基於 Informer 模型驗證網路新聞標題對於大盤預測的有效性

Abstract

因為大盤指數為反映整體市場的指標，預測其走勢是投資大眾感興趣的議題。而網路新聞由於方便取得，經常被用來搭配大盤歷史數據以預測其未來走勢。本論文利用經過時間與空間優化的Informer模型和自由時報財經新聞與聯合新聞網股市新聞的新聞標題來驗證網路新聞標題是否能夠幫助預測臺股大盤走勢。實驗以過去45個30分鐘的資料來預測下個30分鐘的收盤價，而測試資料為2024年的交易日，並以平均絕對百分比誤差（MAPE）做為評估指標。 實驗結果顯示，使用單一來源的新聞標題來協助預測30分鐘後的大盤走勢能夠幫助MAPE平均降低0.162 % ~ 0.474 %，說明網路新聞標題確實包含額外的資訊。由於本論文從自由時報與聯合新聞網挑選的新聞類別不同，且每日各自隨機選取10則新聞標題，造成兩種來源選取到的新聞標題存在議題上的差異，因此使用兩個來源新聞標題來協助預測30分鐘後的大盤走勢即使是最佳結果也僅使MAPE平均降低0.125 %。

The stock index is a proxy of the whole market so predicting its trend attracts many investors. Easily accessible online news is often utilized along with historical market data to help predict the trend of stock index. This thesis utilizes the Informer model, which is time‐ and space‐optimized, and online news headlines from Liberty Times’ financial news and United Daily News’ stock market news to assess whether online news headlines help predict the trend of the Taiwan stock index (TAIEX). The experiments use 45 preceding 30-minute data to predict the closing price 30 minutes from now. The test data covers all trading days in 2024. Performance is measured by the mean absolute percentage error (MAPE). The experimental results show that utilizing headlines from a single source to help predict the stock index can reduce MAPE by 0.162% ~ 0.474% on average, demonstrating that online news headlines do contain additional information. Since this thesis selects different categories from Liberty Times and United Daily News and then randomly chooses 10 headlines from each source every day, the headlines chosen from the two sources differ in topics. Therefore, even the best result from the combination of two sources reduces MAPE by a mere 0.125% on average.