Atlas · 1d / 7d / 30d forecasts

Multi-horizon forecast index

Three honest forecast horizons per country — each with its own XGBoost model, its own SHAP attribution, and its own 90% conformal interval. Click a country for per-horizon detail.

Trained 2026-05-21 · Recommendation: ship_all_three · model info

1d LOCO AUC

0.907

inflated — see scope note

7d LOCO AUC

0.883

inflated — see scope note

30d LOCO AUC

0.845

inflated — see scope note

What this forecast is — and is not

The 7-day forecast is a current-regime risk signal: it reliably tells you which countries are currently in a sustained internet-blocking regime. It is not a shutdown-onset predictor — it does not reliably call a new shutdown before it happens.

On an honest forward-temporal split (train on the past, test only on the strictly-future window) the production model scores AUC 0.589 on its raw label — near chance.
The label target_7day is a 7-day sliding window, so it is 98.9% autocorrelated day-to-day. A trivial predict-yesterday baseline scores AUC 0.957 — high numbers here mean “what is blocked stays blocked,” not predictive skill.
On the rows where a shutdown actually begins (transitions) the forecast scores AUC ~0.33 — below chance.

Use it to see where censorship is entrenched today and to read the SHAP drivers + conformal interval — not as an early-warning siren. Read the full honest re-evaluation →

Multi-horizon feed temporarily unavailable. See /status.

Why three horizons?

A single 7-day number is what most journalists ask for, but it hides two important things: the current-regime risk level tomorrow, and the operational ceiling over the next month. We train three independent XGBoost models and publish all three with LOCO AUC numbers and per-horizon SHAP attribution. Note: each horizon uses the same target_Nday sliding-window label as the production v1 forecast, so the LOCO AUCs are inflated by label autocorrelation the same way — these are current-regime risk signals, not shutdown-onset predictors. See the onset-skill finding.

The monotonicity check is a free honesty signal: longer windows must contain shorter windows, so if P(1d) > P(30d) the three models are disagreeing and you should treat the headline numbers with caution. See /atlas/models for the full registry, or /methodology for the full pipeline.

/atlas/forecast — calibrated 7-day forecast index
/atlas/score-v2 — level-aware composite score (uses 30d avg)
/atlas/changelog — model registry history
/sentinel/calibration — live model-honesty dashboard

Why three horizons?

Related