Firefighter Intervention Predictive Modeling: Reliability Assessment

This research contributes to a comprehensive exploration aimed at refining the assessment of predictive model reliability for firefighting interventions. Motivated by the recognition that predictive models, while facilitating intervention anticipation, may lack reliability in critical scenarios where prioritization is essential, we seek to develop a tool that provides a confidence index for predicted results. Given the urgency inherent in firefighting situations, our investigation delves into how predictive model reliability assessment influences firefighting interventions over an eight-year period from 2015 to 2023. The annual review of firefighters’ interventions in France has revealed particularly long response times in certain departments, underscoring the critical nature of our research. Our methodology encompasses data preparation, in-depth analysis, and the application of the resilient XGBoost predictive model known for its outlier robustness. The iterative training pipeline selects features that enhance the RMSE score over a 24-hour horizon. Subsequently, a thorough evaluation of prediction reliability is conducted through two approaches: the combination of the XGBoost and LightGBM models to measure their proximity and quantile regression to gauge the confidence interval width [0.05, 0.95]. The key conclusion drawn from this study is that the quantile regression approach is advantageous, particularly in terms of execution speed and reliability stability. This precision in understanding the models associated with predictive reliability diagnostics offers a practical approach to anticipate and optimize resource management, improve firefighters’ response times, and contribute to saving lives by mitigating intervention failures during major incidents.