Prediction of Closing Prices for NASDAQ Listed Stocks: A Comparative Study Based on Gradient Boosting Models

Abstract

The forecasting of trends in U.S. stocks remains a focal point in the financial domain. Within the realm of fine-grained forecasting tasks, predicting price trends with pronounced volatility in the last ten minutes of the closing cross-auction presents a distinctive challenge. This study develops four forecasting models—LightGBM (LGB), XGBoost (XGB), CatBoost (CBT), and a weighted fusion model—utilizing data from the final-market order book and closing auction of 200 NASDAQ stocks. The research aims to conduct a comprehensive comparison of the performance of these four models on the same prediction task. The study results reveal Mean Absolute Error (MAE) values of 5.195, 5.197, 5.200, and 5.202 for the respective models, signifying the integrated model as the most precise predictor. This comparative analysis facilitates a nuanced understanding within the research community regarding the performance of each model, offering a valuable reference for future investigations. Simultaneously, it underscores the inherent value of machine learning in financial forecasting, aiding investors in enhancing decision-making efficiency during critical moments on trading days.