An innovative method for accurate NBA player performance forecasting and line-up optimization in daily fantasy sports

Abstract

This study presents a novel approach for predicting NBA players' performance in Fantasy Points (FP) by developing individualized models for 203 players, using advanced basketball metrics from season 2011–2012 up to season 2020–2021 from reliable sources. A two-step evaluation and validation process secured validity, while applying linear optimization methodology, considering constraints such as salary and player position to recommend an eight-player line-up for Daily Fantasy Sports (DFS). Four scenarios with 14 machine learning models and meta-models with a blending approach with an ensembling methodology were evaluated. Using individual per-player modeling, standard and advanced features, and different timespans resulted in accurate, well-established, and well-generalized predictions. Standard features improved MAPE results by 1.7–1.9% in the evaluation and 0.2–2.1% in the validation set. Additionally, two model selection cases were developed, with average scoring MAPEs of 28.90% and 29.50% and MAEs of 7.33 and 7.74 for validation sets. The most effective models included Voting Meta-Model, Random Forest, Bayesian Ridge, AdaBoost, and Elastic Net. The research demonstrated practical application using predictions in a real-life DFS case evaluated in a DFS tournament on a specific match day. Among 11,764 real users, our Daily Line-up Optimizer ranked in the top 18.4%, and profitable line-ups reached the top 23.5%. This unique approach proves the proposed methodology's effectiveness and emphasizes its profitability, as the optimizer process delivers positive results.