Turning
data into
decisions.

The Standardised Precipitation Evapotranspiration Index (SPEI)

SPEI captures weather severity as a single comparable number by subtracting Potential Evapotranspiration (PET) from rainfall — measuring how thirsty the soil actually is, not just how much rain fell. A timescale of k=3 months is used here, making it suitable for monitoring soil moisture and crop yield. Because it is a standardised normal variable (mean 0, SD 1), its values are directly comparable across all ten regions.

Key findings

Three overarching conclusions emerge from the 73-year record, each with direct implications for insurance product design and agricultural policy.

ENSO and IOD — climate drivers of Ugandan rainfall

Two large-scale oceanic phenomena drive Uganda's most severe agricultural weather events. Understanding their interaction is essential for forward-looking insurance design.

Decade-by-decade SPEI volatility by ZARDI (%)

Volatility measures the proportion of SPEI observations outside the predictable farming zone [−1.5, +1.5]. The 1980s were the most stable decade across all zones. The 2020s already show dangerous increases with only four years of data recorded.

* 2020s data is truncated — only 4 out of 10 years recorded at time of analysis. Highlighted in red: highest values (>20%). Green: More stable (Below 10%).

Seasonal risk of extreme events

Using aggregated monthly SPEI values per ZARDI, a running cumulative sum detects whether conditions have crossed the extreme thresholds of +2.0 (extreme wetness) or −2.0 (extreme drought) and remained there for at least three consecutive months. The table below shows the number of such confirmed extreme seasonal anomalies per ZARDI per decade, split by season (M = MAM long rains, S = SON short rains) and peril (Dry / Wet). The 2020s reflect truncated data — four out of ten years recorded.

Insurance trigger structures

Parametric insurance pays out when an objective index crosses a predefined threshold — no field visit or loss adjustment required. Three structures are proposed, each suited to different ZARDI risk profiles.

Policy and institutional recommendations

Three actionable steps for government, insurers, and development finance partners — designed for immediate implementation with existing infrastructure.

• R1
  Establish a National Agricultural Climate Risk Registry

  A joint initiative between the Ministry of Finance, MAAIF, UBOS, and the Uganda Meteorology Authority should record SPEI anomalies, ENSO and IOD event classifications, and crop yield data at the ZARDI level annually. This centralised registry would serve as a shared foundation for insurance pricing, disaster response, and development finance — reducing the cost of fragmented responses to the same risks.
• R2
  Formally adopt SPEI as Uganda's standard agricultural insurance index

  The Uganda Insurers Association, in collaboration with MAAIF, should supersede single-variable rainfall thresholds with SPEI. Rainfall-only triggers would not activate during a drought driven by heat stress despite normal rainfall — leaving farmers underinsured. SPEI's standardised properties mean it is statistically consistent across all ZARDIs regardless of differing baseline rainfall, making cross-portfolio pricing manageable.
• R3
  Integrate IOD and ENSO signals into seasonal agricultural advisory

  ICPAC, NOAA, and the Australian Bureau of Meteorology publish probabilistic IOD and ENSO forecasts between June and August — three to six months before the MAM long rains. These forecasts should be propagated into ZARDI-level planting guidance distributed to farmers, with explicit reference to the risk elevations documented in this report.

Dynamic Model Design

The previous sections identified historical climatic shifts, seasonal and geographic asymmetries, and correlations with climate drivers (ENSO and IOD) to understand their impact on staple crop yields. The use of historical loss tables or burn rate to determine appropriate pricing is becoming increasingly unreliable due to climate change, and the risk of farmer claims being underpaid or overpaid by insurers will remain unmitigated. To address the increasing unpredictability caused by climate change, this section explores a dynamic model framework that uses SPEI as a responsive, self-updating, and forward-looking trigger.

Layer 1 – Set up the baseline period

• L1.A
  Quantile Delta Mapping (QDM) for tail estimation

  Using the structural break at approximately 2000, QDM harmonises the pre-2000 record onto the post-2000 quantile structure. This produces a full 1950-2023 series, whose extreme quantiles reflect current climate magnitudes rather than the blended 73-year average. This approach assumes that the post-2000s is the more relevant baseline for insurance product pricing; however, the expanding variance through the 2020s creates greater residual uncertainty.
• L1.B
  Rolling window for annual trigger recalibration

  Each year, the trigger is recalculated using the most recent 20-25-year window of SPEI data rather than the full historical record to make it self-updating. This track identifies the threshold used as the trigger for the next season.

Layer 2 – Calculate conditional probabilities

• L2.A
  IOD/ENSO conditional probability tables

  For each ZARDI, season, and climate driver combination, the empirical probability that SPEI crosses the trigger threshold is computed. The output is a table comprising the pre-season risk signals, relevant for Horizon 1.
• L2.B
  Logistic regression for continuous probability scoring

  Using lagged climate drivers (DMI observed in JJA and RONI in the prior season), the binary outcome of whether SPEI crosses the trigger threshold in the upcoming MAM or SON season is predicted. This provides continuous probability scoring that scales with the magnitude of the IOD/ENSO signal.
• L2.C
  Copula modelling for coincident event tails

  When sample sizes of coincident events are small, a bivariate copula models the joint tail dependence between the DMI and RONI without requiring the full co-occurrence record. Given the expected lower-tail dependence on coinciding climate drivers, a Clayton copula is more appropriate.

Layer 3 – Forward Risk projection

• L3.A
  Horizon 1: Seasonal pre-warning (3-6 months)

  At the start of JJA each year, the Australian Bureau of Meteorology and NOAA CPC provide respective probabilistic DMI and RONI forecasts, and ICPAC is ingested for the Greater Horn of Africa seasonal outlook. These forecasts are propagated through the conditional probability tables using a weighted average across forecast scenarios, yielding a ZARDI-level trigger probability for the upcoming SON or MAM season. Retrospective validation is available within the existing dataset. For each year, JJA-observed DMI and RONI are used as a proxy for the seasonal forecast. The conditional probability table is applied, and the resulting seasonal outlook is compared to the actual SPEI outcome. This gives a hindcast skill score, like the Brier Skill score, quantifying how much predictive value IOD or ENSO conditioning adds over a climatological baseline.
• L3.B
  Horizon 2: Monthly in-season monitoring

  SPEI is updated monthly as new precipitation and temperature data arrive during an active season. An insurer monitors whether the SPEI is trending toward the trigger threshold in real time, and a trajectory-based alert occurs when the current SPEI is in the warning zone. This allows them to assess whether a payout is likely and reduce payout timelines by processing claims before the trigger formally fires. The required data can be sourced from monthly CHIRPS actuals, ERA5-Land, or the World Bank CCKP API updates.
• L3.C
  Horizon 3: Sub-seasonal crop-stage monitoring

  Using sub-seasonal forecast products such as CHIRPS-GEFS to monitor SPEI trajectory during agronomically critical growth stages, this horizon helps understand where moisture stress has the greatest impact. However, this requires switching the timescale from 3 to 1 for the SPEI calculations to resolve the 15-day signals of CHIRPS-GEFS. This can aid in preventative payouts for irrigation or early harvest before crop loss occurs.

Layer 4 – Dynamic Recalibration and Basis Risk Scoring

• L4.A
  Rolling event coincidence analysis

  Building on the static event coincidence analysis, this version re-runs annually as new FAOSTAT yield data and SPEI actual values arrive to check whether triggered years coincide with yield-loss years.
• L4.B
  Basis risk scoring

  Type 1 basis risk is where a trigger is met without a significant yield loss, causing the insurer to overpay. Type 2 basis risk is where significant yield loss occurs without a trigger being met, leaving the farmer unprotected.
• L4.C
  Quantile Regression for trigger boundary uncertainty

  This gives a confidence band, mitigating the uncertainty that a single threshold would give in a non-stationary climate. This technique is important for regulatory documentation, ensuring robust auditability.