以星狀生成對抗網路(STARGAN)解決股票量價聚合關係預測股票價值之研究:兼論以實驗計畫法調整超參數

Abstract

股票市場中量與價具有聚合（joint）作用，但傳統之統計數值分析模型無法直接同時考慮兩者關係，僅能將其分開處理。本研究利用星狀生成對抗網路(Star GAN)多面向轉換的優點將證券的量價關係做結合，能夠建立模擬證券市場狀況的創新模型。我們繼續將Star GAN輸出的隱含量價之資料輸入常用於預測的LSTM模型預測未來1天或5天的交易量與價資料，達到股價預測之目的。在深度學中，超參數是一個難於解決之問題，我們採用田口實驗計畫法來確定最佳之超參數。實驗結果發現，使用25天的交易量與交易價資料當作輸入，預測1天後的股票資料效果最佳，在預測單一股票時誤差大約會在0.5~0.9%之間，而在預測多家公司股票時誤差會在0.4~0.6%之間。一般數值分析與機器學習方法，對於進行動態短預測有困難，本研究結合系統工程之動態系統來解決此一問題。實驗結果發現結合動態系統在單股預測之準確性(Accuracy)會有所提升，但在多股預測不會有明顯的提升，最重要的是會對預測精確性(Precision)明顯提升。整體而言，本研究運用Star GAN多面向轉換特性可以成功的處理證券量價關係以提升預測的準確性，並利用動態系統提升短期預測之準確率。

Joint effect exists between the volume and price of stock, but statistical stock analysis models cannot directly handle the joint effect but only consider the volume and price independently. This research uses the advantages of Star GAN, i.e. multi-faceted transformation, and uses result with the joint effect for the predictive model that can simulate the behavior of stock market. We feed the volume and price into Star GAN to obtain the output of potential volume and price, and then forward the results into LSTM model that is used for prediction of the volume and price in the next 1 or 5 days. However, there is great difficulty to predict the behavior of stock value in the short-term, thus. Further, Deep learning suffers from the difficulty of determination of the hyper-parameters. Thus, we apply Taguchi method for design of experimental to determine the optimal hyper-parameters. According to the experiment result, we found that using 25-day trading volume and trading price data as input to predict stock value for the next 1 day can obtain the best effect. The error lies between 0.5 and 0.9% by predicting a single stock and the error lies between 0.4 and 0.6% by predicting the prices of multiple company. It is also difficult to predict the stock price in short-term precisely, for that we introduce the dynamic system model in terms of system engineering to improve the precision of prediction. The experiment results reveal that it can improve precision of single-stock prediction, but cannot improve the precision of the multi-stock prediction. As a whole, this research use the multi-faceted transformation feature of Star GAN to improve the accuracy of stock price prediction and deal with the short-term stock price prediction for a single company with dynamic system successfully.