Time-Weighted LSTM Model with Redefined Labeling for Stock Trend Prediction

Various techniques have been applied to predict stock market trends. However, the results are not quite satisfactory due to stock market's complexity. Many approaches either lack a clear and reasonable definition of trend or neglect the uniqueness of time attribute in stock data, treating them like other attributes, and use one-size-fits-all models to solve such a typical time-series problem. In this paper, we attempted to exploit the time attribute of stock data to improve prediction accuracy. Firstly, instead of treating data indiscriminately, we used time weight function to carefully assign weights to data according to their temporal nearness towards the data to be predicted. Secondly, the stock trend definitions were formally given by referencing financial theories and best practices. Lastly, Long Short-Term Memory (LSTM) network was customized to discover the underlying temporal dependencies in data. The trials of different time-weighted functions showed that the relation between the importance of data and their time-series is not constant. Instead, it falls within linear and quadratic, roughly a quasilinear function. Equipped with the time-weighted function, LSTM outperformed other models and can be generalized to other stock indexes. In the test with CSI 300 index, we achieved 83.91% in accuracy when fed with the redefined trends.