基于机器学习的大豆粕期权收益预测、可解释性与经济价值研究——来自中国的证据

摘要

We examine how an expanded set of features, including local supply chain and weather data, affect delta-hedged soybean meal option returns in China. For the selected nonlinear ensemble, the full predictor set produces an out-of-sample R2 of 11.2%. Option characteristics contain most of the predictive information, while supply chain and weather variables provide additional information beyond option and futures predictors. Predictive performance is higher when geopolitical risk and investor attention are low. Weather information is particularly valuable during the local soybean growing season, when it delivers stronger out-of-sample predictive gains. Finally, we show that these forecasts can generate profitable trading strategies; however, this economic viability is highly conditional. Specifically, it heavily depends on applying strict threshold rules to manage market frictions.