Return predictability in equity options : a supervised machine learning analysis of delta-hedged call option returns

Abstract

This thesis aims to explore whether machine learning models can predict cross-sectional variation in next-day delta-hedged returns for at-the-money (ATM) call options and whether these predictions translate into economically exploitable information. The empirical work depends on a snapshot dataset constructed from two consecutive trading days for a group of actively traded U.S. stocks. The analysis integrates option, stock, and market-level variables. The study evaluates three linear models (OLS, Ridge, LASSO) and two nonlinear models (Extreme Gradient Boosting and a feed-forward Multilayer Perceptron using an 80/20 train-test split with nested five-fold cross-validation. The findings indicate that nonlinear models, particularly XGBoost, outperforms linear models, achieves highest out-of-sample accuracy and shows stable performance under liquidity-based robustness checks for this study sample. To assess the economic usefulness of the XGBoost model, predicted returns were sorted into quantiles to examine whether the model establishes a monotonic relationship between predicted and realised delta-hedged returns. The monotonic ordering test confirmed that higher predicted returns are systematically associated with higher realised returns, reflecting meaningful cross-sectional ranking ability. Interpretability for the model is examined using Shapley Additive Explanations (SHAP), which identify the most influential predictors and reveal that gamma, realised volatility, option price, and implied volatility are primary drivers of predicted return, while lower order Greeks and market variables contribute minimally. Overall, the thesis demonstrates that machine learning can uncover short-horizon predictability in delta-hedged call option returns and that these predictions contain economically exploitable information within the limitations of the snapshot-based setting.