An adaptive feature selection schema using improved technical indicators for predicting stock price movements
Download 2.42 Mb. View original pdf
|
| 6. Experiment and analysis 6.1. Model performance In this study, the stock movement direction forecasting task was treated as a dichotomy problem. The common dichotomy model evaluation indexes were used to evaluate the performance of the model, including accuracy, precision, recall rate, and F score. Their meanings are shown in Table Another useful indicator is the Receiver Operating Characteristic (ROC) curve. Each point on the ROC curve reflects the sensitivity to the same signal stimulus. 6.2. Experimental result of denoising Firstly, four basic functions, Sym4, Db, Coif, and Haar, are used to conduct denoising experiments on data from different sources, and MES values were used to evaluate the denoising results. The performances of different wavelet basis functions on four data sets are shown in Table The table shows that Sym4 exerted the best effect on the three data sets SSEC, DJI, SP 500, while COIF had the best effect on HSI. The wavelet basis function with the best effect in each data set was selected for denoising in the following experiment. The trend before and after wavelet denoising is shown in Fig. Based on the conclusions of the previous experiment, we used wavelet analysis to process the original data. We calculated the technical indicators with the stock closing prices before and after wavelet denoising, and obtained the technical indicators before and after the improvement. Furthermore, the impact of improved technical indicators on the performance of the forecasting model has been investigated. We took the data from the improved and the unimproved technical indicators to predict the fluctuations after three days. The results are shown in Table As can be seen from the results in the table, compared with the SSEC data set, the accuracy and precision of the model after wavelet denoising were improved by 25.56% and 25.90%, respectively, and the recall rate and F score were improved by 25.39% and 34.48%. On the HSI data set, the accuracy is improved by 34.95%, the precision is improved by 40.54%, the recall rate and the F score were improved by 36.12% and 41.56%, respectively, Meanwhile on the DJI data set, the accuracy was improved by 31.16%, the precision is improved by 31.51%, the recall rate was improved by 31.34%, and the F score was improved by 34.48%. On the SP 500 data set, the accuracy and precision were increased by 32.07% and 33.27%, respectively. The recall rate was improved by 32.07%, and the F score is improved by 32.75%. Overall, these results indicate that using the wavelet analysis method to denoise stock closing price data, and obtaining improved technical indicators as features can greatly improve the performance of the model. Furthermore, this method was valid for stock data from different sources, among which the HSI from Hong Kong showed the best effect. 6.3. Experimental result of forecast target The forecast target of all experiments in this article is the movement direction of stock price after three days. We attempt to reduce the impact of emergencies on stock prices, because research shows that emergencies will have a greater impact on stock prices in the first three days ( Mittal and Goel, 2012 ). Taking the SSEC data set as an example, we conducted a confirmation experiment and explained the results. We used the improved indicators of the day to predict the direction of stock movements for 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9 days from the current time. Considering the randomness of a single experiment caused by various factors like the initial weights settings of the model, we conducted multiple experiments and took the average. The results are shown in Table 6 and Fig. As can be seen from the results in the table and the figure, taking the Download 2.42 Mb. Share with your friends:
|