Time series prediction of stock price using deep belief networks with intrinsic plasticity

In recent years, the stock market plays an important role, which has attracted more and more attentions. The key problem of the stock market prediction is how to design a method to improve the prediction performance. As we know, the biggest challenge is that the stock time series is essentially dynamic, nonlinear, complicated, nonparametric and chaotic. In this paper, we propose a novel method to predict the stock closing price based on the deep belief networks (DBNs) with intrinsic plasticity. In the experiments, the stock in S&P 500 is used to examine the performance. The back propagation algorithm is used for output training to make minor adjustments of structure parameters. The intrinsic plasticity (IP) is also applied into the network to make it have adaptive ability. It is believed that IP learning for adaptive adjustment of neuronal response to external inputs is beneficial for maximizing the input-output mutual information. Our results show that the application of IP learning can remarkably improve the prediction performance. Moreover, the effects of two kinds of IP rules on the performance of prediction are examined. Compared with Triesch's IP and without IP, DBN with Li's IP learning has much better prediction performance than the others. These results may have important implications on the modeling of neural network for complex time series prediction.