Pakistan (PK)
7-day internet-shutdown probability from shutdown_risk_v9. Trained and validated on Access Now KeepItOn STOP shutdowns (journalist-verified). Live JSON: /v1/shutdown-risk/PK.
Current 7-day probability
23.5%
Risk band
elevated
As of
2026-06-13
Elevated threshold
18.0%
KIO shutdowns (5y)
19
verified count
KIO recent (7d)
0
last 7 days
Probability components
The live model combines a structural-risk score (v4 = isotonic(kio_hist_5y + 2 × kio_recent_7d)) with the OONI-trajectory forecast (honest_forecast_v1); the served version is shown above. A logit-space blend (the v9 variant, sigmoid(0.6·logit(v4)+0.4·logit(hf))) that lifts within-country median AUC ~0.73 → ~0.74 is validated and auto-gated — it serves only on days it durably beats the live model, leave-one-country-out validated.
- ensemble_raw
- 0.381
- hf_prob
- 73.4%
- sr_prob
- 18.5%
Honest validation
Same validation block this endpoint returns inline — never computed at request time, never adjusted per-country.
Full-panel AUC
0.897
cross-country
Within-country median AUC
0.735
23 countries
Within-country mean AUC
0.625
paired with median
Countries ≥ 0.65
13 / 23
within-country AUC
Model honesty caveats (5)
- FORWARD-VALIDATED (the real product): in a strict past->future temporal holdout the CROSS-COUNTRY ranking holds — full-panel AUC ~0.87, country-ranking AUC ~0.90, 7.8x PR-lift over base rate. The model reliably tells you WHICH countries are heading into danger, forward in time. BUT the within-country median AUC is CROSS-SECTIONAL (full panel, in-sample isotonic) — it measures day-ranking when the model has seen the whole period, NOT forward skill. In a strict past->future temporal holdout, within-country timing AUC is ~0.38 (below chance), for BOTH v5 and v9, driven by non-stationarity (shutdowns cluster in time). USE THIS MODEL FOR *WHICH* COUNTRY IS AT RISK (cross-country full-panel AUC 0.90, which IS forward-meaningful), NOT as a within-country *WHEN* oracle.
- v9 replaces v5's multiplicative product with a logit-additive blend (alpha=0.6 on structural risk). The product collapsed to ~0 when honest_forecast was near-zero for a country, annihilating a good structural signal; the logit blend cannot zero out either factor.
- Within-country median AUC 0.735 vs v5 0.729. Improvement confirmed by leave-one-country-out alpha selection (all 23 folds chose alpha=0.60; held-out median +0.0095, mean +0.0131, 18/23 countries up).
- Still country-level; sub-national shutdowns (~half of KeepItOn events) remain harder. Low-N countries are sample-size limited, not model-limited.
- Same two honestly-validated base models as v5 (v4 structural + honest_forecast_v1). No new external features (unlike the gated v7 CF-Radar and v8 GDELT attempts) — purely a better combiner.