Machine Learning for Business Process Risk Prediction and Operational Performance Optimization

Abstract

This study develops and evaluates a machine learning framework for predicting business process delays and supporting risk-based operational prioritization under limited review capacity. Using the BPI Challenge 2017 event log, comprising 31,509 loan-application cases and 1,202,267 events, the study applies chronological, case-disjoint validation to assess predictive performance. Following data screening, 16,481 training, 4,025 validation, and 6,275 testing cases were selected for event-prefix analysis. Prolonged processing was defined using the 75th-percentile duration threshold (30.9 days) derived exclusively from training data. Logistic regression, random forest, and histogram gradient boosting models were evaluated against an elapsed-time baseline. The validation-selected logistic regression model achieved an average precision of 0.394 (95% CI: 0.371–0.414) and ROC-AUC of 0.686 on the independent test set. Under a simulated 10% review-capacity constraint, risk-based prioritization identified 17.3% of prolonged cases, compared with 16.4% using elapsed-time ranking and 10.1% through random selection. Although the improvement over the elapsed-time baseline was statistically inconclusive, the findings demonstrate the feasibility of predictive analytics for retrospective business process risk assessment. The study provides a reproducible framework for operational decision support while highlighting the need for prospective validation before claims of improved process efficiency can be established.