遷移學習對股價預測的影響：以台股為例

Abstract

深度學習近年來被廣泛地應用於多個領域，但是使用深度學習時需要較為大量的訓練資料，因此若想將深度學習應用於對新上市的股票進行股價預測必須解決資料量不足的問題。 遷移學習常被用於處理訓練資料不足的問題，但在金融領域中較常用於文本分析，部分對於市場指數或股價漲跌趨勢預測的研究又缺少使用遷移學習前後的模型表現比較，因此無法說明遷移學習在金融領域中對於模型表現的影響。本論文以台股為對象，嘗試以不同實驗觀察遷移學習是否能緩解新上市股票資料不足的問題。根據本論文的實驗結果，遷移學習普遍而言可以提升對新上市股票的股價預測表現，但部分結果也說明並非選擇任何股票作為模型預訓練的資料在遷移學習上都能有好表現，因此需謹慎選擇。

Although deep learning has been widely applied in many fields recently, it needs plenty of training data, thus it is difficult to apply to stock price prediction for newly listed companies due to lack of historical data. Transfer learning is a common technique to deal with the problem of scarcity of data. But in financial domain, it is often applied to textual analysis. Some studies on market index or stock trend prediction that apply transfer learning do not compare the performances with and without transfer learning, making the net effects of transfer learning hard to isolate. In this thesis, we design experiments to track the effects of transfer learning on model performances in Taiwan stock market. The results reveal that transfer learning can generally raise the performance on predicting prices of newly listed stocks, but some special cases also demonstrate that which stock is used for model pre-training before applying transfer learning has significant impacts on the performance.