TCN-based Futures Prediction Using Financial Indices, Bargain Chips, and Forum Messages

Abstract

Traditional research on the stock and/or futures price prediction mostly uses the past stock/future prices and technique indicators, such as KD, RSI, and MACD, as features. Very few studies consider the forum messages or bargaining chips as stock and/or futures price prediction features. In this research, discussion messages from both PTT and CMoney forums are converted into daily sentimental vectors using the retrained BERT. The daily sentimental vector as well as three bargaining chips are then used as features to train the GRU and TCN models. The experiment results show that the TCN performs better than the GRU-based RNN model in terms of MAE, MAPE, RMSE, and accuracy. In addition, both of the bargaining chips and forum messages are verified to be useful in the futures price prediction. The market simulations based on the historical futures price show that a simple investment strategy using the TCN models with techniques, bargaining chips, and forum messages can earn more than 7 times of the investment in the period of one year.