Incident Duration Reliability Assessment Using Monte-Carlo Simulation and Kernel Density Estimation of Machine Learning-Based Models

Abstract

Traffic incidents are a major cause of non-recurrent congestion and delays, making accurate incident duration (ID) prediction essential for effective traffic management. While machine learning (ML) and deep learning (DL) models have been developed to predict ID and its sub-periods (verification, response, and clearance times), their reliability has not been systematically assessed. This study introduces a novel framework to evaluate the reliability of these predictions using 4,000 traffic incident records. Non-parametric Kernel Density Estimation effectively captured variations in the data, outperforming traditional parametric methods. Monte Carlo simulation was then used to assess prediction reliability. Among the models tested, bagged ensemble trees provided the best balance between accuracy and complexity, showing strong reliability for predicting total ID and sub-periods. Adding 5% to 25% buffer adjustments further improved reliability by accounting for prediction uncertainties. This framework offers a robust tool for assessing prediction reliability, is adaptable to various ML and DL models, and represents a significant step forward in traffic incident management.