機器學習在波動風險溢酬策略中的應用探討

Abstract

本研究旨在探討機器學習技術於波動風險溢酬（Volatility Risk Premium, VRP）交易策略中的應用，透過結合行為金融與總體經濟變數，以提升策略獲利表現與穩定性。本研究應用了兩種選擇權策略：Delta-hedged short straddle與Delta-hedged short put策略，並分別以「Raw VRP」與「Excess VRP」（扣除標普500報酬後之VRP）為預測目標，評估多種機器學習模型之預測效能與交易表現，包括Linear Regression、Random Forest與Gradient Boost模型。本研究納入市場情緒、意見分歧、動能因子、總體經濟指標與波動率等多元變數，並透過傳統Feature Selection方法及SHAP（SHapley Additive exPlanations）解釋性方法，分析模型在不同維度下之穩健性與敏感度。實證結果顯示，於兩類策略中，Delta-hedged short put策略相對具有較高之報酬潛力與風險調整後績效；而以Excess VRP為目標變數之模型，整體而言具備更高之穩定性與預測能力。研究結果顯示，行為與總體經濟變數可有效強化VRP策略之預測性，並提供實務上進行動態部位調整之依據；同時，本研究亦揭示在市場環境變遷下維持模型穩健性所面臨之挑戰。綜上所述，本論文提出一套結合可解釋性機器學習方法與Delta-hedged選擇權策略之整合性架構，對於風險溢酬策略之研究與實務應用具一定貢獻。

This thesis examines the application of machine learning models to predict and optimize the profitability of volatility risk premium (VRP) strategies through adaptive delta-hedging techniques informed by behavioral and macroeconomic indicators. Specifically, we construct and evaluate delta-hedged short straddle and put strategies on the S&P 500 index. By comparing two target formulations — raw VRP and excess VRP over the underlying index — we assess the predictive performance and trading efficacy of linear regression, random forest, and gradient boosting models. Our methodology encompasses comprehensive feature engineering, incorporating sentiment, disagreement, momentum, macroeconomic surprises, and volatility structure signals. We implement both traditional and SHAP-based feature selection techniques to evaluate the sensitivity of model performance to input dimensionality. Among the two strategy types, the put-only strategy displays higher return potential and stronger Sharpe ratios. At the same time, excess VRP emerges as a more stable and predictive target variable than raw VRP. The findings demonstrate the value of incorporating behavioral finance and macroeconomic insights into quantitative models, highlighting the practical challenges of maintaining model robustness in shifting environments. This study contributes to the growing literature on machine learning applications in asset pricing by proposing a framework that combines delta-hedged options trading with interpretable predictive modeling.