ER · per-country backtest
ER forecast vs reality
Every Voidly Sentinel shutdown-risk forecast for ER, plotted against what actually happened. 28 (predicted, observed) pairs from the rolling 30-day evaluation window.
Updated every 30 min · CC BY 4.0 · Raw JSON · Current forecast →
Forecasts evaluated
28
since Apr 17
Accuracy @ 0.5
89.3%
25/28 correct
Brier score
0.002
lower is better
Observed positive rate
0%
mean predicted 4%
Forecast time series
Blue line: forecast probability. Green ✓ markers: forecast was right. Red ✗ markers: forecast was wrong. Dashed line at 0.5 is the binary decision threshold.
Y axis: forecast probability · Faint tick: distance to observed outcome (0 or 1)
All 28 predictions (newest first)
| Eval date | Forecast | Pred ≥ 0.5? | Observed? | Correct? |
|---|---|---|---|---|
| May 14 | 4.6% | — | no event | ✓ |
| May 13 | 5.4% | ↑ | no event | ✗ |
| May 12 | 4.9% | — | no event | ✓ |
| May 11 | 4.9% | — | no event | ✓ |
| May 10 | 4.9% | — | no event | ✓ |
| May 9 | 2.5% | — | no event | ✓ |
| May 8 | 4.4% | — | no event | ✓ |
| May 7 | 2.7% | — | no event | ✓ |
| May 6 | 4.8% | — | no event | ✓ |
| May 5 | 4.0% | — | no event | ✓ |
| May 4 | 3.6% | — | no event | ✓ |
| May 3 | 3.9% | — | no event | ✓ |
| May 2 | 3.9% | — | no event | ✓ |
| May 1 | 4.9% | — | no event | ✓ |
| Apr 30 | 3.3% | — | no event | ✓ |
| Apr 29 | 3.7% | — | no event | ✓ |
| Apr 28 | 4.3% | — | no event | ✓ |
| Apr 27 | 5.2% | ↑ | no event | ✗ |
| Apr 26 | 4.0% | — | no event | ✓ |
| Apr 25 | 3.7% | — | no event | ✓ |
| Apr 24 | 4.7% | — | no event | ✓ |
| Apr 23 | 5.4% | ↑ | no event | ✗ |
| Apr 22 | 3.9% | — | no event | ✓ |
| Apr 21 | 4.9% | — | no event | ✓ |
| Apr 20 | 3.8% | — | no event | ✓ |
| Apr 19 | 4.5% | — | no event | ✓ |
| Apr 18 | 4.0% | — | no event | ✓ |
| Apr 17 | 3.8% | — | no event | ✓ |
How to read this
- Each row is one historical forecast. We made the prediction at eval_date, then waited the 7-day horizon, then graded against the observed outcome.
- Forecast % is the calibrated probability we published that day. Post-recalibration (2026-05-20) these now match actual observed rates much better — see the refit finding.
- Pred ≥ 0.5 shows the binary alert decision. Note we usually fire alerts at a lower threshold (see /v1/sentinel/global_heatmap for the live cutoff) — the 0.5 column here is for backtest scoring consistency with the global confusion matrix.
- Correct = both (pred ≥ 0.5) and (observed) agree.
Related
- /atlas/forecast/er — current 7-day calibrated forecast for ER with SHAP drivers
- /er — ER country page (current state, recent incidents)
- /sentinel/backtest — global reliability diagram + per-country comparison table
- Calibration refit writeup