Calibrated Machine Learning with Temporal Leakage Controls for Rare-Event Forecasting of Port Activity Anomalies: Evidence from Selected African Countries and Indonesia

Port activity anomalies are rare but may signal operational constraints relevant to supply chains. The aim of the study was to develop and empirically evaluate a calibrated machine learning framework for forecasting rare, algorithmically defined port activity anomalies over horizons of 1, 3, 7 and 14 days, under explicit temporal leakage controls and while assessing spatial transferability, the incremental value of port co-movement proxies and the operational utility of warnings. A total of 303,140 port–day observations were used for 115 ports in six selected African countries and Indonesia for the period 2019–2026. A seasonal benchmark, logistic regression and two variants of the Light Gradient Boosting Machine (LightGBM)—including one with a port co-movement proxy—were compared. Separate training, calibration and final-holdout periods were applied, along with isotonic calibration, port bootstrapping, transferability tests and threshold sensitivity analysis. In the main specification, logistic regression demonstrated a bootstrap-supported advantage for 1- and 3-day horizons, whilst differences between the models for 7- and 14-day horizons were not conclusive. A conservative rerun excluding 15 anomaly-derived predictors preserved the bootstrap-supported 1-day advantage, whereas the 3-day margin over the seasonal benchmark became inconclusive. The inclusion of a port co-movement proxy did not yield a confirmed improvement in predictive quality, whilst higher sensitivity required a marked increase in false alarms. Because the target is algorithmically defined and was not validated against an independent event registry, the results concern predictive signals of activity anomalies rather than confirmed port disruptions. The framework provides a transparent research benchmark but is not yet ready for autonomous implementation.