CERES
Calibrated Early-warning & Risk Estimation System
CERES is a probabilistic famine early warning system producing weekly 90-day IPC Phase 3+/4+/5 risk forecasts across 43 countries. Every prediction is publicly verifiable through a write-once prospective grading ledger — the first of its kind in famine intelligence.
Open data published on United Nations OCHA Humanitarian Data Exchange under CC BY 4.0. arXiv:2603.09425.
43 countries · Weekly forecasts · 87 IPC transitions · Public grading ledger · CC BY 4.0
43
Countries covered
~95%
of active IPC Phase 3+ caseload
87
IPC transition records
Weekly
Forecast cadence
WHY CERES
What makes CERES different
Existing famine early warning systems — FEWS NET, IPC, WFP HungerMap, INFORM — do not publish prediction-level accuracy records. CERES does.
Probabilistic outputs
CERES produces P(IPC 3+), P(IPC 4+), and P(IPC 5) probability distributions — not ordinal ratings. Every forecast includes 90% sensitivity intervals from 2,000 input-perturbation draws.
Public verification ledger
Every prediction carries a unique ID (CERES-HYP-{ISO3}-{YYYYMMDD}-{hex}). At T+90 days each is graded against IPC outcome data. The grading record is public via the /v1/grades endpoint and tamper-evidenced by Git commit history.
Open data, open API
Predictions are published weekly on UN OCHA Humanitarian Data Exchange under CC BY 4.0. A rate-limited public REST API provides machine-readable access to all forecasts, hypotheses, and graded outcomes.
Multi-source convergence
6 independent data sources: precipitation anomaly (CHIRPS), vegetation stress (MODIS), conflict events (ACLED), IPC classification, food consumption scores (WFP VAM), and cereal price index (FAO/WFP). Convergence scoring detects cross-signal alignment.
METHODOLOGY
7-stage forecasting pipeline
Runs every Monday at 06:00 UTC. Typical runtime 45–90 seconds. Outputs published automatically to public API and OCHA HDX.
01
Ingest
Pull from 6 authoritative sources: CHIRPS v2.0 precipitation, MODIS MOD13A3 vegetation, ACLED conflict events, IPC classification, WFP VAM food consumption scores, FAO/WFP cereal price index.
02
Normalise
Standardise signals to z-scores against historical baselines. Admin0, Admin1, and Admin2 spatial resolution. Temporal alignment to weekly cadence.
03
Signals
Derive anomaly scores per signal per region. Parametric perturbation generates sensitivity intervals (5th–95th percentile) for uncertainty quantification.
04
Convergence
Multi-source convergence scoring detects cross-signal alignment. Weighted combination: IPC 0.25 · Conflict 0.20 · Drought 0.20 · Food access 0.15 · Vegetation 0.10 · Price 0.10.
05
Score
Composite risk score mapped to P(IPC 3+), P(IPC 4+), P(IPC 5) probability distributions. 90% sensitivity intervals from 2,000 input-perturbation draws. Validated against 87 historical IPC transitions.
06
Hypotheses
Generates structured FamineHypothesis objects with unique IDs, sensitivity intervals, contributing signal breakdown, and falsifiability criteria. Pre-registered for T+90 day prospective grading.
07
Publish
Predictions published to public API, OCHA HDX dataset (CC BY 4.0), and internal grading ledger. Every prediction pre-registered for T+90 day prospective grading.
Data sources and signal weights
| Source | Provider | Signal | Weight |
|---|---|---|---|
| CHIRPS v2.0 | UCSB Climate Hazards Group | Precipitation anomaly | 20% |
| MODIS MOD13A3 | NASA Earthdata | NDVI vegetation stress | 10% |
| ACLED | Armed Conflict Location & Event Data | Conflict events + fatalities | 20% |
| IPC Classification | IPC Global Platform | Current/projected IPC phase | 25% |
| Food Consumption Score | WFP VAM | Food access + household coping | 15% |
| Cereal Price Index | FAO / WFP | Market price stress | 10% |
PROOF
Verification architecture
Northflow follows strict proof-before-claim discipline. Every CERES prediction is pre-registered and graded against outcomes — publicly, automatically, permanently.
Unique prediction IDs
Every forecast carries a permanent identifier:
CERES-HYP-{ISO3}-{YYYYMMDD}-{6-char hex}T+90 day grading
At T+90 days each prediction is automatically graded against IPC outcome data. The grading record is public via the /v1/grades endpoint and tamper-evidenced by Git commit history.
Model performance
- AUC = 0.84 (95% CI [0.63, 1.00] — reported for completeness)
- n=87 historical IPC transitions
- 2,000 input-perturbation draws (sensitivity intervals)
- Brier score: 0.0066 vs 0.0494 baseline
Back-validation results
| Event | Year | Result |
|---|---|---|
| Somalia famine | 2011 | TIER-1 correct |
| South Sudan famine | 2017 | TIER-1 correct |
| Ethiopia Tigray crisis | 2022 | TIER-1 correct |
| Yemen food crisis | 2021 | TIER-1 correct |
Prospective verification begins March 2026. First graded outcomes expected May 2026.
Alert tiers
P(IPC 4+) ≥ 0.50 OR CRITICAL multi-source convergence
P(IPC 3+) ≥ 0.45 OR WARNING convergence
Watch — elevated multi-source signal below TIER-2 threshold
Pre-registered calibration targets
Brier Score target
June 2026In-sample Brier score 0.0066 vs 0.0494 persistence baseline
TIER-1 precision/recall
September 2026First prospective outcomes available May 2026; full evaluation September 2026
Full CRPS evaluation
March 2027Continuous Ranked Probability Score across full 12-month prospective cohort
DATA ACCESS
Open API & data access
CERES operates an open public API with no authentication required. All predictions are published on OCHA Humanitarian Data Exchange under CC BY 4.0.
API endpoints
Base URL: ceres-core-production.up.railway.app
/v1/predictionsAll 43 countries — current forecasts
/v1/predictions/{region_id}Single country forecast
/v1/admin1/predictionsProvince-level forecasts
/v1/admin2/predictionsDistrict-level forecasts
/v1/hypothesesFull FamineHypothesis objects
/v1/archive/latestLatest archived prediction set
/v1/gradesGraded prediction outcomes (public)
/v1/grades/metricsAggregate accuracy metrics
Access tiers
Public
Free- Public REST API (rate limited)
- CC BY 4.0 weekly predictions
- OCHA HDX dataset access
- /v1/grades public ledger
- GitHub repository (ceres-core)
Institutional
Subscription- TIER-1 alert webhooks
- 1,000 requests / hour
- Weekly PDF briefings
- Parquet export
- Priority support
- Sub-national Admin2 data
United Nations OCHA Humanitarian Data Exchange
Dataset: "Global — CERES Famine Risk Predictions" · License: CC BY 4.0 · Updated weekly
View on UN OCHA Humanitarian Data Exchange →ENGAGEMENT
Research access & institutional engagement
CERES data is freely available. For institutional briefings, subscription access, integration support, or research collaboration, use the pathways below.
Live system
Access the CERES dashboard, current forecasts, and interactive country views at the live production system.
Open ceres.northflow.no →Research paper
Read the full methodology in the arXiv preprint. Planned for submission to Nature Food following first prospective validation results (Q3–Q4 2026).
arXiv:2603.09425 →Institutional subscription
TIER-1 alert webhooks, 1,000 req/hr API access, weekly PDF briefings, and Parquet export. Designed for humanitarian organisations, government agencies, and institutional early warning teams.
Request access →Open data — OCHA HDX
Weekly predictions published on UN OCHA Humanitarian Data Exchange under CC BY 4.0. Machine-readable CSV, updated every Monday.
Open on HDX →Request a briefing
For humanitarian organisations, government agencies, academic researchers, or funding bodies interested in CERES methodology, integration, or prospective validation programme.
Request CERES briefing