強化學習下動態調整風險偏好之投資組合配置:以台灣50指數為例

Abstract

Markowitz(1952)提出現代投資組合理論，透過均數-變異數模型(Mean-Variance Model) 為投資人進行資產配置，並設風險趨避參數調整報酬率和風險之間的比例，但在實務中，此風險趨避參數難以動態調整。本研究使用強化學習(Reinforcement Learning) 中的近端策略優化(Proximal Policy Optimization,PPO)，依據不同市場變化，動態調整每一天風險趨避參數，當市場情況好時，投資人偏好承擔較高風險，獲得更高報酬，當市場情況壞時，投資人風險偏好趨於保守。本研究以台灣 50 指數當作整體市場走勢，比較強化學習輸入過去不同時間週期資訊之結果，研究結果顯示，不論輸入時間週期長短，強化學習績效皆能贏過固定風險趨避參數下均數-變異數模型，說明利用強化學習，能解決實務上風險趨避參數難以動態調整之問題。

Markowitz (1952) proposed Modern Portfolio Theory, which used the Mean-Variance Model to allocate assets for investors, and set the risk aversion parameter to adjust the ratio between return and the risk. But in practice, this risk aversion parameter is difficult to adjust dynamically. In our paper, we use Proximal Policy Optimization in reinforcement learning to dynamically adjust daily risk aversion parameters according to different market changes. In a bull market, investors prefer to take higher risks and get higher returns. On the other hand, in a bear market, investors` risk appetite tends to be conservative. This study uses the Taiwan 50 Index as the overall market trend, and compares the results of inputting different time periods of information into the model. The results show that regardless of the length of the input time period, the performance of the model can outperform the mean-variance model under fixed risk aversion parameters. Explain that the use of reinforcement learning can solve the problem of difficulty in dynamic adjustment of risk aversion parameters in practice.