情緒指標與台灣股市短期震盪之預測

Abstract

在主流經濟文獻中，效率市場假說（The efficient market hypothesis, EMH）假設投資人為理性，從而技術指標應能提供合理的報酬。然而文獻指出這樣的假說並不能適用於劇烈震盪的股市。行為經濟學家指出情緒在證券價格決策中也佔有地位，因而引申出兩個問題：如何有效的擷取情緒以及情緒如何影響市場價格。 我們嘗試了以下三個模型：「純技術指標模型」、「情緒字典模型」、「自標註模型」，在不同的移動平均與最小文本長度下，「純技術指標模型」在一日移動平均與最小文本長度提供了最好的預測結果，其RMSE表現好於其他模型10-30%。 在情緒影響市場方面，我們提出以情緒字典與手標資料來微調（fine-tuning）BERT模型來做為情緒分類器。此法改善擷取情緒的效率，同時在我們的測時集中達到了93%的F1分數。 實驗結果顯示，加入情緒會使的預測的指數向上偏移，亦即對於現實世界的交易而言，加入情緒並不是個好選擇。

The efficient market hypothesis (EMH) in mainstream financial literature assumes that market behavior is rational; thus, technical indicators should provide reasonable predictions for the fluctuations of stock markets. However, existing literature points out that the hypothesis may not be sufficient for predicting highly volatile stock fluctuations. Behavioral economists have proposed that the emotions of investors can also influence stock prices. Therefore, we arrive at two questions: how to extract emotions more effectively and how emotions affect predictions. For the first question, we test three different models for stock price predictions, namely “technical indicator only model ,” “dictionary emotion model,” and “self-labeled emotion model,” with the inputs of different moving averages and minimum text lengths. The results show that the “technical indicator only model” performed the best with a 1-day moving average and a minimum length of three words. It outperformed the other two models by 10–30% in RMSE. For the second question, we proposed extracting emotions by fine-tuning BERT as the emotion classifier with dictionary and hand-labeled data. The method compensates for disadvantages of common ways to extract emotions, and achieves over a 93% F1 score in our testing data. The results show that adding emotions to the prediction process can cause predictions to shift upward, and imply that adding emotions to price predictions can be a false choice for real-world trading.