Momentum prediction method based on CUSUM and LightGBM-SHAP model

This paper introduces a novel approach for predicting momentum shifts in time-series data by integrating an enhanced Cumulative Sum (CUSUM) algorithm with a Light Gradient Boosting Machine (LightGBM) model, augmented with SHAP values. Our methodology refines CUSUM through a composite sliding window to improve detection accuracy and utilizes LightGBM for efficient prediction, with SHAP values providing interpretability of model predictions. Tested on a dataset of 1701 tennis matches from the 2023 Wimbledon Championships, our method demonstrated superior performance, achieving near-perfect prediction accuracy with an R2 score of 0.999, and significantly lower prediction errors (MAPE and RMSE) than traditional models. These results underline the method's effectiveness in not only accurately predicting momentum shifts but also offering detailed insights into the driving factors behind these changes. This advancement opens new avenues for time-series analysis across various domains, extending well beyond the initial sports analytics application.