emperor-mew/voidly-forecast-v1-multi-horizon

Voidly Atlas Multi-Horizon Forecast v1

Version: v1 | Trained: 2026-05-21T03:40:41.231835Z | License: CC BY 4.0

Three separate XGBoost + isotonic-calibration models predicting the probability of a censorship event at horizons 1 day, 7 days, and 30 days. Each model has per-horizon SHAP top-5, a 90% conformal interval, and a monotonicity-consistency check across horizons.

Eval (leave-one-country-out)

| Horizon | AUC | F1 | Brier | n_features |

|---------|-----|----|-------|------------|

| 1d | 0.9074 | 0.4481 | 0.0278 | 39 |

| 7d | 0.8828 | 0.6597 | 0.0551 | 39 |

| 30d | 0.8449 | 0.7012 | 0.1077 | 39 |

Decision: ship all three (per-horizon honest gate all green).

Intended use

GET https://api.voidly.ai/v1/forecast/{cc}/multi-horizon returns { 1d, 7d, 30d } calibrated probabilities + 90% conformal intervals

• per-horizon SHAP top-5 contributors. Legacy single-horizon endpoint /v1/forecast/{cc}/7day remains backwards-compatible.

Honest caveats

• 30-day F1 is strongest (0.70) — but Brier triples vs 1d (0.028 → 0.108), so 30-day probabilities are noisier.
• AUC monotonically decreases with horizon, as expected (1d > 7d > 30d).
• Conformal coverage tracked online via Adaptive Conformal Inference (Gibbs & Candès 2021) — see voidly-forecast-aci-v1 card.
• 20 spotlight countries — performance on long-tail countries unverified for this multi-horizon variant.

Reproducibility

python3 scripts/build-forecast-multi-horizon-labels.py

Citation

@misc{voidly_voidly_forecast_v1_multi_horizon,
  title  = {Voidly Atlas: voidly-forecast-v1-multi-horizon (v1)},
  author = {Voidly},
  year   = {2026},
  url    = {https://huggingface.co/emperor-mew/voidly-forecast-v1-multi-horizon},
  note   = {Open censorship-research ML stack. CC BY 4.0.}
}