Predicting Arrest Outcomes and Mapping 25-Year Enforcement Patterns in Chicago (2001-2026): A Corrected Machine Learning, Longitudinal, and Geospatial Analysis

This paper presents two distinct contributions applied to the City of Chicago public crime catalog: 7,877,981 incident records across 22 attributes spanning 2001-2026. The first contribution is methodological: a fully corrected machine learning pipeline that eliminates four systematic errors present in prior analyses of this dataset. A controlled ablation study quantifies the largest error, target leakage from crime-type identifiers, showing XGBoost accuracy drops 17.25 percentage points when leaky features are removed (87.29% to 70.04%). After correction, Random Forest achieves 79.77% accuracy [95% CI: 79.72-79.82%] and XGBoost achieves 52.16% arrest-class F1. All pairwise differences are statistically significant (McNemar test, all chi-squared > 95,000, p < 0.001). Temporal validation on 2019-2026 data yields 78.30% accuracy and 35.56% F1, with rolling year-by-year analysis documenting concept drift from post-2020 policing changes. The second contribution is longitudinal: a 25-year characterization of enforcement patterns only possible at this temporal scale. Narcotic arrest ratios sustain 99.3% through 2019 then collapse following documented policy shifts. Sexual offense arrest ratios decline 50 percentage points monotonically from approximately 75% in 2001 to below 25% by 2025, a finding not previously characterized at this granularity. Theft clearance remains structurally stable at 9.1% across all 25 years. K-Means geographic clustering identifies four coherent spatial crime regimes.