以機器學習分類方法結合選擇權資訊優化美國股市投資組合

Abstract

本研究探討了納入選擇權資訊能否提升機器學習模型在預測個股未來報酬率方面的能力。結果顯示，當在已有的九個金融市場變數基礎上加入CP Spread、RI Spread 和Skew 等選擇權資訊後，模型的樣本外R 平方（R2 OOS）明顯提升。XGBoost 和LightGBM 模型在預測報酬率方面優於隨機森林和梯度提升迴歸樹，且單層類神經網路的R2_OOS 表現也較為優異。本研究比較了三種預測方法：報酬率數值預測、漲跌分類預測及本研究提出之TopMidBot 組別分類預測，結果發現後兩者能更有效地提供投資策略，特別是TopMidBot方法中的預測機率P(Bot)與TMB（P(Top) − P(Bot)）以及漲跌分類預測之預測上漲機率P(Up)能顯著辨識出負向三因子超額報酬，幫助優化投資組合的風險與回報。本研究提出的「去除高P(Bot)、低 TMB 個股或是低 P(Up) 個股」投資策略，相較於基準投資組合，能有效降低風險並提升夏普比率，特別在小型股與成長股提升幅度最為明顯。

This study investigates whether incorporating option Information can enhance the performance of machine learning models in predicting individual stock future returns. The results indicate that adding option Information such as CP Spread, RI Spread, and Skew to the existing nine financial market variables significantly improves the out-of-sample R-squared (R2_OOS) of the models. Among the models tested, XGBoost and LightGBM outperform Random Forest and Gradient Boosting Regression Trees in return prediction. The single-layer neural network also shows superior R2_OOS performance. This research compares three prediction methods: return prediction, stock price trend classification, and TopMidBot group classification we proposed. The findings reveal that the latter two methods provide more effective investment strategies. Notably, in the TopMidBot method, the predicted probability P(Bot), TMB (P(Top) − P(Bot)), and the price trend classification prediction P(Up) can significantly identify negative three-factor excess returns, thus helping to optimize the risk and return of investment portfolios. The proposed investment strategies—excluding stocks with high P(Bot), low TMB, or low P(Up)—effectively reduce risk and improve the Sharpe ratio compared to benchmark portfolios, particularly for small-cap stocks and growth stocks.