Atlas forecast
Follow the
signal.
Explore possible changes in OONI measurements.
Choose a country. Inspect its record.
Seven-day event estimates
Source JSONUnusual changes in a country’s OONI signals. An event estimate does not establish a full national shutdown or its exact date.
honest_forecast_v1- EthiopiaET Elevated85.2%
- BelarusBY Elevated79.0%
- EgyptEG Elevated78.8%
- PakistanPK Elevated78.2%
- NigeriaNG Elevated75.3%
- SyriaSY Elevated69.0%
- Saudi ArabiaSA Watch65.1%
- KazakhstanKZ Watch60.3%
The board does not supply per-country observation dates. Open a country’s source response to check its “as of” date.
All 23 event-model records
| Country | Event estimate | Model | Source band | Threshold | Threshold flag | Source |
|---|---|---|---|---|---|---|
| EthiopiaET | 85.2% | honest_forecast_v1 | Elevated | 67.5% | Above | Country JSON |
| BelarusBY | 79.0% | honest_forecast_v1 | Elevated | 67.5% | Above | Country JSON |
| EgyptEG | 78.8% | honest_forecast_v1 | Elevated | 67.5% | Above | Country JSON |
| PakistanPK | 78.2% | honest_forecast_v1 | Elevated | 67.5% | Above | Country JSON |
| NigeriaNG | 75.3% | honest_forecast_v1 | Elevated | 67.5% | Above | Country JSON |
| SyriaSY | 69.0% | honest_forecast_v1 | Elevated | 67.5% | Above | Country JSON |
| Saudi ArabiaSA | 65.1% | honest_forecast_v1 | Watch | 67.5% | Not above | Country JSON |
| KazakhstanKZ | 60.3% | honest_forecast_v1 | Watch | 67.5% | Not above | Country JSON |
| UzbekistanUZ | 56.8% | honest_forecast_v1 | Watch | 67.5% | Not above | Country JSON |
| MyanmarMM | 45.4% | honest_forecast_v1 | Watch | 67.5% | Not above | Country JSON |
| CubaCU | 44.0% | honest_forecast_v1 | Watch | 67.5% | Not above | Country JSON |
| VenezuelaVE | 24.1% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| SudanSD | 16.4% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| VietnamVN | 8.8% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| RussiaRU | 3.3% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| IranIR | 3.1% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| TurkeyTR | 1.8% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| BangladeshBD | 1.3% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| ChinaCN | 1.2% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| IndiaIN | 0.8% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| IndonesiaID | 0.3% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| ThailandTH | 0.3% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
| TurkmenistanTM | 0.1% | honest_forecast_v1 | Low | 67.5% | Not above | Country JSON |
7 records from other or unspecified models
These records remain separate because their model identity or forecast basis differs. Use the corresponding source notes and validation, not the event-model evaluation below.
| Country | Source value | Model | Source band | Threshold | Threshold flag | Source |
|---|---|---|---|---|---|---|
| LebanonLB | 27.1% | Not supplied | Band not supplied | 35.0% | Not above | Country JSON |
| BrazilBR | 8.7% | Not supplied | Band not supplied | 5.0% | Above | Country JSON |
| EritreaER | 5.1% | Not supplied | Band not supplied | 5.0% | Above | Country JSON |
| NicaraguaNI | 4.8% | Not supplied | Band not supplied | 27.5% | Not above | Country JSON |
| PhilippinesPH | 4.2% | Not supplied | Band not supplied | 5.0% | Not above | Country JSON |
| North KoreaKP | 4.1% | Not supplied | Band not supplied | 5.0% | Not above | Country JSON |
| MalaysiaMY | 3.1% | Not supplied | Band not supplied | 22.5% | Not above | Country JSON |
The earlier Sentinel v1 forecast is a current-regime signal with unreliable new-shutdown onset skill. A low output does not show that a country has an open internet.
Read the legacy forecast evaluationData service checked at .
Put the model
beside its baselines.
Performance on later data matters more than a good fit to the past.
Inspect the model cardRolling-origin validation · precision–recall AUC
Higher is better on the same evaluation. PR-AUC is a ranking measure, not the percentage of correct predictions. The recent-event baseline counts event days in the seven-day input window.
Model card generatedSee the wider
pattern.
Inspect regional aggregates and their model scope.
Follow the evidenceFind what
changed.
Explore recent observations and source records.
Understand the record.
What this model’s event means
The event model uses OONI measurement signals. Its target is an anomaly-rate spike relative to a country’s own recent baseline within the next seven days. It is not trained on the same target as the separate KeepItOn shutdown-risk model.
The model card specifies a 28-day baseline, a z-score threshold of 2, an anomaly-rate floor of 8.0%, and a minimum event-day measurement count of 50.
Stable blocking can produce a low change estimate. Sparse observations and differences between countries limit interpretation. Read standing censorship separately, and verify any incident against its evidence.
Evaluation and uncertainty
The comparison above uses strict temporal rolling-origin evaluation for honest_forecast_v1. Country history and measurement coverage matter; transferring to a new country is a different test.
- Rolling-origin ROC AUC
- 0.824
- Mean validation precision
- 29.5%
- Mean validation recall
- 39.0%
- Brier score
- 0.153
- Leave-one-country-out mean AUC
- 0.742
Precision and recall are averages across validation folds, using thresholds selected from each fold’s training data. They do not describe a single current threshold.
These are evaluation results, not guarantees for an individual country. The heatmap supplies no confidence interval. No interval or exact-day trajectory is inferred for the dots above.
Existing country pages include legacy daily trajectories, attributions and interval fields. Check the model identity and source notes before applying them to the event-model headline.
Archived legacy-model movement
This is a separate comparison of archived v1 scores. It does not measure movement in the event-model probabilities above. Prior and current values use the same legacy basis.
Earlier snapshot:
Later snapshot| Country | Earlier v1 score | Later v1 score | Change |
|---|---|---|---|
| Myanmar | 73.5% | 31.7% | -41.8 pp |
| Ethiopia | 29.1% | 70.4% | +41.3 pp |
| Russia | 36.9% | 1.0% | -35.9 pp |
| Cuba | 76.5% | 44.0% | -32.5 pp |
| Iran | 32.9% | 3.6% | -29.2 pp |
| Brazil | 32.0% | 8.7% | -23.3 pp |
| Kazakhstan | 66.9% | 50.4% | -16.5 pp |
| Uzbekistan | 43.5% | 58.8% | +15.3 pp |
| Nigeria | 52.7% | 67.3% | +14.6 pp |
| Pakistan | 67.4% | 81.7% | +14.3 pp |
| Sudan | 5.5% | 16.4% | +10.9 pp |
| Belarus | 69.2% | 79.8% | +10.6 pp |
Model recalibration can change scores without establishing a change on the network. Raw source notes distinguish the legacy deltas from the newer headline values.
Related forecasts, alerts and reuse
Election context, case studies and legacy trajectories have their own scope. A worked alert example does not establish forecast timing skill.
Voidly-original records and annotations are CC BY 4.0. Raw upstream measurements retain their source-specific license. Read the data terms before reuse.