1 Key Findings: 2026 Trigger Development
1.1 Introduction
This chapter summarizes the development of the 2026 anticipatory action trigger for drought in Afghanistan’s northern provinces. The trigger system has two activation windows:
- March: Early warning based on SEAS5 seasonal precipitation forecast
- April: Primary trigger based on observed indicators (Combined Drought Index)
The core challenge: define what “drought” means for prediction, then build indicators that can predict it early enough to enable anticipatory action.
1.2 Defining Drought: Why Return Period 4?
1.2.1 ASI as the Outcome Variable
We use the FAO Agricultural Stress Index (ASI) as the outcome variable — a satellite-derived measure of vegetation health anomalies during the growing season. For agricultural livelihoods, reduced vegetation correlates with crop failure and food insecurity.
Drought is defined as years where ASI exceeds a return period (RP) threshold. A 4-year RP means “worse than 1 in 4 years” — approximately the worst 25% of years.
1.2.2 Validating Against CERF Allocations
To test whether ASI captures “real” droughts worth acting on, we compared ASI thresholds against CERF Rapid Response allocations — an independent, external validation of drought severity representing the humanitarian community’s real-time assessment that conditions warranted emergency funding.
ASI achieves ROC-AUC = 0.92 against CERF Rapid Response allocations, indicating our outcome variable captures droughts that warranted humanitarian response historically. This validates ASI as a reasonable proxy for “drought severe enough to act on.”
See Figure 13.1 and Table 13.1 in Chapter 14 for the full analysis.
1.2.3 Threshold Selection
We evaluated how well different ASI thresholds (not trigger thresholds) predict CERF allocations — this tells us which RP definition of “drought” best matches historical humanitarian response.
Key finding: RP 3 and RP 4 produce identical CERF alignment (F1 = 0.714, perfect recall). No years fall exactly between these thresholds, so both flag the same drought years. Performance degrades at RP ≥ 5 (missed CERF years) and RP ≤ 2 (too many false positives).
See Figure 13.1 in Chapter 14 for exact metrics by threshold.
We chose RP 4 because:
- Stricter threshold: Leaves room for SEAS5 forecast uncertainty in the combined trigger
- Operational alignment: Consistent with typical AA activation frequencies (every 3-5 years)
- Sufficient positive cases: 10 drought years in 42-year record for model training
1.3 Building the April Trigger: Combined Drought Index
1.3.1 Model Approach
We use ridge regression to predict ASI RP ≥ 4 using early-season indicators available by April. Ridge regularization prevents overfitting on our small sample (n=42 years).
1.3.2 CDI Components and Weights
The Combined Drought Index (CDI) is a weighted sum of four standardized indicators:
| Component | Description | Weight |
|---|---|---|
| VHI | Vegetation Health Index (March) | ~35% |
| mixed_fcast_obsv | Observed Mar precip + forecasted Apr/May | ~28% |
| snow_cover | Snow cover fraction (March) | ~22% |
| ASI | Agricultural Stress Index (March) | ~15% |
Higher CDI = higher drought risk. The F1-optimized threshold corresponds to approximately RP 3.9 — close to our policy target of RP 4.
See Figure 11.1 and Table 11.1 in Chapter 12 for the CDI formula derivation, component contributions, and historical reanalysis.
1.3.3 Performance
Leave-one-out cross-validation (LOOCV) provides honest out-of-sample estimates:
- F1 Score: ~0.82
- Precision: ~82% of triggered years are actual droughts
- Recall: ~82% of droughts are correctly triggered
See the “Model Validation (LOOCV)” section in Chapter 12 for the full confusion matrix.
1.4 Adding Early Warning: March SEAS5
1.4.1 The Value of Earlier Triggers
The SEAS5 seasonal precipitation forecast (March-April-May) is available one month before the April CDI. Adding SEAS5 using OR logic (trigger if CDI OR SEAS5 exceeds threshold) provides early warning for droughts.
Critically: every drought caught by SEAS5 is already caught by CDI. SEAS5 adds no unique detection — its value is purely in lead time.
1.4.2 SEAS5 Threshold Options
| SEAS5 Threshold | Droughts with Early Warning | False Alerts Added | Early:False Ratio |
|---|---|---|---|
| RP ≥ 3 | 7/10 (70%) | 5 | 1.4:1 |
| RP ≥ 4 | 6/10 (60%) | 4 | 1.5:1 |
| RP ≥ 5 | 4/10 (40%) | 2 | 2:1 |
| RP ≥ 6 | 4/10 (40%) | 0 | ∞ |
| RP ≥ 7 | 3/10 (30%) | 0 | ∞ |
RP ≥ 6 adds zero false positives while still providing early warning for 40% of droughts. This conservative threshold:
- Avoids triggering on false alarms from forecast uncertainty
- Provides March warning for 4 of 10 historical drought years
- Maintains high precision in the combined trigger
See Figure C.1 in the SEAS5 Baseline Sensitivity appendix for the full analysis.
1.5 Final Trigger Specification
March Window (Early Warning)
- Signal: SEAS5 Mar-Apr-May precipitation forecast
- Threshold: RP ≥ 6
- Logic: If SEAS5 exceeds threshold → early trigger
April Window (Primary)
- Signal: Combined Drought Index (CDI)
- Components: VHI, mixed forecast/observation precipitation, snow cover, lagged ASI
- Threshold: F1-optimized (~RP 4)
- Logic: If CDI exceeds threshold → trigger
Combined Logic: OR (either signal can trigger action)
1.6 Summary Metrics
| Metric | Value | Source |
|---|---|---|
| ASI vs CERF ROC-AUC | 0.92 | Ch 14 |
| CDI LOOCV F1 | ~0.82 | Ch 12 |
| SEAS5 RP≥6 early warning | 40% of droughts | Ch 13 |
| SEAS5 RP≥6 false positives added | 0 | Ch 13 |
| Combined trigger RP | ~3.5 years | Ch 13 |
1.7 What This Means Operationally
- In most drought years, the April CDI will trigger action
- In ~40% of drought years, SEAS5 provides one month earlier warning
- False alarm rate is controlled by using RP 4 for CDI and RP 6 for SEAS5
- Combined activation frequency is approximately every 3.5 years — within the acceptable range for anticipatory action
The trigger system balances early warning (SEAS5) with reliable detection (CDI), using independent signals at different lead times to maximize both coverage and precision.