上市個股預測台灣加權指數高頻趨勢

Abstract

股市指數通常由多家上市公司的股票組合而成。然而其價格決定於複雜的交易過程，並且資產間時常具有非線性的相依性。因此為了預測指數未來的短期走勢，我們考慮以卷積神經網路處理非線性多維時間序列。我們將模型用於預測指數的上漲或下跌，並實證模型的預測性。市場高頻資料 (每5秒) 包含了 2015 年至 2017 年日內的臺灣加權指數與個股價格。實驗結果顯示，模型使用在日內交易預測能得到良好的準確率與績效。

A stock index consists of a board selection of stocks. Stylized facts show non-linear correlations exist between stock prices. Therefore, predicting the direction of an index involves modeling and analyzing sophisticated multi-dimensional time series. We employ the convolutional neural network to forecast the intra-day price movement of the Taiwan Stock Exchange Weighted Index (TAIEX). Furthermore, we verify the prediction performance of our model on the high-frequency data (5 sec) of the TAIEX and its constituent share prices from 2015 to 2017.