基於行為經濟學與價量分析使用增強式學習演算法建立臺灣股票指數期貨交易策略

本研究試圖基於行為經濟學及價量分析理論對於金融交易市場的動態行為進行分析，並試圖使用機器學習演算法建立交易策略。我們先試圖從事件研究中，尋找合適的進場時機，在此我們使用了K-means進行對價格與成交量進行分群，基於行為經濟學市場過度樂觀及過度恐慌的理論，將帶量大漲及帶量大跌視為事件，並且在分群前對成交量的正規化計算時，設計不同的實驗，分析其找到的事件前後的價格變動情形。而在出場時機的部分，我們採用了增強式學習演算法，增強式學習的概念是基於對環境的觀察與環境互動取得報酬，並且具有延遲報酬的特性，我們將此演算法對映到了交易策略的決策上，希望訓練出一個模型，尋找到好的出場時機，在此我們針對增強式學習中的Observation設計實驗，分析不同的Observation對於模型的影響。最後我們會對我們的建立交易策略及買進持有策略（Buy and Hold）進行回測，並且比較兩者的績效表現。

關鍵字

交易策略；臺股期貨；增強式學習； Q-learning ；機器學習；行為經濟學；事件研究

並列摘要

This study attempts to analyze the dynamic behavior of the financial trading market based on behavioral economics and price analysis theory, and attempts to establish trading strategies using machine learning algorithms. We first tried to find suitable entry opportunities from the event study. We used K-means to cluster prices and volume, and based on the theory of excessive optimism and excessive panic in the behavioral economics market, we took the increase in volume and the fall in volume as an event, and we normalized the volume before clustering. In the normalization calculation, different experiments were designed to analyze the price changes before and after the events they found. In the part of the timing of closing the position, we choose reinforcement learning algorithm. The concept of reinforcement learning is based on observation of the environment and interaction with the environment to obtain reward, and has the characteristics of delayed reward. We have mapped this algorithm to the trading strategy. In the decision-making, we hope to train a model and find a good time for closing the position. We focus on the Observation design experiments in the reinforcement learning and analyze the effects of different observations on the model. Finally, we will backtest our established trading strategies and buy and hold strategies and compare their performance.