voidly

ChinaCN

falling (-8.0pp)

30-day day-by-day shutdown-risk trajectory. Each point is a per-horizon XGBoost prediction (h=1..30) with a 90% conformal band derived from absolute LOCO residuals. Peak risk highlighted.

Peak day
51.7%
d10 · 2026-06-18
Floor day
7.3%
d16 · 2026-06-24
30-day delta
-8.0pp
d30 vs d1 probability

30-day trajectory with 90% conformal band

model trajectory_v1
0%25%50%75%100%d1d7d14d21d30peak d10 52%
mean P(shutdown by day)90% conformal bandpeak day

Drivers (SHAP top 3 from the 7-day model — same features feed all horizons)

day 120.3%
2026-06-09
  • block_rate_roll7_std-0.011
  • gdelt_unrest_30d-0.008
  • high_urgency_signals_7d-0.005
day 712.6%
2026-06-15
  • block_rate_roll7_std-0.011
  • gdelt_unrest_30d-0.008
  • high_urgency_signals_7d-0.005
day 1417.5%
2026-06-22
  • block_rate_roll7_std-0.011
  • gdelt_unrest_30d-0.008
  • high_urgency_signals_7d-0.005
day 3012.2%
2026-07-08
  • block_rate_roll7_std-0.011
  • gdelt_unrest_30d-0.008
  • high_urgency_signals_7d-0.005

Honest caveats

  • Per-horizon XGBoost stacked into a trajectory — not a true joint distribution.
  • Conformal bands are absolute-residual q90 from LOCO predictions on 18 eval countries.
  • Long horizons (h>=14d) have wider effective uncertainty than the conformal q90 suggests; the same lag features feed all horizons.
  • Bands floor at [0,1]; probabilities below 0.05 are at the model's noise floor.
  • SHAP attribution above is from the 7-day point model. The trajectory uses 30 separate per-horizon XGBoost models that share the same feature set; we have not yet computed per-horizon SHAP.

Sample horizons

DayDateP(shutdown)90% interval
d12026-06-0920.3%[0%, 64%]
d72026-06-1512.6%[0%, 77%]
d142026-06-2217.5%[0%, 86%]
d302026-07-0812.2%[0%, 91%]
Spotlight countries