Use Hybrid Neural Networks To Predict Stock Trends And Mine Predictive Candlestick Patterns

Due to its tremendous value in both theory and application area,the forecast for the future trend of stocks using historical data has been widely studied by academia and industry researchers.Recently,with deep learning methods regain public's attention,more and more scholars and industry researchers try to explore how to use neural networks to solve the problem,and they have made some great progress.However,while neural networks have strong ability in fitting data,they are also criticized for lacking interpretability.Since a neural network has numerous parameters and complex structure,the academic community cannot explain the prediction logic of most neural networks,cannot perform significant tests on model parameters and cannot summarize some laws that humans can understand and use.Therefore,neural network is often seen as a predictive but unexplainable “black box”.Inspired by the theory of technical analysis in the field of stock investment,this paper uses neural network models with different characteristics to extract effect feature combinations from historical stock data including short-term price fluctuation data,long-term stock price trend data and long-term turnover rate data.Then we use the feature combinations to train a hybrid neural network.In addition,on the basis of ensuring that the prediction effect of the model is not lower than that of mainstream models,this paper uses the attention mechanism to further mine predictive K-line patterns.On the one hand,it summarizes the available judgement experience for human researchers,on the other hand,it explains the prediction logic of the hybrid neural network.This paper uses the historical volume and price data of the Shanghai and Shenzhen 300 Index component stocks in the past ten years,and proves the validity of the model through a large number of experiments.In terms of prediction,the hybrid neural network showed in this paper has a higher prediction effect on future stock trends on other models,and the trading strategy constructed based on the predictive results is superior to the derivative strategies of other models under each evaluation metric in financial area.In explaining the prediction logic of the model,the K-line pattern mined by the attention mechanism has been experimentally proved to have more significant predictive power than the general K-line pattern,which explains the prediction basis of the hybrid neural network and contributes to crack the “black box disadvantages” of neural networks.