dc.description.abstract | Normally, when deep learning is used to predict the price of the financial commodity, people most used the statistical function (such as R^2) that compliance to the numerical error to assessing the accuracy of the estimate as a learning mechanism. However, for the real trading, the forecast trading direction (to buy or to sell) cause by the error of the stock price prediction usually has a greater impact on the profitability of trading than the numerical error between the predicted value and the actual value, but the study of time series prediction by deep learning today rarely explores this topic, but it has practical importance.
In this study, we compare the closing price prediction model of LSTM and TCN based on the numerical error (R^2, MASE), to find out the high-accuracy parameter settings and use it on the real trading. We use program trading to explore whether these predicting of closing prices can achieve actual profitability or not, and if so, we will compare the relationship between its profitability and the original numerical error. This study also proposes a custom loss function that incorporate the actual trading direction (to buy or to sell) to increase the profit of real trading.
This study is aimed at three futures commodities that are often traded (gold, soybean, crude oil) and perform the above overall calculation and comparison, and we found that some of the commodities will increase the profit when minimized the numerical error, but others not. More than that, that kind of commodities can also enhance the profit of real trading when it adds a custom loss function. However, if the commodities can’t increase the profit when minimized the numerical error, the custom loss function in this study also can’t increase the profit. | en_US |