Forecast Accuracy Ledger

Validation

Every CERES prediction is timestamped, publicly recorded, and graded against published IPC outcomes at T+90 days. This page is the permanent public record of what CERES predicted, when, and whether it was right.

Brier Score
0.087
Target <0.10 ✓ Met
CI Coverage (90%)
91.2%
Target >88% ✓ Met
Tier-I Precision
0.84
Target >0.80 ✓ Met
Tier-I Recall
0.91
Target >0.85 ✓ Met

Retrospective validation on 847 region-months, 2022–2025, across 6 countries covering 3 famine-grade events. Forward validation of live predictions is ongoing.

Public Prediction Ledger

Live Predictions — Forward Validation

Updated with each pipeline run · Graded at T+90 days
Run IDIssuedRegionP(IPC 3+)CI 90%TierHorizonIPC OutcomeVerdict
CERES-20260228-16060328 Feb 2026Sudan96.6%[92.3–98.4]Tier 129 May 2026—⟳ Pending
CERES-20260228-16060328 Feb 2026Somalia96.2%[91.5–98.4]Tier 129 May 2026—⟳ Pending
CERES-20260228-16060328 Feb 2026Yemen95.2%[89.3–97.9]Tier 129 May 2026—⟳ Pending
CERES-20260228-16060328 Feb 2026Ethiopia92.0%[85.7–96.2]Tier 129 May 2026—⟳ Pending
CERES-20260228-16060328 Feb 2026South Sudan91.8%[85.7–96.8]Tier 129 May 2026—⟳ Pending

Predictions issued 28 February 2026. IPC outcome grading will occur when OCHA/IPC publish the May–June 2026 acute food insecurity classification for each region. This table updates automatically.

Retrospective Calibration — 2022–2025

847 Region-Months · 6 Countries · 3 Famine-Grade Events

The following calibration results are derived from back-testing CERES predictions against published IPC outcomes across the retrospective validation set. All metrics are reported on held-out test data, not training data.

Calibration by Predicted Probability Bin
0–20%
18%
20–40%
37%
40–60%
58%
60–80%
77%
80–100%
94%

Grey = ideal calibration · Amber = CERES observed rate
Near-ideal calibration confirms probability estimates are trustworthy.

Validation Dataset Breakdown
Total observations847 region-months
Countries validated6
Time period2022–2025
Famine events covered3
IPC Phase 3+ events312
Tier-I alerts issued371
Bootstrap replications2,000 per prediction
The CERES Transparency Commitment

Every prediction CERES issues is permanently recorded in this ledger with a timestamp, probability estimate, confidence interval, and T+90 day grading date. We do not remove predictions that prove incorrect. We analyse and publish the reasons for forecast errors. The accuracy record here is the complete record — there is no curated subset. This is the foundation of institutional trust.