Abstract
<title>Abstract</title> <p>Kenyan magistrate courts resolve more than eighty per cent of the country's judicial business, and they carry a correspondingly large share of its backlog. The Judiciary's Integrated Case Management System records what has happened to a case but cannot indicate which newly filed cases are likely to become delayed. This study develops and validates a three-stage machine learning pipeline that supplies that missing capability, using a census extraction of 744,340 closed cases filed between January 2011 and March 2026, of which 672,974 were retained after validation. Because 28.98 per cent of cases were resolved on their filing day, a same-day disposal classifier was introduced as the pipeline's first stage, followed by a filing-time duration model and a model that revises its estimate as procedural history accumulates. Random Forest, XGBoost, LightGBM and CatBoost were tuned under grouped five-fold cross-validation and compared at every stage. LightGBM performed best at same-day triage (ROC-AUC 0.9148) and at filing-time prediction (RMSE 444.25 days, R² 0.616), explaining approximately sixty-two per cent of duration variance from information available at filing alone. CatBoost performed best on the chronologically held-out progression test (RMSE 80.83 days). SHAP analysis identified caseload per magistrate, legal representation, case nature, first-mention delay and accumulated hearings as the dominant drivers of delay. An empirically derived four-tier risk scheme translates predictions into administrative categories, and the models are served to court users through a working web application.</p>