Stock Price Forecast Based on LSTM and DDQN

Abstract

The prediction of time series data is very difficult. For example, the price of stocks belongs to time series. Small fluctuations in society, politics, economy and culture may affect the stocks in the stock market. In the stock market, it is very important for people to have a general judgment on stocks. Therefore, the study of stocks has practical significance. This experiment confirms that the results are affected by the data set and statesize. Statesize is predicted by the closing price of several days.On the premise that the appropriate size of statesize makes the final profit the highest, and on the premise that improved algorithm of Q value based on DQN adds regularization (DDQN), it is proved that under different data sets, adding Long Short-Term Memory (LSTM) and full connection layer are better than only full connection layer. DQN is composed of neural network and Q-learning. Q-learning is a basic algorithm in reinforcement learning. And it is proved that DDQN algorithm is better than DQN on the premise that the appropriate statesize makes the final profit the highest, and on the premise of adding regularization and LSTM. Finally, it is also proved that under certain preconditions, the combination of LSTM and DDQN is better than only DQN and full connection layer. The only indicator of this experiment is the total profit. At the same time, this paper uses the closing price to predict.