Seasonal Year Multi Cover shows partial payouts occurring during the 2007-
2008, 2008-2009, 2012-2013, 2016-2017 and 2017-2018 seasonal years (Figure 41). This covers
payouts do not correlate strongly with periods of experienced drought. The Product 3 - Dry
Rolling Multi Seasonal Year Cover Shows partial payouts occurred 2007-2008, 2008-2009 and
2017-2018. However, the 2007-2008 and the 2017-2018 payouts are negligible and would not
have incurred cost recovery (Figure 42). This covers payouts do not correlate strongly with
periods of experienced drought.
168. Regarding the Product 4 - Dry Dry Season with Contingency Kick-in Cover partial losses are
observed in the 2007-2008, 2008-2009, 2011-2012, 2015-2016 and 2017-2018 experienced dry
season. All payouts bar the 2015-2016 are attributed to the contingency payout in risk period 1.
The 2015-2016 cover received a near full payout (Figure 43). This cover has a large payout that
coincides with the drought experienced in 2015-2016.
169. Within the Product 5 - Dry Seasonal Year with Contingency Kick-in Cover where a dry wet
season triggers a contingency payout and the second structure during the dry season partial
losses are observed during the 2007-2008, 2008-2009, 2012-2013 and 2017-2018 seasonal
years, these losses are attributed to contingency payouts for a dry wet season as well as a Dry
Dry season payout for the 2007-2008 experienced dry season (Figure 44). Which does not
coincide with any historic drought during the data period.
64
Technical Report by CelsiusPro AG, March 2021
Product 1 - Dry Dry Season Multi Cover RMI North
800 100%
700 90%
80%
Percentage Payout
600
70%
Rainfall mm
500 60%
400 50%
300 40%
30%
200
20%
100 10%
0 0%
Year (Index Ending)
Payout Structure 1 01/12-12/02 Payout Structure 2 13/02-30/04
Cumulative Rainfall 01/12-30/04 Rainfall Structure 1 01/12-12/02
Rainfall Structure 2 13/02-30/04 Cumulative Payout 01/12-30/04
Figure 40: Dry Season Multi Structure RMI North Showing Rainfall and corresponding Payouts
Product 2 - Dry Seasonal Year Multi Cover RMI North
3500 90%
3000 80%
Percentage Payout
70%
2500
Rainfall mm
60%
2000 50%
1500 40%
30%
1000
20%
500 10%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-30/11 Payout Structure 2 01/12-30/04
Cumulative Payout 01/05-30/04 Rainfall Structure 1 01/05-30/11
Rainfall Structure 2 01/12-30/04
Figure 41: As-if payouts Multi Structure Dry Wet Season/Dry Dry Season with payouts Product
Running from 01/05 to 30/11 for the wet season and 01/12 to 30/04(index ending year) for dry
season - RMI North
65
Technical Report by CelsiusPro AG, March 2021
Product 3 - Dry Rolling Multi Seasonal Year Cover RMI North
3500 100%
Payout Percentage
3000 80%
Rainfall mm
2500
2000 60%
1500 40%
1000
500 20%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-30/11 Payout Structure 2 01/05-31/01
Payout Structure 3 01/05-14/03 Payout Structure 4 01/05-30/04
Cumulative Payout 01/05-30/04 Rainfall Structure 1 01/05-30/11
Rainfall Structure 2 01/05-31/01 Rainfall Structure 3 01/05-14/03
Rainfall Structure 4 01/05-30/04
Figure 42: Multi Cover running throughout to protect progressively worse year, showing rainfall
and payouts – RMI North
Product 4 - Dry Dry Season with Cont. Kick-in RMI North
800 120%
700
100%
Percentage Payout
600
Rainfall mm
80%
500
400 60%
300
40%
200
20%
100
0 0%
Year (Index Ending)
Payout Structure 1 01/12-12/02 Payout Structure 2 13/02-30/04
Cumulative Payout 01/12-30/04 Rainfall Structure 1 01/12-12/02
Rainfall Structure 2 13/02-30/04 Cumulative Rainfall 01/12-30/04
Figure 43: Dry Season Kick-in Contingency cover with payout in risk period 1, Payouts observed in
the 2007-2008, 2008-2009, 2011-2012, 2015-2016 and 2016-2017 experienced dry season – RMI
North
66
Technical Report by CelsiusPro AG, March 2021
Product 5 - Dry Seasonal Year with Cont. Kick-in RMI North
3’500 25%
3’000
20%
Percentage Payout
2’500
Rainfall mm
2’000 15%
1’500 10%
1’000
5%
500
0 0%
Year (Index Ending)
Payout Structure 1 01/05-30/11 Payout Structure 2 01/12-30/04
Cumulative Payout 01/05-30/04 Rainfall Structure 1 01/05-30/11
Rainfall Structure 2 01/12-30/04 Cumulative Rainfall 01/05-30/04
Figure 44: As-if payouts Kick-in with Contingency Payout risk period 1 Wet Season triggers dry
season structure. Rainfall represented but the lines - RMI North
Table 18: RMI North Pricing per Product Multi Structures
Type of Risk Risk Risk Risk Strike 1 Strike 2 Strike 3 Strike 4 ROL
Cover Period period period period
1 2 3 4
Product 01/12- 13/02- N/A N/A 64mm 49mm N/A N/A 15.9%
1 - Dry 12/02 30/04
Dry
Season
Multi
Cover
Product 01/05- 01/05- 01/05- 01/05- 1430mm 1237mm 1095mm 1460mm 5.8%
2 - Dry 30/11 01/01 14/03 30/04
Seasonal
Year
Multi
Cover
Product 01/05- 01/12- N/A N/A 1430mm 115mm N/A N/A 14.5%
3 - Dry 30/11 30/04
Rolling
Multi
Seasonal
Year
Cover
67
Technical Report by CelsiusPro AG, March 2021
Table 19: RMI North Pricing per Product Kick-in Structures
Type of Cover Risk Risk Trigger Strike Exit ROL
Period period
1 2
Product 4 - Dry Dry Season 01/12- 13/02- 71mm 49mm 44mm 12.8%
with Contingency Kick-in 12/02 30/04
Cover
Product 5 - Dry Seasonal 01/05- 01/12- 1431mm 115mm 102mm 4.7%
Year with Contingency Kick- 30/11 30/04
in Cover
170. Basis Risk
The experienced droughts in RMI North during the years we have access to data to, 2001 (20
years) were 2006-2007, 2013 and 2015-2016. As 2015-2016 has been highlighted as the year of
most risk as it was the worst experienced drought period it is recommended that this be used as
a basis for drought protection. For RMI North Product 1 - Dry Dry Season Multi Cover payout
has observed payouts for the drought year of 2007-2008, for this cover the insured would have
received a late payout for a dry second half of the dry season. However there is a full payout
observed in the 2016-2017 dry season, this is a concern as this is a non-drought year and will
increase ROL.
171. The Product 2 - Dry Seasonal Year Multi Cover shows wet season and dry season payout in year
running from 2007-2008, the insured would have received partial payment after the first
structure and partial payment after the second struture. However as there are payouts are
observed in 2008-2009 and 2016-2017 non-drought years the ROL would increase due to
payments made during non-drought years. As these structures have payouts in experienced
year of drought these products are recommended.
172. However due to large payouts in non-drought years namely 2016-2017 for the Product 1 - Dry
Dry Season Multi Cover, Product 4 - Dry Dry Season with Contingency Kick-in Coverand Product
2 - Dry Seasonal Year Multi Cover and in 2008-2009 for the Product 5 - Dry Seasonal Year with
Contingency Kick-in Cover, it increases the ROL by having payouts in non-drought years. Product
3 - Dry Rolling Multi Seasonal Year Cover that accumulates rainfall as the seasonal year
progresses observed full payout in the 2008-2009 seasonal year.
173. Product 4 - Dry Dry Season with Contingency Kick-in Cover observed partial payout during the
2007-2008 seasonal year and this coincides with observed drought. The insured would have
received a contingency payout at the end of the first structure and another payout at the end of
the second structure. However as there is a large payout observed during 2016-2017 dry season
and this does not coincide with a historic drought year this will increase the ROL. The Product 5
- Dry Seasonal Year with Contingency Kick-in Cover for RMI North observes payouts for the
2007-2008 drought year. The insured would have received a partial contingency payout and a
partial payout at the end of the cover. However, as payouts are observed during the 2008-2009
seasonal year and this is a year of non-experienced drought the ROL would increase for payouts
on non-drought periods.
68
Technical Report by CelsiusPro AG, March 2021
174. Testing the five product options, using multiple different structures, based on RMI South
and RMI North historical GPM IMERG data results in the various pricing range and trigger
thresholds, which are geared towards ROL with a minimum of 5% and a maximum of 16%. The
triggers are determined based on the years of loss, available data, historical losses, seasonality
and the inherent strain on water resources during the dry season.
175. Recommendation
For RMI North there isn't really an ideal cover as there are a lot of payouts in years of non-
drought. Out of all the product options, Product 5 - Dry Seasonal year kick-in as this has highest
payout in years of drought and less payouts in years of non-drought. However, this structure is
still not recommended given the divergence between experienced drought and the GPM IMERG
data. Given the concentration of RMI’s population in the South, the project team recommends
focusing any drought cover to support the government’s drought emergency relief on this
regions, not factoring the North region within the product structure.
4.6 Product Analysis - Samoa
176. Samoa’s rain season takes place from November to April and the dry season from May to
October. Widespread water shortages from the El Niño related drought/dry periods of 1982-
1983, 1997-1998, 2001-2002 and 2002-200314, Samoa has also experienced drought conditions
in 201115 and 201416. For the basis of this analysis a “Seasonal year” is on that runs from the
beginning of the experienced wet season until the end of the experienced dry season. For
Samoa this is the 1st November to the following 31st October (e.g. 01/11 2004 to 31/10 2005).
177. Several as-if payouts are observable when analysing the GPM IMERG data, reflecting the actual
drought experienced in Samoa. Observed losses are accounted for in all covers. For Samoa,
Product 1 - Dry Dry Season Multi Cover shows partial payout during the 2011 and 2015 dry
seasons (Figure 45). The partial payment in 2011 coincides with historic experienced drought in
Samoa.
178. Product 2 - Dry Seasonal Year Multi Cover shows partial payouts occurring during 2010-2011,
2014-2015 and 2018-2019 seasonal years (Figure 46). This coincides with the drought
experienced in Samoa during 2011 and 2015. Product 3 - Dry Rolling Multi Seasonal Year Cover
shows partial payouts occurred 2004-2005, 2010-2011, 2014-2015 and 2018-2019 (Figure 47).
This coincides the drought experienced in Samoa in 2011 and 2015.
179. Regarding Product 4 - Dry Dry Season with Contingency Kick-in Cover partial losses are observed
in the 2011 experienced dry season. This payout is attributed to the contingency payout in RP1.
(Figure 48). This partial payment coincides with the drought experienced in 2011.
14
https://climateknowledgeportal.worldbank.org/country/samoa/vulnerability
15
https://reliefweb.int/disaster/dr-2011-000146-tuv
16
https://reliefweb.int/report/samoa/declaration-meteorological-drought
69
Technical Report by CelsiusPro AG, March 2021
180. Within Product 5 - Dry Seasonal Year with Contingency Kick-in Cover where a dry wet season
triggers a contingency payout and the second structure during the dry season partial losses are
observed during the 2004-2005 and 2018-2019 seasonal years, these losses are attributed to
contingency payouts for a dry wet season (Figure 49). This cover does not coincide with any
experienced drought in Samoa.
Product 1 - Dry Dry Season Multi Cover Samoa
2500 90%
80%
2000
Percentage Payout
70%
Rainfall mm
60%
1500
50%
40%
1000
30%
500 20%
10%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-31/07 Payout Structure 2 01/08-31/10
Cumulative Payout 01/05-31/10 Rainfall Structure 1 01/05 -31/07
Rainfall Structure 2 01/08-31/10 Cumulative Rainfall 01/05-31/10
Figure 45: Dry Season Multi Structure Samoa Showing Rainfall and corresponding Payouts
Product 2 - Dry Seasonal Year Multi Cover Samoa
2500 90%
80%
2000 70% Percentage Payout
Rainfall mm
60%
1500
50%
40%
1000
30%
500 20%
10%
0 0%
Year (Index Ending)
Payout Structure 1 01/11-30/04 Payout Structure 2 01/05-31/10
Cumulative Payout 01/11-31/10 Rainfall Structure 1 01/11-30/04
Rainfall Structure 2 01/05-31/10
Figure 46: As-if payouts Multi Structure Dry Wet Season/Dry Dry Season with payouts Product
Running from 01/11 to 30/04 for the wet season and 01/05 to 31/10 (index ending year) for dry
season – Samoa
70
Technical Report by CelsiusPro AG, March 2021
Product 3 - Dry Rolling Multi Seasonal Year Cover Samoa
5000 100%
Percentage Payout
4000 80%
Rainfall mm
3000 60%
2000 40%
1000 20%
0 0%
Year (Index Ending)
Payout Structure 1 01/11-30/04 Payout Structure 2 01/11-30/06
Payout Stucture 3 01/11-31/08 Payout Structure 4 01/11-31/10
Cumulative Payout 01/11-31/10 Rainfall Structure 1 01/11-30/04
Rainfall Structure 2 01/11-30/06 Rainfall Structure 3 01/11-31/08
Rainfall structure 4 01/11-31/10
Figure 47: Multi Cover running throughout year to protect progressively worsening drought year,
showing rainfall and payouts – Samoa
Product 4 - Dry Dry Season with Contingency Kick-in Samoa
2’500 25%
2’000 20%
Percentage Payout
Rainfall mm
1’500 15%
1’000 10%
500 5%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-31/07 Payout Structure 2 01/08-31/10
Cumulative Payout 01/05-31/10 Rainfall Structure 1 01/05-31/07
Rainfall Structure 2 01/08-31/10 Cumulative Rainfall 01/05-31/10
Figure 48: Dry Season Kick-in Contingency cover with payout in risk period 1, Payouts observed in
the 2011 dry season – Samoa
71
Technical Report by CelsiusPro AG, March 2021
Product 5 - Dry Seasonal Year with Contingency Kick-in Samoa
4’500 25%
4’000
20%
Percentage Payout
3’500
Rainfall mm
3’000
15%
2’500
2’000
10%
1’500
1’000 5%
500
0 0%
Year (Index Ending)
Payout Structure 1 01/11-30/04 Payout Structure 2 01/05-31/10
Cumulative Payout 01/11-31/10 Rainfall Structure 1 01/11-30/04
Rainfall Structure 2 01/05-31/10 Cumulative Rainfall 01/11-31/10
Figure 49: As-if payouts Kick-in with Contingency Payout risk period 1 Wet Season triggers dry
season structure. Rainfall represented but the lines - Samoa
Table 20: Samoa Pricing per Product Multi Structures
Type of Risk Risk Risk Risk Strike 1 Strike 2 Strike 3 Strike 4 ROL
Cover Period period period period
1 2 3 4
Product 1 01/05- 01/08- N/A N/A 216mm 209mm N/A N/A 8.7%
- Dry Dry 31/07 31/10
Season
Multi
Cover
Product 2 01/11- 01/11- 01/11- 01/11- 1514mm 1691mm 1828mm 2205mm 11.1
- Dry 30/04 30/06 31/08 31/10 %
Seasonal
Year
Multi
Cover
Product 3 01/11- 01/05- N/A N/A 1402mm 573mm N/A N/A 14.5%
- Dry 30/04 31/10
Rolling
Multi
Seasonal
Year
Cover
72
Technical Report by CelsiusPro AG, March 2021
Table 21: Samoa Pricing per Product Kick-in Structures
Type of Cover Risk Risk Trigger Strike Exit ROL
Period 1 period 2
Product 4 - Dry Dry Season 01/05- 01/08- 216mm 209mm 127mm 4.4%
with Contingency Kick-in 31/07 31/10
Cover
Product 5 - Dry Seasonal Year 01/11- 01/05- 1514mm 573mm 378mm 4.8%
with Contingency Kick-in 30/04 31/10
Cover
181. Basis Risk
The basis risk for Samoa is hard to assess and extremely limited as there have been no recorded
economic losses due to drought in Samoa. However, each product relating to Samoa does
capture some year(s) of loss when using the GPM IMERG dataset. Actual damages would need
to be recorded to assess whether they concur with Samoa's current as-if analysis. Actual loss
needs to be recorded and aligned with the products proposed to lower Samoa's basis risk.
182. The experienced droughts in Samoa during the years we have access to data to, were 2001,
2003, 2011 and 2014. For Samoa Product 1 - Dry Dry Season Multi Cover payout observed a
partial payout during the 2011 dry season, this is attributed to a payout during structure 1.
However, a payout is also observed during the 2015 dry season which isn’t reported to be a
year of drought.
183. The Product 2 - Dry Seasonal Year Multi Cover shows payouts 2010-2011 seasonal year where
drought was observed. This is attributed to having a dry dry season (RP2). There is payouts
during the 2014-2015 seasonal year for this cover but this is attributed to a dry dry season
during 2015 where no drought wasn’t experienced. The Product 3 - Dry Rolling Multi Seasonal
Year Cover that accumulates rainfall as the seasonal year progresses observed payouts in the
2010-2011 seasonal year, payouts are also observed during the 2004-2005 and 2014-2016
seasonal years. The payouts for the 2010-2011 seasonal year are attributed to a worsening
drought throughout the seasonal year. The payouts for the 2004-2005 and 2014-2015 seasonal
years do not coincide with experienced drought in Samoa.
184. Product 4 - Dry Dry Season with Contingency Kick-in Cover for Samoa observes partial payout
2011. This is due to the contingency payout as part of structure one of the cover. The Product 5
- Dry Seasonal Year with Contingency Kick-in Cover shows payouts in years of non-experienced
droughts. As the observed payouts for this structure does not align with observed drought his
product is not recommended.
185. Testing the five product options, using multiple different structures, based on Samoa
historical GPM IMERG data results in the various pricing range and trigger thresholds, which are
geared towards ROL with a minimum of 4.4% and a maximum of 14.5%. The triggers
are determined based on the years of loss, available data, historical losses, seasonality and the
inherent strain on water resources during the dry season.
73
Technical Report by CelsiusPro AG, March 2021
186. Recommendation
Possible product options for Samoa are Product 2 - Dry Seasonal Year Multi and Product 3 - Dry
Rolling Seasonal Year Multi. Among these two options, the project team recommends Product 3
- Dry Rolling Seasonal Year Multi cover due to each structure within the cover being
independent and have observed payouts in drought years. The ROL of 14.5% (see Table 20) can
be reduced by extending the cover length to unprecedented rainfall levels, this will reduce the
ROL by 2-4% but it is ultimately up to the risk taker and how they price unprecedented risk.
Note that the pricing calculator has a function to change the strikes in order to arrive at suitable
cover that meets the budget.
4.7 Product Analysis - Tonga
187. Tonga has a wet season from December to April and a dry season from May to November. The
last three major droughts that have occurred in Tonga in 1983, 1998, and 2006 have been
directly linked to the May 1982–June 1983, May 1997–April 1998, and September 2006–January
2007 El Niño events17. Tonga has also experienced drought during 2015-2016 and 201818.
188. For the basis of this analysis a “Seasonal year” is one that runs from the beginning of the
experienced wet season until the end of the experienced dry season, for Tonga this is the 1st
December to the following 30th November (e.g 01/12 2004 to 30/11 2005). Several as-if payouts
are observable analyzing the GPM IMERG data. Observed losses are accounted for in all covers.
189. For Tonga the Product 1 - Dry Dry Season Multi Cover experiences payouts during the 2010,
2014, 2015 and 2018 dry seasons (Figure 50). This coincides with droughts experienced in
Tonga. The Product 2 - Dry Seasonal Year Multi Cover shows partial payouts occurring during
2013-2014, 2014-2015 and 2015-2016 seasonal years (Figure 51). This cover coincides with
drought experienced in drought. The Product 3 - Dry Rolling Multi Seasonal Year Cover shows
payouts occurred 2002-2003, 2014-2015, 2016-2017 and 2019-2020(Figure 52). This cover
coincides with drought experienced in Tonga.
190. Regarding the Product 4 - Dry Dry Season with Contingency Kick-in Cover partial losses are
observed in the 2010,2014, 2015 and 2018 experienced dry season (Figure 53). This cover does
not coincide with experienced drought in Tonga. Within the Product 5 - Dry Seasonal Year with
Contingency Kick-in Cover where a dry wet season triggers a contingency payout and the
second structure during the dry season partial losses are observed during the 2014-2015, 2015-
2016 and 2019-2020 seasonal years (Figure 54). This coincides with experienced drought in
Tonga.
17
https://climateknowledgeportal.worldbank.org/country/tonga/vulnerability
18
https://reliefweb.int/report/tonga/el-ni-o-and-drought-watch-tonga-ento
74
Technical Report by CelsiusPro AG, March 2021
Product 1 - Dry Dry Season Multi Cover Tonga
2500 100%
13-Feb 12-Feb
90%
2000 80%
Percentage Payout
70%
1500 60%
30-Apr
50%
1000 40%
30%
500 20%
Rainfall mm1-Dec
Period
10%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-15/08 Payout Structure 2 16/08-30/11
Cumulative Payout 01/05-30/11 Rainfall Structure 1 01/04-15/08
Rainfall Structure 2 16/08-30/11 Cumulative Rainfall 01/05-30/11
Figure 50: Dry Season Multi Structure Tonga Showing Rainfall and corresponding Payouts
Product 2 - Dry Seasonal Year Multi Cover Tonga
2500 90%
80%
2000
Percentage Payout
70%
60%
Rainfall mm
1500
50%
40%
1000
30%
500 20%
10%
0 0%
Year (Index Ending)
Payout Structure 1 01/12-30/04 Payout Structure 2 01/05-30/11
Cumulative Payout 01/12-30/11 Rainfall Structure 1 01/12-30/04
Rainfall Structure 2 01/05-30/11
Figure 51: As-if payouts Multi Structure Dry Wet Season/Dry Dry Season with payouts Product
Running from 01/12 to 30/04 for the wet season and 01/05 to 30/11 (index ending year) for dry
season – Tonga
75
Technical Report by CelsiusPro AG, March 2021
Product 3 - Dry Rolling Multi Seasonal Year Cover Tonga
3500 100%
Percentage Payout
3000 80%
Rainfall mm
2500
2000 60%
1500 40%
1000
500 20%
0 0%
Year (Index Ending)
Payout Structure 1 01/12-30/04 Payout Structure 2 01/12-31/07
Payout Structure 3 01/12-30/09 Payout Structure 4 01/12-30/11
Cumulative Payout 01/12-30/11 Rainfall Structure 1 01/12-30/04
Rainfall Structure 2 01/12-31/07 Rainfall Structure 3 01/12-30/09
Rainfall Structure 4 01/12-30/11
Figure 52: Multi Cover running throughout year to protect progressively worsening drought year,
showing rainfall and payouts – Tonga
Product 4 - Dry Dry Season with Contingency Kick-in Tonga
2’500 100%
90%
2’000 80%
Percentage Payout
70%
Rainfall mm
1’500 60%
50%
1’000 40%
30%
500 20%
10%
0 0%
Year (Index Ending)
Payout Structure 1 01/05-15/08 Rainfall Structure 2 16/08-30/11
Cumulative Payout 01/05-30/11 Rainfall Structure 1 01/05-15/08
Rainfall Structure 2 16/08-30/11 Cumulative Rainfall 01/05-30/11
Figure 53: Dry Season Kick-in Contingency cover with payout in risk period 1, Payouts observed in
the 2010, 2014, 2015 and 2018 dry season – Tonga
76
Technical Report by CelsiusPro AG, March 2021
Product 5 - Dry Seasonal Year with Contingency Kick-in Tonga
3’500 90%
3’000 80%
Percentage Payout
70%
2’500
Rainfall mm
60%
2’000 50%
1’500 40%
30%
1’000
20%
500 10%
0 0%
Year (index Ending)
Payout Structure 1 01/12-30/04 Payout Structure 2 01/05-30/11
Cumulative Payout 01/12-30/11 Rainfall Structure 1 01/12-30/04
Rainfall Structure 2 01/05-30/11 Cumulative Rainfall 01/12-30/11
Figure 54: As-if payouts Kick-in with Contingency Payout risk period 1 Wet Season triggers dry
season structure. Rainfall represented but the lines - Tonga
Table 22: Tonga Pricing per Product Multi Structures
Type of Risk Risk Risk Risk Strike Strike 2 Strike 3 Strike 4 ROL
Cover Period period period period 1
1 2 3 4
Product 01/05- 16/08- N/A N/A 308mm 246mm N/A N/A 9.7%
1 - Dry 15/08 30/11
Dry
Season
Multi
Cover
Product 01/12- 01/12- 01/12- 01/12- 968mm 1333mm 1483mm 1885mm 10.7%
2 - Dry 30/04 31/07 30/09 30/11
Seasonal
Year
Multi
Cover
Product 01/12- 01/05- N/A N/A 748mm 686mm N/A N/A 13.9%
3 - Dry 30/04 30/11
Rolling
Multi
Seasonal
Year
Cover
77
Technical Report by CelsiusPro AG, March 2021
Table 23: Tonga Pricing per Product Kick-in Structures
Type of Cover Risk Risk Trigger Strike Exit ROL
Period 1 period 2
Product 4 - Dry Dry Season 01/05- 16/08- 308mm 246mm 214mm 10.6%
with Contingency Kick-in 15/08 30/11
Cover
Product 5 - Dry Seasonal 01/12- 01/05- 968mm 686mm 499mm 11.3%
Year with Contingency Kick- 30/04 30/11
in Cover
191. Basis Risk
The basis risk for Tonga is hard to assess and extremely limited as there has been no recorded
economic losses due to drought in Tonga. However, each product relating to Tonga does
capture some year(s) of loss when using the GPM IMERG dataset. Actual damages would need
to be recorded to assess whether they concur with Tonga's current as-if analysis. Actual loss
needs to be recorded and aligned with the products proposed to lower Tonga's basis risk.
192. The experienced droughts in Tonga during the years we have access to data to, 2001 (20 years)
were 2006, 2015-2016 and 2018. The Product 1 - Dry Dry Season Multi Cover and Product 4 -
Dry Dry Season with Contingency Kick-in Cover shows payouts in years of non-experienced
droughts.
193. The Product 2 - Dry Seasonal Year Multi Cover shows payouts 2013-2014, 2014-2015 and 2015-
2016 seasonal year where drought was observed. The payouts observed are attributed to a dry
dry season for both the 2013-2014 and 2014-2015 seasonal year. The payout observed during
the 2015-2016 is attributed to a dry wet season and relates back to a continuing drought from
the 2015 drought during the 2015 dry season. The Product 3 - Dry Rolling Multi Seasonal Year
Cover that accumulates rainfall as the seasonal year progresses observed payouts in the 2014-
2015 seasonal year. There is a total payout observed for this cover and it is attributed to a
progressively worsening drought in Tonga for the 2014-2015 seasonal year. As these structures
have payouts in experienced year of drought these products are recommended.
194. The Product 5 - Dry Seasonal Year with Contingency Kick-in Cover for Tonga shows payouts in
2014-2015 and 2015-2016. The insured would have received a contingency payment for both
2014-2015 and 2015-2016 seasonal years, the insured would have also received a payment for
the triggered structure 2 for the 2014-2015 seasonal year, this would be attributed to a dry dry
season.
195. Testing the five product options, using multiple different structures, based on Tonga
historical GPM IMERG data results in the various pricing range and trigger thresholds, which are
geared towards ROL with a minimum of 9% and a maximum of 13.9%. The triggers
are determined based on the years of loss, available data, historical losses, seasonality and the
inherent strain on water resources during the dry season.
78
Technical Report by CelsiusPro AG, March 2021
196. Recommendation
Possible product options for Tonga are Product 3 - Dry Rolling Seasonal Year Multi and Product
5 - Dry Seasonal Year Kick-in with contingency payout. Among these two options the project
team recommends focusing on Product 3 - Dry Rolling Seasonal Year Multi cover due to each
structure within the cover being independent and have observed payouts in drought years. The
ROL of 13.9% (see Table 22) can be reduced by extending the cover length to unprecedented
rainfall levels, this will reduce the ROL by 2-4% but it is ultimately up to the risk taker and how
they price unprecedented risk. Note that the pricing calculator has a function to change the
strikes in order to arrive at suitable cover that meets the budget.
4.8 Overall Feasibility Assessment
197. A GPM IMERG based parametric drought insurance product is deemed to be possible, aside
from RMI North, given the correlation between a lack of rainfall and historical drought events
across climatic zones of the countries in question. Further consultation with countries will be
required to determine the most suitable trigger points and product structure. Moreover, a
confirmation of the definition of wet and dry seasons is needed. Reflecting different wet and
dry season triggers helps tailor the product to respond to severe droughts, which generally
occur when below average rainfall in both wet and dry seasons coincide.
198. As each country has different rainfall patterns and seasons a “one size fits all” approach will not
work when determining a cover for each country. Covers where payouts occur that coincide
with experienced drought are recommended for each individual country or region. The
approach taken with drought contrasts that of an excess rain event as drought is a slow moving
catastrophe. Different covers suit different localities with product structures and
recommendations adjusted depending on their alignment with historic droughts.
199. Across the five product options analysed, the primary recommendation for RMI South focuses
on Product 1 - Dry Dry Season Multi Cover while for Samoa and Tonga Product 3 - Dry Rolling
Multi Seasonal Year Cover is targeted. Whenever possible, the project team recommends
focusing on structures with independent covers for the different risk periods, ensuring that
payout calculations can be assessed at the end of each respective period. Such structures
ensure multiple payouts will be received by the country, increasing in size as the drought
increases in severity over the risk periods. It is not advised to pursue a specific product
recommendation for RMI North, given the mismatch with historical droughts.
200. Secondary product options noted across the countries reviewed include Product 2 - Dry
Seasonal Year Multi Cover which also align relatively well with historical droughts. The two
product options with kick-in structures, Product 4 - Dry Dry Season with Contingency Kick-in
Cover and Product 5 - Dry Seasonal Year with Contingency Kick-in Cover are also found to have
certain matches with past droughts. However, the dependency between the risk periods,
requiring the first to be triggered, results in a more complex product structure.
79
Technical Report by CelsiusPro AG, March 2021
201. Ultimately, all the five product options would ensure the Government of the respective
countries would receive financial support to mitigate the impact of a severe drought on the
local economy and population. The figures included in the report are initial indicative pricing
scenarios, these are subject to change with changing structures and definitions of seasons. With
these different trigger points, final pricing will come down to the risk taker.
202. It is imperative to be transparent with the client about basis risk and to agree on a suitable
product that represents the client’s expectations. Moving forward some of the limits that could
be encountered are high premiums put forward by the reinsurer. Another is varying definitions
of seasonality; however, this can be easily addressed by recalculating the above structures to
the appropriate time period. One issue with adjusting the definition of the seasons is that the
payouts may not represent what has happened historically.
203. As recent droughts are observed in the data within the last five years this heavily weighs on the
pricing of risk associated with further drought. As the loss is in relation to drought and the
above pricing scenarios reflect this, we believe it is feasible for GPM IMERG data to be used to
create a drought relief product for each of the PICs above. Further assessment also needs to be
made in relation to loss incurred by drought in Samoa and Tonga as there doesn't appear to be
any literature or hard numbers in relation to loss relating to drought.
204. Further pricing fluctuations can be attributed to the reinsurer and their associated loadings in
relation to the ENSO cycle, increased or decreased in relation to years of El Nino and La Nina.
4.9 Next Steps Towards Product Development
205. The following next steps are suggested towards establishing a drought product for the PICs:
Confirm the definition of seasonality in all three countries and adjust parametric covers
accordingly.
Have an open dialogue about years of loss for base cover especially for Tonga & Samoa.
Demand assessment: Confirm interest of countries, what is the budget available for the
drought premium. Get preliminary buy-in from governments and begin the process of
requesting subsidies via the World Bank.
Have an initial dialog with selected group of reinsurers to test product structure and receive
feedback on estimated premium levels.
Implement loss pricing to get a final figure as opposed to the ROL.
Develop policy wording for both PCRIC policy and outward reinsurance placement. Begin
formal market consultations and calculate the reinsurance capacity required.
Finalize operational set-up required to issue the policy and monitor the product inc.
selecting a calculation agent.
After the first policy year, an in-depth analysis of the product and its functionality can be
conducted.
80
Technical Report by CelsiusPro AG, March 2021
Annex
Annex 1 - Data Expansion Method Report
206. GPM IMERG – Benefits and Limitations
The GPM IMERG data set, presently used in this study, extends from June 2000 to present day
and is available on a 0.1-degree (approximately 10-km) global grid. The baseline time resolution
used in this project is 3-hourly precipitation accumulations (although it is also available at half-
hourly resolution).
207. Record Length
This dataset has many benefits, including a short latency (14 hours for the IMERG Late Run used
here), relatively high spatial and temporal resolution, observation-based and not modelled, and
continuous in time and space (as opposed to rain gauge observations, which have significant
spatial and, on occasions, temporal gaps in the data).
208. However, the limitation of using GPM IMERG for the estimation of extreme precipitation is the
record length. The period June 2000 to November 2020 (as used here) allows for 19 annual
maxima peaks to be extracted. This can reasonably be extrapolated to the 80-year return period
level (approximately four times the data length), but is limited in its ability to represent the 150-
year return interval, as may be required for the exhaustion point for this parametric excess
rainfall insurance product.
209. Climate Drivers
Additionally, the sample period used to define rainfall return periods should capture some level
of stationarity in the climate drivers responsible for heavy rain events. The typical baseline
period adopted in climate studies to define "present-day" conditions (e.g. in CMIP5, CMIP6
climate model ensembles, and the IPCC in their Assessment Reports) is the period 1980 - 2020.
210. In the Pacific, the primary modes of climate variability impacting flood and cyclone risk is El
Niño Southern Oscillation (ENSO, on cycles of 2 to 7 years), and over the longer term, the PDO
(Pacific Decadal Oscillation). The PDO can be viewed as a longer-term ENSO background state
that persists for multi-decadal periods and on which individual ENSO events are superimposed.
For our application, it is important to consider the phases of the PDO which we are sampling
from when deriving extreme values. The period 1979 – 2009 was PDO positive, 2005 – 2015
PDO negative, and 2015 - present PDO positive.
81
Technical Report by CelsiusPro AG, March 2021
Figure 55 Long-term trend of the Pacific Decadal Oscillation (PDO) over the cyclone season (Nov-
Mar). Red bars denote PDO positive (La Niña-like background state) and blue bars show PDO
negative (El Niño-like background state). Red line denotes five-year running
211. With the GPM data we have approximately 12 full years of predominantly PDO positive
conditions (2001-2008 and 2015-2019) and seven years of PDO negative (2009-2015) . If we
were to extend this back to 1979, by blending the GPM data with reanalysis data, we would
have approximately 30 years of predominantly PDO positive conditions (1979 - 2008, 2015 -
2019) and seven years of PDO negative (2009 - 2015). PDO positive reinforces La Niña-like
conditions in the Pacific basin. La Niña is generally associated with higher rain, flood and cyclone
risk in the Southwest Pacific region. Using such a baseline (1979 – 2019) could be interpreted as
more risk-averse in terms of assigning extreme values, as we are weighting more towards a La
Niña-like background state. This period is also consistent with the baseline adopted by the IPCC.
212. As can be seen in Figure 55, there is significant inter-annual variability in the PDO phase within
the longer, multi-decadal warm and cool phases. It should also be noted that we are in PDO
positive at present, so an extended time series - in addition to reducing the statistical
uncertainty associated with high (extrapolated) return periods - may also be a better
representation of the near-future flood/cyclone risk for the Pacific region.
213. ERA5 Climate Reanalysis – Benefits and Limitations
A climate reanalysis gives a numerical description of the recent climate, produced by combining
models with observations. It contains estimates of atmospheric and surface parameters (such as
rainfall). Climate forecast models and data assimilation systems are used to 'reanalyse' archived
observations, creating global data sets describing the recent history of the atmosphere, land
surface, and oceans. In this way, the weather models are constrained by, and calibrated to,
observations, at regular time intervals.
82
Technical Report by CelsiusPro AG, March 2021
Figure 56 Example of the coverage of EXMWF’s two main reanalysis products; ERA5 (top image)
and ERA5-Land (bottom image). Source: Copernicus (2016).
214. ERA5
ERA5 is the latest in a series of global climate reanalysis products produced by the European
Centre for Medium-range Weather Forecasts (ECMWF). ERA5 provides a snapshot of the
atmosphere, land surface and ocean waves for each hour from 1979 onwards (and eventually
from 1950). ERA5 covers the whole of the globe (both land and ocean) on a 0.25-degree
(approximately 25-km) grid. ERA5 is one of the most widely used and validated reanalyses for
research, government and business applications, and is freely available for both commercial and
non-commercial use under a Creative Commons licence. A large range of atmosphere, ocean
and land parameters are available to download at either hourly or monthly time resolutions.
215. ERA5-Land
ERA5-Land is a reanalysis dataset providing a consistent view of the evolution of land variables
from 1980 to present at an enhanced resolution compared to ERA5. ERA5-Land has been
produced by replaying the land component of the ECMWF ERA5 climate reanalysis. ERA5-Land
uses ERA5 atmospheric variables as input to control the simulated land fields, such as air
temperature and air humidity. Compared to ERA5, ERA5-Land:
• Only covers land cells, and no over-water cells (see Figure 56)
• Is at a higher spatial resolution (0.1-degrees, or approximately 10-km)
• Accounts for topographic influence on atmospheric variables, such as wind and rain
The temporal and spatial resolutions of ERA5-Land makes this dataset very useful for all kind of
land surface applications such as flood or drought forecasting.
83
Technical Report by CelsiusPro AG, March 2021
216. Observations Assimilation
Reanalysis requires special developments to ensure the best possible temporal consistency of
its products, which can be adversely affected by biases in models and observations, and by the
ever-changing observing system. For example, the assimilation of satellite-based observations
into reanalyses by approximately 1979/1980 dramatically increased the amount of
observational data included in the modelling process, thus better constraining and validating
the model output. While there are several reanalysis product that extent back further than the
1980s, they will increasingly have less observational data assimilated the further back in time
they extend. For this reason, at present, the period 1979 – present is typically regarded as the
extent of the homogenous reanalysis record and aligns with the climate baseline period (1980 –
2020) used by the IPCC. It should be noted that the GPM IMERG satellite data is directly
included as an assimilated observational data source in ERA-5 and - by dynamical downscaling -
into ERA-5 Land.
217. Evaluation of ERA5 and ERA5-Land for Pacific Island States
The two reanalysis products, ERA5 and ERA5-Land were evaluated for suitability across the
seven Pacific Islands Countries (PICs) of interest (Fiji, Vanuatu, Samoa, Tonga, Cook Islands,
Marshall Islands, and Solomon Islands). It was found that, while ERA5-Land is preferable in
terms of spatial resolution and the inclusion of topographic effects on rainfall, in many cases
islands were not classified as “land” in ERA-Land and instead regarded as ocean, therefore
having no data for these locations. This was generally the case for island units having a spatial
area of less than 100 square-kilometres, which is sub-grid-scale for ERA-Land.
This, unfortunately, renders ERA-Land unusable for our application as we require data covering
all islands across the PICs of interest. Rainfall metrics cannot be properly aggregated to
provincial level if there are some cells with missing data. We therefore need to proceed with
using ERA-5 and statistically downscaling this data to the native resolution of GPM IMERG (i.e.,
from 0.25-degree to 0.1-degree resolution). ERA-Land, however, remains a potentially useful
data product for World Bank and other organisations to consider when looking to undertake
similar rainfall-based analyses in other global locations.
218. ERA5 – Data Preparation and Quality Control - Data Acquisition
ERA5 parameters are available for download at either hourly or monthly time resolutions. The
native resolution used for the GPM data was 3-hourly accumulated precipitation, and the aim
was to align the ERA5 data to the same time steps and accumulation periods. ERA5 rain data
were downloaded at hourly time steps between 1 January 1979 and 30 November 2020 for all
global grid cells (approximately 730 GB of data). Time series data for grid cells that intersected
with the seven PICs of interest were then extracted.
219. Preliminary ERA5T and Quality Control
Recent data in ERA5 are considered preliminary and are called ERA5T. This data takes up
approximately twice the memory of the final release data as it includes pre-validation and
uncertainty information. It was found that some of the downloaded ERA5 rain data between
September and November 2020 were erroneous (having very high values, several orders of
magnitude greater than the long-term mean). ERA5 was therefore only used up to 31 August
2020 in this work, to ensure that all data used were final and had undergone proper quality-
control.
84
Technical Report by CelsiusPro AG, March 2021
220. Alignment with GPM IMERG
The GPM IMERG data used in this work were three-hourly rainfall accumulations at 02:00,
05:00, 08:00, 11:00, 14:00, 17:00, 20:00 and 23:00 UTC for each calendar day, where the
timestamp represents the start time of the three-hour accumulation period. The three-hourly
data were summed over an eight-period (24-hour) moving window, and then the maximum of
the 24-hour accumulations for each calendar day was then used as the rain metric. The same
procedure was undertaken for ERA5, with an additional first step of accumulating and aligning
the hourly data to the same three-hourly periods used in GPM (i.e., 02:00, 05:00 etc…).
221. Downscaling and Bias-Correcting ERA5 to GPM for PICs
Reanalysis is increasingly used to complement observational data in parametric insurance to
improve spatial and temporal coverage, and in turn, to ameliorate product structuring and the
reduction of basis risk. An essential component of utilising reanalysis data alongside
observations is the homogenisation – or bias-correction – of the reanalysis record such that it
has a comparable mean and variability (and, in our case, upper quantiles) to the observational
data. Reanalysis will have inherent biases compared to the control data (in our case, GPM) for a
number of reasons including; weather model bias, interpolation, observation assimilation
schemes, and availability of assimilated observations.
222. Only after proper bias-correction can reanalysis be used to complement the control data. A
component implicit to this procedure – if the resolution of the reanalysis and control data vary –
is the downscaling of the ERA5 data to the native resolution of the GPM data. There are a
number of different methods that can be used for downscaling and bias-correction. Here, we
undergo sensitivity testing of four different approaches, and assess their performance based on
the ability of the bias-corrected ERA5 data to replicate the empirical return periods of the GPM
data over the control period (2000 – 2020). These include:
• Polynomial fitting
• Quantile mapping using empirical cumulative distribution functions (CDF)
• Quantile mapping using blended empirical and extreme value CDFs
223. General Approach – Cell by Cell and Monthly
Bias-correction was undertaken on a cell-by-cell and monthly basis. First, all ERA5 and GPM cells
covering one of the 42 identified PIC provinces were collated. Next, ERA5 was truncated to
include only those data covering the control period (June 2000 – May 2020). Then, the nearest
ERA5 grid cell to each GPM cell was located and the bias-correction was undertaken between
these pairs of data (Figure 57). Since the resolution of ERA5 is 0.25-degrees and GPM is 0.1-
degrees, each ERA5 cell is matched to and bias-corrected against several GPM cells. This is the
basis of the statistical downscale, and results in the same number of bias-corrected ERA5 cells
as exists for GPM cells.
85
Technical Report by CelsiusPro AG, March 2021
Figure 57 Illustration of the process of downscaling ERA5 data to the resolution of GPM data.
Those GPM cells that intersect with ERA cells are located and the bias-correction is repeated for
each pair of cells, in effect statistically downscaling the ERA5 data f
224. The data pairs (ERA cell-data and GPM cell-data over the control period) were divided into
monthly segments, in order for the bias-correction to be performed separately for each
calendar month (i.e., 12 times, for all January data, then all February data etc…). This was done
to remove the effect of seasonal variability on the performance of the bias-correction method
and resulted in a fit that was specific to each calendar month. The monthly bias-correction
functions are then applied to monthly segments of the ERA5 data over the control period (2000
– 2020), and also for those that precedes the control period (i.e. 1979 – 2000). The corrected
ERA5 data for the control period 2000 – 2020 is used to validate the performance of the bias-
correction procedure (see Section 5.5). The corrected ERA5 data for the period 1979 – 2000 is
blended with the GPM IMERG data (2000 – 2020) to create an extended observational record
spanning the period 1979 to 2020 for each GPM grid cell. The daily mean across all of the grid
cells that make up a PIC province are then calculated for the period 1979 - 2020. This single,
provincial-level time series represents a temporally-extended version of Metric 1 – Simple
Mean, as referred to previously in this work.
225. Polynomial Fitting
This method involves fitting a polynomial function to the GPM data over the control period and
then adjusting the ERA data using this polynomial fit. The order of the polynomial is determined
based on the correlation and slope of the corrected data that it produces. Only 2nd to 5th-order
polynomials were tested to avoid over-fitting. This approach tends to provide good results for
data within the control period, but because polynomials are unbounded, corrections using this
method can be erroneous towards the tail of the observations and for values higher that those
experienced during the control period. For this reason, the curves may diverge rapidly outside
the fitted range.
226. Quantile Mapping using Empirical CDFs
This method involves determining the empirical cumulative distribution function (CDF) of the
GPM and ERA data over the control period and correcting the ERA empirical CDF by mapping
the quantiles across from the GPM CDF (Figure 58). Once the CDFs are matched, corrections are
made on the ERA time series data based on the transfer function derived for each bin in the CDF
86
Technical Report by CelsiusPro AG, March 2021
histogram. A bin size of 0.001 in the range 0 to 1 was used in this analysis, giving 1001 transfer
functions to choose from.
Figure 58 Example of the quantile mapping routine to match the cumulative probabilities
between the ERA and GPM data over the control period. The CDFs may be derived either
empirically, or after distribution fitting.
227. A limitation of this approach is that events outside the fitted range are usually corrected with
the bin correction at that end of the range, and thus may not well represent the correction for
extreme values that are much larger than those seen during the control.
228. Quantile Mapping using Blended Empirical and Extreme Value CDFs
This method involves portioning the control data into modal and extreme vales, with the former
corrected using the above-described empirical CDF approach, and the latter using an extreme
value distribution. Distribution functions need to be bounded with a fixed area under the curve
(they essentially asymptote to zero). This means that, theoretically, an appropriate extreme
value distribution should better model the corrections associated with extreme rain events that
are higher than experienced during the control period. For each cell and month, the most
appropriate extreme value distribution was selected from:
• Gumbel
• Weibull
• Fréchet
• Pareto
• Exponential
• Log-normal
87
Technical Report by CelsiusPro AG, March 2021
229. The two-sample Kolmogorov-Smirnov test was used to determine the most appropriate
extreme value distribution for each cell/month bias correction. The two-sample Kolmogorov-
Smirnov test is a nonparametric hypothesis test that evaluates the difference between the cdfs
of the distributions of the two sample data. In this case, the extreme value cdf was compared
with the empirical cdf and the best fitting distribution was chosen for the correction of extreme
values.
230. Assessing the Performance of Bias-Correction Methods
Annual Maxima for the GPM data and for the corrected ERA5 data over the control period (June
2000 – August 2020) were used to derive empirical estimates of return periods out to the 19-
year interval (using 19 years of full annual data, from 2001 – 2019). The return period R (in
years) of an event describes the average period of time expected to pass before an event of a
given magnitude is exceeded again. The empirical method to estimate R is to first rank the
annual maxima in decreasing order of magnitude from the largest m = 1 to the smallest m = N,
where N is the number of years in the dataset. Then the return period in its simplest form can
be calculated as:
231. The goodness-of-fit of the return period curve using the corrected ERA5 data, compared to
using the GPM data over the control period, was assessed visually for each of the 42 provinces
for the seven PICs (Figure 56).
Figure 59 Example of comparison of empirical return periods for the ‘Eua province in Tonga,
using the empirical CDF bias-correction approach. The blue line shows the return period
estimates using the uncorrected ERA5 data over the control period (2000 – 2020), an
88
Technical Report by CelsiusPro AG, March 2021
232. Figure 59 shows an example of this process. In this example, it can be seen that the uncorrected
ERA5 return periods (blue line) significantly underestimate those derived using the GPM data
(red line), but when corrected (black line), provide a much better match. Once the corrected
ERA5 data is blended with the GPM data (green line), empirical return periods out to the 41-
year interval – representing the climate of the period 1979 – 2019 – can be plotted. These
plotting probabilities are then used in the final stage to fit to an extreme value distribution and
derive final extended return periods for each province to the 150-year interval.
233. Results – Bias Correction
The bias-correction method resulting in the best match between the GPM and ERA5 empirical
return periods is given in Table 24 Best-performing bias-correction method evaluated using
empirical return periods.for each of the 42 provinces.
Table 24 Best-performing bias-correction method evaluated using empirical return periods.
Country Province Best performing bias correction method
'Cook Islands' Rarotonga' Blended empirical cdf / extreme value
'Cook Islands' North' Blended empirical cdf / extreme value
'Cook Islands' South' Empirical cdf
'Fiji' 'Western' Empirical cdf
'Fiji' 'Northern' Empirical cdf
'Fiji' 'Central' Blended empirical cdf / extreme value
'Fiji' 'Rotuma' Empirical cdf
'Fiji' 'Eastern' Empirical cdf
'Marshall Islands' South' Empirical cdf
'Marshall Islands' North' Blended empirical cdf / extreme value
'Samoa' 'A''ana' Empirical cdf
'Samoa' 'Va''a-o-Fonoti' Empirical cdf
'Samoa' 'Aiga-i-le-Tai' Blended empirical cdf / extreme value
'Samoa' 'Gaga''emauga' Blended empirical cdf / extreme value
'Samoa' 'Gagaifomauga' Blended empirical cdf / extreme value
'Samoa' 'Satupa''itea' Blended empirical cdf / extreme value
'Samoa' 'Palauli' Empirical cdf
'Samoa' 'Vaisigano' Empirical cdf
'Samoa' 'Tuamasaga' Empirical cdf
'Samoa' 'Fa''asaleleaga' Blended empirical cdf / extreme value
'Samoa' 'Atua' Empirical cdf
'Solomon Is' 'Malaita' Empirical cdf
'Solomon Is' 'Central' Empirical cdf
'Solomon Is' 'Makira Ulawa' Empirical cdf
'Solomon Is' 'Temotu' Empirical cdf
'Solomon Is' 'Western' Empirical cdf
'Solomon Is' 'Isabel' Blended empirical cdf / extreme value
'Solomon Is' 'Choiseul' Blended empirical cdf / extreme value
'Solomon Is' 'Rennell and Bellona' Empirical cdf
'Solomon Is' 'Guadalcanal' Empirical cdf
'Solomon Is' 'Honiara' Blended empirical cdf / extreme value
'Tonga' 'Vava''u' Empirical cdf
89
Technical Report by CelsiusPro AG, March 2021
'Tonga' 'Ha''apai' Empirical cdf
'Tonga' 'Tongatapu' Blended empirical cdf / extreme value
'Tonga' 'Eua' Empirical cdf
'Tonga' 'Niuas' Blended empirical cdf / extreme value
'Vanuatu' 'Sanma' Empirical cdf
'Vanuatu' 'Malampa' Empirical cdf
'Vanuatu' 'Tafea' Blended empirical cdf / extreme value
'Vanuatu' 'Penama' Empirical cdf
'Vanuatu' 'Shefa' Empirical cdf
'Vanuatu' 'Torba' Empirical cdf
234. Extreme rainfall events occurring in most provinces (27 out of 42) were best described when the
ERA5 data was bias-corrected using the empirical CDF quantile mapping method. 15 of 42
provinces benefitted from the addition of blending the empirical CDF correction with an
extreme value distribution-based correction for the top 5 % of rain events. In general, those
provinces experiencing extreme rain events in the period preceding the control period (i.e.
during the period 1979 – 2000), that were much higher than those experienced during the
control period (2000 – 2020), benefitted from the additional extreme value correction. Those
provinces where rainfall events in the earlier period did not significantly exceed those
experienced during the control period, were best described by the empirical CDF method. This
is reflective of the fact that the empirical CDF method can reasonably correct extreme data
close to that observed using the correction of the last histogram bin, but becomes erroneous
(often over-predicts) when it attempts to correct values that are much greater than those seen
during the control period.
235. Results – Return Periods - Extreme Value Distribution Fitting and Parameters
236. The empirical return periods from the blended GPM / corrected ERA annual maxima were used
as plotting probabilities to fit an extreme value distribution and extrapolate to the 150-year
return period. As during bias correction, a range of extreme value distributions were trialled
(Gumbel, Weibull, Fréchet, Pareto, Exponential and Log-normal) and the best fitting distribution
was determined using the two-sample Kolmogorov-Smirnov test. Table 25 details the extreme
value distribution used for each province, and their parameters in order that they can be fitted
again at a later stage if required.
237. As can be seen, extreme rain events in most (31 of 42) provinces are best described by a Fréchet
distribution type, while about a quarter (11 of 42) are best fitted using a Weibull distribution.
While other distributions were tested, it appears none of the provinces’ data follow a Gumbel,
Pareto, Exponential or Log-normal distribution.
Table 25 Extended Dataset - Best fitting distribution to annual maxima 1979 – 2019 for each
region and their parameters. k = shape parameter, sigma = scale parameter, mu = location
parameter.
Country Province Best fitting distribution k sigma mu
'Cook Islands' Rarotonga' Weibull -0.23 86.55 232.56
'Cook Islands' North' Weibull -0.06 30.73 97.79
'Cook Islands' South' Fréchet 0.06 53.31 163.02
90
Technical Report by CelsiusPro AG, March 2021
'Fiji' 'Western' Fréchet 0.12 54.78 146.48
'Fiji' 'Northern' Weibull -0.01 44.42 125.31
'Fiji' 'Central' Weibull -0.09 42.21 131.98
'Fiji' 'Rotuma' Fréchet 0.25 74.6 203.9
'Fiji' 'Eastern' Fréchet 0.03 39.91 115.04
'Marshall Islands' South' Weibull -0.03 20.53 87.98
'Marshall Islands' North' Fréchet 0.03 25.62 77.82
'Samoa' 'A''ana' Fréchet 0.18 53.67 142.5
'Samoa' 'Va''a-o-Fonoti' Fréchet 0.24 69.45 160.59
'Samoa' 'Aiga-i-le-Tai' Fréchet 0.21 47.34 136.72
'Samoa' 'Gaga''emauga' Weibull -0.14 45.82 139.69
'Samoa' 'Gagaifomauga' Fréchet 0.01 46.67 141.7
'Samoa' 'Satupa''itea' Fréchet 0.22 37.95 128.16
'Samoa' 'Palauli' Fréchet 0.2 47.1 141.1
'Samoa' 'Vaisigano' Fréchet 0.29 56.23 140.35
'Samoa' 'Tuamasaga' Fréchet 0.18 58.76 145.86
'Samoa' 'Fa''asaleleaga' Fréchet 0.04 48.81 140.87
'Samoa' 'Atua' Fréchet 0.12 66.8 161.59
'Solomon Is' 'Malaita' Fréchet 0.05 29.29 99.11
'Solomon Is' 'Central' Fréchet 0.3 27.58 100.53
'Solomon Is' 'Makira Ulawa' Fréchet 0.06 41.98 127.64
'Solomon Is' 'Temotu' Fréchet 0.06 34.53 132.78
'Solomon Is' 'Western' Weibull -0.2 26.53 103.2
'Solomon Is' 'Isabel' Weibull -0.05 23.55 93.14
'Solomon Is' 'Choiseul' Fréchet 0.07 22.98 92.93
'Solomon Is' 'Rennell and Bellona' Fréchet 0.22 56.27 137.83
'Solomon Is' 'Guadalcanal' Fréchet 0.29 32.92 104.67
'Solomon Is' 'Honiara' Fréchet 0.26 36.75 104.91
'Tonga' 'Vava''u' Weibull -0.08 67.68 205.36
'Tonga' 'Ha''apai' Fréchet 0.01 70.33 186.22
'Tonga' 'Tongatapu' Fréchet 0.22 55.54 159.6
'Tonga' 'Eua' Fréchet 0.21 63.47 171.23
'Tonga' 'Niuas' Weibull -0.02 69.7 204.5
'Vanuatu' 'Sanma' Fréchet 0.33 52.75 126.81
'Vanuatu' 'Malampa' Fréchet 0.41 40.26 112.1
'Vanuatu' 'Tafea' Weibull -0.1 40.79 139.76
'Vanuatu' 'Penama' Fréchet 0.06 54.8 126.01
'Vanuatu' 'Shefa' Fréchet 0.25 53.69 133.62
'Vanuatu' 'Torba' Fréchet 0.01 60.27 149.61
Table 26 Extended Dataset - Best fitting distribution to annual maxima 1979 – 2019 for each
country and their parameters. k = shape parameter, sigma = scale parameter, mu = location
parameter.
Country Best fitting distribution k sigma mu
'Cook Islands' Weibull -0.10 38.06 117.95
'Fiji' Fréchet 0.04 25.99 98.45
91
Technical Report by CelsiusPro AG, March 2021
'Marshall Islands' Fréchet 0.15 17.34 65.73
'Samoa' Fréchet 0.13 48.51 137.15
'Solomon Is' Fréchet 0.10 20.29 73.23
'Tonga' Fréchet 0.03 44.55 122.06
'Vanuatu' Fréchet 0.28 33.51 96.56
238. In addition to identifying the best fitted distribution for the extended dataset, the same
approach was used for the 2001 – 2019 period related to the GPM IMERG data only. The
regional and national parameters are included in Table 27 and Table 28.
Table 27 GPM only - Best fitting distribution to annual maxima 2001 – 2019 for each region and
their parameters. k = shape parameter, sigma = scale parameter, mu = location parameter.
Country Province Best fitting distribution k sigma mu
'Fiji' 'Central' 'Weibull' -0.04 46.31 120.32
'Fiji' 'Eastern' 'Frechet' 0.04 40.81 114.36
'Fiji' 'Northern' 'Gumbel' 0.04 43.47 120.39
'Fiji' 'Rotuma' 'Weibull' -0.03 73.31 204.53
'Fiji' 'Western' 'Frechet' 0.20 46.35 128.24
'Solomon Is' 'Central' 'Frechet' 0.36 25.37 106.18
'Solomon Is' 'Western' 'Weibull' -0.13 22.25 107.23
'Solomon Is' 'Choiseul' 'Frechet' 0.13 20.17 98.27
'Solomon Is' 'Guadalcanal' 'Frechet' 0.30 37.31 107.47
'Solomon Is' 'Honiara' 'Frechet' 0.22 50.26 127.23
'Solomon Is' 'Isabel' 'Weibull' -0.31 20.87 104.14
'Solomon Is' 'Makira Ulawa' 'Weibull' -0.11 43.69 135.85
'Solomon Is' 'Malaita' 'Weibull' -0.01 26.74 105.33
'Solomon Is' 'Rennell and Bellona' 'Frechet' 0.18 57.38 149.45
'Solomon Is' 'Temotu' 'Frechet' 0.21 30.37 129.78
'Tonga' 'Eua' 'Frechet' 0.03 60.55 173.28
'Tonga' 'Ha''apai' 'Weibull' -0.05 78.35 204.40
'Tonga' 'Niuas' 'Frechet' 0.08 57.43 211.36
'Tonga' 'Tongatapu' 'Weibull' -0.07 60.90 173.45
'Tonga' 'Vava''u' 'Weibull' -0.02 67.71 217.67
'Vanuatu' 'Malampa' 'Frechet' 0.67 26.80 103.43
'Vanuatu' 'Penama' 'Frechet' 0.18 39.36 126.36
'Vanuatu' 'Sanma' 'Frechet' 0.21 34.88 125.06
'Vanuatu' 'Shefa' 'Weibull' -0.09 47.21 131.18
'Vanuatu' 'Tafea' 'Weibull' -0.07 42.94 146.49
'Vanuatu' 'Torba' 'Frechet' 0.01 39.40 152.56
'Samoa' 'A''ana' 'Frechet' 0.10 56.52 142.82
'Samoa' 'Va''a-o-Fonoti' 'Frechet' 0.16 62.30 162.99
'Samoa' 'Vaisigano' 'Weibull' -0.02 45.73 146.73
'Samoa' 'Aiga-i-le-Tai' 'Frechet' 0.15 46.09 143.41
'Samoa' 'Atua' 'Frechet' 0.10 61.56 155.56
'Samoa' 'Fa''asaleleaga' 'Weibull' -0.28 46.94 151.52
'Samoa' 'Gaga''emauga' 'Weibull' -0.15 45.46 147.64
92
Technical Report by CelsiusPro AG, March 2021
'Samoa' 'Gagaifomauga' 'Weibull' -0.07 44.22 146.83
'Samoa' 'Palauli' 'Weibull' -0.12 42.54 139.76
'Samoa' 'Satupa''itea' 'Frechet' 0.01 40.67 133.55
'Samoa' 'Tuamasaga' 'Frechet' 0.14 55.11 146.46
'Cook Islands' 'North' 'Weibull' -0.35 35.22 125.27
'Cook Islands' 'South' 'Weibull' -0.08 54.37 163.24
'Cook Islands' 'Rarotonga' 'Weibull' -0.41 76.77 240.56
'Marshall Islands' 'North' 'Frechet' 0.03 17.97 75.95
'Marshall Islands' 'South' 'Weibull' -0.08 16.21 88.29
Table 28 GPM only - Best fitting distribution to annual maxima 2001 – 2019 for each country and
their parameters. k = shape parameter, sigma = scale parameter, mu = location parameter.
Country Best fitting distribution k sigma mu
'Fiji' 'Frechet' 0.03 21.20 94.03
'Solomon Is' 'Frechet' 0.06 21.43 78.52
'Tonga' 'Weibull' -0.19 45.57 136.06
'Vanuatu' 'Frechet' 0.51 21.76 94.17
'Samoa' 'Weibull' 0.00 45.62 136.66
'Cook Islands' 'Weibull' -0.36 28.42 115.67
'Marshall Islands' 'Weibull' -0.19 11.64 65.13
239. Fitted Return Periods
The annual maxima were fitted to the most appropriate extreme value distribution (see Table 2)
to estimate return periods out to the 150-year interval. Figure 58 shows an example of this for
the Central province in Fiji.
Figure 60 Annual maxima data of Metric 1 for the Central province in Fiji fitted to a Weibull
distribution. The blue line and crosses represents the 19-year GPM data (full years 2001 – 2019),
93
Technical Report by CelsiusPro AG, March 2021
and the red line and crosses represent the 41-year (full years 1979 – 2019). 95 % uncertainty
bounds are shown as dashed lines.
240. As can be seen in Figure 60 (and as was the case for most provinces) the addition of 22 years of
more data onto the 19 full years of GPM data did not fundamentally shift the position of the
extreme value curve. Rather, the effect was to reduce the confidence bounds of the return
period estimates and refine the shape of the curve. The reduction in uncertainty bounds is most
apparent for long (i.e. extrapolated) return periods, whereas the refining of the shape of the
curve is most prominent for shorter return periods (i.e. more frequent rain events). This
suggests that, in most cases, the period 2001 – 2019 is broadly a good baseline period from
which to derive rainfall return intervals for the Pacific region, and is largely representative of the
rainfall climate of the past 41 years (i.e. 1979 – 2019).
241. In instances where there was divergence between the 1979 – 2019 and the 2001 – 2019 curves,
it was almost always the case that the 1979 – 2019 curve lay above the 2001 – 2019 curve,
essentially raising the rain volume associated with a given return period. This points to the fact
that there were, on average, higher rain events that occurred over the extended period 1979 –
2000 than over the period 2001 – 2019. While this has not be investigated in detail, it is
consistent with the sampling bias towards PDO positive, or La Niña-like conditions in the Pacific,
which are associated with higher flood and cyclone risk in the Southwest Pacific.
94