運用深度學習建構個股預測模型
-以台積電為例

在現今這個普遍低薪水高物價的環境中，投資也成為現在這一世代必學的必修課。如今，各種大數據與人工智慧的演算法應育而生，也讓金融領域的預測成為時下熱門的研究議題之一。而股市資料屬於時間序列型資料，且過去研究通常使用傳統倒傳遞類神經網路為預測模型，因此本研究分別比較倒傳遞類神經網路(BPN)、時間序列倒傳遞類神經網路(TS-BPN)與深度學習中的遞歸神經網路長短期記憶模型(RNN-LSTM)，來探討何種類神經網路對時間序列資料能有更好的學習與預測。而在訓練類神經網路時，模型的學習參數是左右模型訓練好壞的關鍵，所以本研究除了使用田口法規劃實驗組合外，也以田口法為基礎延伸為改良田口法，探討參數水準層外之新參數解，並分別比較何種參數組合優化方法能找出更好的學習參數組合，提高績效。最終實驗結果表示，本研究延伸之改良田口法也優於傳統田口法，除了能有效取代試誤法減少實驗組合外，也能有系統地找出學習參數水準層移動方向，找出新學習參數水準層解。而RNN-LSTM在預測股市時間序列資料上有最好的預測能力，漲跌方向能達到81.9%，判斷模型誤差績效的RMSE只有0.007505，為本研究綜合績效最好的預測模型。

關鍵字

深度學習；長短期記憶網路；倒傳遞類神經網路；股市預測；特徵篩選；田口法

並列摘要

The data of the stock market belongs to a time series data, the research in the past usually use traditional backed propagation neural network as predictive model. Therefore, this study compares the "Back Propagation Neural Network (BPN)" and the improved time series data in this study. The "Time Series Back Propagation Neural Network (TS-BPN)" and "Recurrent Neural Network Long Short Term Memory Model (RNN-LSTM)" in deep learning, are being used to determine the suitable type of neural networks that can learn better, and predict the time series data. In the training of neural networks, the learning parameters of the model are the key to determining the quality of the model training. Therefore, the Taguchi method is also being used to plan experimental combinations, and it would be extended to the improved (Taguchi) method. Based on the Taguchi method, exploring new parameter solutions outside the parameter level, and comparing the different types of parameter combination optimization methods, can find better combinations of learning parameters and improve performance. The final experimental results show the improved Taguchi method extended in this study, is also better than the traditional Taguchi method. Also, the RNN-LSTM has the best performance in predicting stock market time series data, the ups and downs can reach 81.9%. The model error of the RMSE is a minimal 0.007505, which is the best predictive model for the comprehensive performance.