| dc.description.abstract | There were many studies that applied deep learning to stocks, but most of them only stopped at the prediction of stock prices. These studies focused on statistical measurements, e.g. MSE; mean square error, accuracy or the directions of price movements and stopped at the level of increasing the accuracy of the predictions. Few people extended the research from the accuracy of the predictions to the application of real-life stock trading and evaluated investment performance derived from deep learning models.
This study represents a novel deep learning stock backtesting framework, which incorporated into two parts. Firstly, using a deep learning model to retrieve the prediction of stock prices. Secondly, applying the prediction of stock prices to real-life trading. We then evaluate investment performance created by the deep learning model. A prior study compared the performance of TCN and LSTM, but there haven’t been studies to compare the performance of TCN and LSTM with residual connections.
Our study compares the performance of TCN and LSTM with residual connections and applies them to four common targets (two of them are benchmarks, and the remaining two are Microsoft stock and JPMorgan Chase stock.) and then evaluates which model produces better investment performance and compares their performance with a buy-and-hold strategy.
We empirically find that no matter which business day after today we choose (i.e., one or five) and no matter which neural network we choose (i.e., TCN or RES LSTM), the fewer the neurons, the better the rate of return will be. But the reward/ risk ratio doesn’t follow the same rule.
Regarding the rate of return, RES LSTM model outperforms TCN model with five out of eight parameter sets. On the other hand, considering risk rewards, TCN model outperforms RES LSTM model with five out of eight parameter sets. Although TCN model takes less time and parameters to train, the average rate of return TCN model produces is higher and the average risk rewards TCN model produces are higher. When using TCN model, more profitable plateaus can be generated. All in all, with the application of the momentum trading strategy, the TCN network is better than the RES LSTM network.
| en_US |