Voidly
All findingsPublished research

Predictive shutdown contagion chain: pairwise XGBoost classifier (0.67 primary AUC)

New predictive piece in the Atlas stack: given country A had a confirmed censorship event today, score every other watched country by P(follow within N days). Pairwise XGBoost on (trigger, follower, horizon) tuples; 0.67 held-out AUC at 7d primary horizon, 0.69 at 14d, 92 trigger-follower pairs…

Read the full published summary

New predictive piece in the Atlas stack: given country A had a confirmed censorship event today, score every other watched country by P(follow within N days). Pairwise XGBoost on (trigger, follower, horizon) tuples; 0.67 held-out AUC at 7d primary horizon, 0.69 at 14d, 92 trigger-follower pairs cross the P>0.30 significance line. Lead-lag is descriptive (who is correlated), causal forest is causal (election ATE) — this is the missing PREDICTIVE piece. Trained 2024-2026 on 311 events across 34 candidate countries; held-out 45-day test window. Top chains concentrate in MENA → Central Asia (OM/MA/EG/SY → JO/UZ/AZ/PK), consistent with the lead-lag findings and a regional regime-imitation prior.

What we shipped

A predictive shutdown contagion / chain model — given that country A had a confirmed censorship event today, score every other watched country by the probability that it will follow with its own event within N days.

This closes a missing piece in the Atlas measurement stack: we already had lead-lag cross-correlation (descriptive — who’s correlated with who, at what lag) and DTW cohorts (similarity by shape). What we did not have was the predictive piece: a model you can actually call on a fresh trigger event to get a ranked list of likely followers.

How it works

A pairwise XGBoost classifier on (trigger, follower, horizon) tuples. For each confirmed-censorship event in country A at time t, we generate one row per candidate follower country B and per horizon H ∈ {3, 7, 14} days. The label is did_follow = 1 if B had any censorship/mixed event in (t, t+H].

Features computed strictly from data before t:

  • Trigger / follower country embeddings (label-encoded categorical).
  • Regional & regime grouping — same_region (MENA, E_EUR, C_ASIA, etc.), same_auth_cluster.
  • Rate features — trigger 30-day events, follower 30 / 90-day events, follower base rate, global 30-day events.
  • Recency — days_since_follower_last_event (capped at 365).
  • Seasonality — month_sin / month_cos.
  • Pair coevent history — count of times both countries had events within a ±14-day window in the prior year (the closest thing to a learned “contagion prior”).

Trained on 2024-01-01 onwards (signal density explodes in 2024+; pre-2024 history is too sparse). Temporal holdout: last 45 days of trigger events. 17,622 train rows / 9,306 test rows, 12% / 25% positive rates.

Results — held-out AUC

HorizonAUCAvg precisionBriern_testPos rate
3d0.64570.3500.1413,10216.8%
7d (primary)0.66970.4930.1863,10225.4%
14d0.69200.5970.2273,10233.5%
Overall (stacked)0.69840.5130.1859,30625.2%

Promote gate: primary AUC ≥ 0.65 AND at least 10 trigger-follower pairs with predicted P > 0.30. Passed: AUC 0.67 at 7d, 92 significant pairs.

Top predicted contagion chains (held-out)

Highest-scoring (trigger → follower) pairs at 7-day horizon on the test set, with the actual hold-out follow rate to ground-truth the score:

TriggerFollowerP(follow within 7d)Actual follow rate
OMJO0.9901.00 (2/2)
OMPK0.9441.00 (2/2)
MAJO0.9360.67 (2/3)
OMAZ0.9261.00 (2/2)
MAUZ0.9220.67 (2/3)
NIAZ0.9211.00 (2/2)
OMUZ0.9111.00 (2/2)
EGAZ0.8900.80 (4/5)

How to call it

  • Live ranked-follower scorer: GET /v1/atlas/contagion-chain/{trigger}?horizon=7 — pulls the last 365 days of events from the Voidly DB, builds features as if the trigger fired today, returns ranked candidate followers.
  • Query-string variant: GET /v1/atlas/contagion-chain?trigger=RU&horizon=7
  • Sidecar / model meta: GET /v1/atlas/contagion-chain/info

Honest caveats

  • This is a temporal point-process surrogate, not a Hawkes process. With only 311 events on the training window across 34 countries, a full Hawkes self-/cross-exciting fit (~50² × 3 horizon parameters) would be hopelessly over-parameterized. The pairwise classifier substitutes regularized gradient boosting for parametric kernels — it captures the same structural insight (lift in intensity after a trigger) without the data-hungry kernel fit.
  • did_follow = 1 includes any coincidental event. The model can only learn that the joint event distribution shifts after the trigger, not establish causal contagion. Use these scores as a forecast prior, not a causal claim — pair with /sentinel/attribute (synthetic difference-in-differences) if you need causality.
  • High-confidence pairs are sparse-support. OM → JO at P=0.99 is based on just 2 trigger events in the held-out window. Treat the top-5 list as qualitative regime evidence (MENA / Central Asia tend to cluster temporally), not as a tightly calibrated probability.
  • Cold-start triggers fail. A trigger country with no events in the candidate set returns an error (we restrict to 34 countries with ≥3 confirmed events 2024-2026).
  • Calibration is not isotonic-refit. The raw XGBoost output is used as the rank score. Brier scores (0.14–0.23) indicate the rank ordering is more trustworthy than the absolute probability.

How this complements the rest of the stack

EndpointWhat it answersType
/v1/atlas/lead-lag Who’s correlated with whom, at what lag? Descriptive
/v1/atlas/cohorts Which countries have the most similar incident shape over time? Descriptive (DTW)
/v1/sentinel/hte What’s the average treatment effect of an election on shutdown risk? Causal (causal forest)
/v1/atlas/contagion-chain Given a trigger event today, which countries are most likely to follow within N days? Predictive
/v1/sentinel/attribute Did THIS specific outage on country X at date Y reflect a real anomaly vs comparable democracies? Counterfactual (SDID)

The source record.

Original source links and labels. Linked responses may have changed since publication.