Multi-horizon shutdown forecast

ML — three honest horizons

How the current-regime shutdown-risk level for ML reads over 1 day, 7 days, and 30 days. Each horizon uses its own XGBoost model with per-horizon SHAP attribution and a 90% conformal interval. These are current-regime risk signals, not onset predictions — see the scope note below.

Generated 2026-05-22 20:30 UTC · model multi-horizon-h-v1 · raw JSON

Probabilities

1 day

27.1%

90% interval [17%, 37%]

7 days

58.8%

90% interval [40%, 78%]

30 days

30d

70.0%

90% interval [20%, 100%]

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 →

Consistency check

OK — monotonicP(1d) ≤ P(7d) ≤ P(30d)

The three horizons are internally consistent: longer windows capture at least the risk of shorter windows. This is the expected behavior — a 30-day window must contain any 7-day or 1-day shutdown.

27.1% ≤ 58.8% ≤ 70.0%

Top-5 SHAP features per horizon

Each horizon has its own learned drivers. Positive SHAP values push the prediction up; negative values push it down. Features differ across horizons because near-term risk responds to acute signals, while month-out risk responds to slower structural factors.

Horizon	Rank	Feature	SHAP
1-day	#1	week_of_year	-0.688
	#2	risk_tier	-0.244
	#3	recent_shutdown	+0.155
	#4	is_friday	-0.139
	#5	day_of_week	-0.080
7-day	#1	risk_tier	-0.441
	#2	week_of_year	-0.364
	#3	recent_shutdown	+0.290
	#4	month	-0.076
	#5	gdelt_unrest_30d	-0.073
30-day	#1	week_of_year	-0.964
	#2	risk_tier	-0.647
	#3	recent_shutdown	-0.126
	#4	gdelt_unrest_30d	-0.069
	#5	month	-0.048

Honest caveats

Each horizon is a separate XGBoost model with its own features and its own honest evaluation. LOCO AUC across spotlight countries: 1d 0.91 · 7d 0.88 · 30d 0.84.
30-day forecasts inherently have wider conformal intervals (more time = more uncertainty). A wide interval is not a bug — it's honesty.
These numbers are not guarantees. A 70% probability does not mean a shutdown is certain; a 20% probability does not mean you're safe.
See /atlas/multi-horizon for the cross-country index, or /atlas/models for the full model registry.

/atlas/forecast/ml — 7-day calibrated forecast with drivers
/ml — ML country profile
/atlas/multi-horizon — all spotlight countries with 30d probability
/v1/forecast/ML/multi-horizon — raw JSON

Probabilities

Consistency check

Top-5 SHAP features per horizon

Honest caveats

Related