Multi-Horizon Rearrest Prediction

Calibrated probability models for rearrest prediction at 1, 2, and 3-year horizons with comprehensive fairness diagnostics. Built on the NIJ Recidivism Forecasting Challenge dataset (~18,000 Georgia parolees).

18K
Parolees
3
Horizons
0.188
Best Brier (Y1)
0.70
AUROC (Y1)

Key Findings

Calibrated predictions work

Best Y1 Brier score is 0.188 (vs base-rate 0.209). Platt-calibrated XGBoost produces well-calibrated probabilities across race and gender subgroups.

Model performance

Fairness gaps are threshold-dependent

Race FPR gap at t=0.5 is 0.002, but peaks at 0.116 near t=0.25. Choosing a threshold changes who bears the errors.

Fairness analysis

Gang affiliation is strongest predictor

SHAP importance 0.163. Associated with +19.1 pp rearrest delta. Age 18-22 (+12.2 pp) and supervision risk score also dominant.

Predictive factors

At a Glance

Brier Score by Model and Horizon

Lower is better. Y2/Y3 are conditional on non-rearrest at prior horizons.

Explore

Model Performance

XGBoost leaderboard, baseline comparisons, seed stability, and COMPAS benchmark.

View models

Fairness

Threshold sweeps, subgroup calibration, and error-rate parity analysis across race and gender.

View fairness

Predictive Factors

SHAP values, feature importance rankings, and subgroup rearrest-rate breakdowns.

View factors

Methods

Data pipeline, feature engineering, model training, calibration, and evaluation methodology.

View methods

FAQ

Common questions about the dataset, models, fairness metrics, and limitations.

View FAQ