Stock prediction method based on ARMA-LSTM model

Abstract

The invention provides a stock prediction method based on an ARMA-LSTM model, and the method comprises the steps: carrying out the regression fitting and prediction of the stock sequence data throughan ARMA model, carrying out the training and prediction of a residual error sequence through an LSTM model, and finally adding the two results as a final prediction result. The stock transaction datais related with the historical data, and comprises a linear correlation part and a nonlinear correlation part, so the method achieves the fitting of the data sequence through employing the ARMA modelin advance, and the linear part of the data sequence is extracted, thereby speeding up the convergence of the LSTM training, and improving the prediction capability of the LSTM for the nonlinear partso as to reduce the local convergence phenomena.

Description

technical field [0001] The invention belongs to the technical field of machine learning and pattern recognition, and in particular relates to a stock prediction method based on an ARMA-LSTM model, which can accurately predict future stock prices by using historical stock price sequences. Background technique [0002] In the vast universe and the long river of history, it seems that all things develop with the change of time. We believe that the development of everything is composed of certain regularity and randomness. As long as we collect corresponding historical data at different points in time, this type of data can reflect the law of a certain thing, phenomenon, etc. over time, and these data are called time series data. Stock data is also a kind of time-series data. Changes in stock prices reflect the trend of the economic market and also reflect the development of the country's economy. Therefore, for the research on stocks, discovering the rules of change can better ...

AI Technical Summary

Problems solved by technology

For the determination of the model order (p, q), there are currently two methods, one is to determine through the autocorrelation diagram ACF and the partial autocorrelation diagram PACF, this method can only be determined empirically, and cannot be quantified accurately Determine; the second is to quantify and determine through the AIC or BIC information criterion (generally, the smaller the value, the better the model), but in actual experiments, it is found that the AIC calculated by many time series data is negative and does not converge to a certain value. A specific order, so it is difficult to accurately find a reasonable p, q to achieve model validity detection

At the same time, for neural networks (such as multi-layer BP neural network, RNN neural network), we know that it solves parameters through iterative operations, which inevitably has disadvantages such as slow training speed, convergence to local optimal solutions, and overfitting.

Images