Stock Prediction based on Bayesian-LSTM

Biao Huang, Qiao Ding, Guozi Sun, Huakang Li

2018 · 33 citations · 20 references

Concepts

Abstract

Fluctuations in stock market represent the changes in national economic objectively. Some machine learning algorithms, such as linear fitting and sequence mining, are employed to predict the stock market. However, linear fitting faces issue of over-fitting and black relationships with historical data, while sequence mining is short in efficiency and lack dynamic adaptations. This paper builds a modified Bayesian-LSTM (B-LSTM) model for stock prediction. Six indicators of the Chinese stock market in every day are the basic input for LSTM. In order to represent the economic wave, we defined a data set unit by week which means the basic unit in LSTM is data in one week. Furthermore, a Bayesian optimization model is proposed to estimate the unit number dynamically in different economic cycle. The experimental results demonstrated that the B-LSTM increase over 25% prediction than the conventional LSTM.

References

20