Why hourly

Until today, the only shutdown forecast Voidly published was daily — a single risk number per country-day. That works for the common case (e.g. Iran has 35% risk this week) but throws away an obvious axis of structure: shutdowns happen on specific hours. Election-morning DNS blocks, midnight protest-suppression cuts, a 06:00 UTC fiber-restoration window — the daily forecast collapses 24 hours of inside-day variance into a single bucket.

We trained three within-day XGBoost classifiers, one per horizon: K = 6, 12, 24 hours. Each predicts P(at least one censorship-class incident starts in the next K hours) at hourly granularity. Live at GET /v1/forecast/{cc}/hourly?horizon=6|12|24.

Features

Fifteen-dimensional, all observable on the country panel without leakage:

• hour-of-day (raw + cyclical sin/cos)
• day-of-week (raw + cyclical sin/cos) + weekend flag
• n_evidence and n_block in the trailing 6h and 24h windows
• block_rate (block / evidence) for 6h and 24h
• asn_diversity in 24h (distinct ASNs touched)
• country_base_rate — leak-free positive rate computed on training half only

Training panel

90-day evidence × 45 countries × 24 hours = 97,245 cells. Drop cells with no recent evidence (MIN_EVIDENCE_24H ≥ 1) → 75,570 training rows. Country set is the union of every country with a confirmed censorship/mixed incident in the window (36) plus the top-30 by evidence volume.

Positive rates: y_6h 1.3%, y_12h 2.6%, y_24h 5.1%. Imbalanced as expected — most hours, in most countries, are uneventful.

Honest caveats — read these before citing

• About 50% of evidence rows cluster at hour 00 UTC because OONI and CensoredPlanet land at daily granularity. The model still uses them but the hourly precision is bounded above by what the sources can reveal. Treat the K=6 number as an UPPER BOUND on what's achievable from this data, not ground truth.
• Incident first_seen is itself often midnight UTC on the day the event landed in our pipeline — so the y target inherits the same coarseness as the features.
• Conformal intervals are not yet wired — current p_low/p_high use a static ±0.15 halfwidth around the calibrated probability. ACI-style adaptive width is the next step.

Evaluation

Temporal holdout: the most recent 30 days (April 21 → May 21 2026) is held out; everything earlier is training. The model has never seen any country's hourly observations after the cutoff. Per-country AUC is graded only on countries whose holdout has both classes AND where the model produced a non-degenerate ranking (AUC ≠ 0.5 exactly — those ties come from flat predictions on country windows where the model can't separate hours, which is real but not informative for ranking quality).

All three horizons clear the 0.60 per-country AUC floor. The promote rule is intentionally lower than for daily (0.70) because hourly granularity is noisier and the labels themselves are coarser.

Per-country, K=6h

Where the model has strong inside-day signal:

• QA — 0.987
• TZ — 0.921
• JO — 0.906
• UZ — 0.882
• AZ — 0.778
• PK — 0.721

Where the model is at or below chance:

• BD 0.496, EG 0.494, KZ 0.492, TN 0.486, BY 0.479

The pattern is clear: countries whose shutdowns concentrate at specific hours (likely tied to election windows, prayer time, or protest hours) are predictable. Countries whose shutdowns are spread evenly across the day are not — and the model honestly admits it.

Endpoints

• GET /v1/forecast/{cc}/hourly?horizon=6|12|24
• GET /v1/forecast/hourly/info — full sidecar metrics + honest caveats

Response includes probability, p_low, p_high, the top-5 features ranked by gain × live value, and the full honest_caveats array on every call. Countries with no recent evidence (MIN_EVIDENCE_24H == 0) get a 404 + an explicit refusal message — the model does NOT score no-data countries.

What we are NOT claiming

• This does not replace the daily 7-day forecast for journalist or policy-research use. Daily is still the right granularity for most reporting; hourly is for systems needing finer windows (rate-limiting, circumvention scheduling, in-day alerting).
• We are not claiming that's the "true" hourly probability — source granularity is the ceiling here, not model quality. If you ingest finer-grained probe data into your own copy of the panel, you should retrain.