本論文分為兩個主軸,第一個主軸透過 panel-data regression 的方式針對台灣上市、上櫃公司進行 Fama–French 三因子模型驗證,探討三個風險因子在不同時間點下對於各投資組合股價報酬的影響,而本論文實驗結果表明三個風險因子在日資料上對股票報酬有顯著的影響;第二個主軸則藉由 panel-data regression 發展出一套當沖的交易策略,針對偏離均衡價格的股票進行操作以及進一步透過 LightGBM 模型針對前述交易策略進行優化,其精神在於將風險因子的估計係數、基本價量指標以及技術分析指標作為模型所需的因子並預測隔天股票的漲跌,因此發展出原始策略與機器學習輔助策略兩種策略模式,並形成一共120種的交易策略。此外,經由回測結果發現,某些策略的績效比台灣 50 ETF 來得好且機器學習輔助策略的模式只適用於小型公司投資組合的策略。
This thesis is divided into two parts. Focusing on Taiwan-listed companies and over-the-counter market companies, the first part uses panel-data regression to verify the Fama–French three-factor model. It discusses the impacts of the three risk factors on stock returns for each portfolio over the period June 4, 2014, through May 31, 2019, and the experimental results show that the three risk factors have a significant impact on daily stock returns. The second part explores daily trading strategies that take advantage of deviations from a stock’s equilibrium price established by panel-data regression, such trading strategies may also be improved through the LightGBM model, which uses the estimated coefficients of risk factors, basic price-volume information, and technical analysis indicators to predict stock’s rise or fall the next day. Finally, a total of 120 trading strategies will be generated with or without the help of machine learning. The backtesting results show that some of the trading strategies beat the Taiwan Top 50 ETF, and machine learning only helps the strategies that trade only small-company stocks.