財報基本面與產業指數對上市公司股價牽引之研究─使用長短期記憶模型與多任務遷移學習

Abstract

隨著資訊科技的發展，新的科技技術與創新應用提供各領域多元解決方案與破壞式之創新。本研究以台灣半導體產業上市公司為例，利用公司財務報表內會計科目作為股價預測依據，以取得基本面之資訊。模型方面，本研究採用LSTM網路並透過多任務學習萃取產業基本面特徵與股價指數潛在結構，並應用於特定公司，以取得相應市場價值，建構出同產業內多家公司股價預測模型，實現單一模型具備多家公司預測之能力。研究顯示，本模型具一定泛化能力，能降低模型誤差。本研究樣本資料期間為十一年，結果顯示模型對於近期預測有顯著且穩定的效果。

New technologies are bringing into solutions and disruptive innovation to our society. This research retrieves the information from published financial statements of listed semiconductor industry in Taiwan as the basis of evaluation the stock prices. Our research combines LSTM neural network with multi-task learning to extract the hidden structure from network of industrial basic features and stock index, which then applies to specific company and get the potential market value in proportion. The results show the capability of generalizing the prediction to other similar companies which might lack for complete financial metrics. Over a span of eleven years collected data, the results also present significant and stable performance especially for the prediction of the recent years.