LSTM stock price prediction optimization method based on 3D-CNN and Hash attention mechanism

Abstract

The invention provides an LSTM stock price prediction optimization method based on 3D-CNN and a Hash attention mechanism. The method comprises the steps: introducing a multi-dimensional convolutional neural network mechanism for high-dimensional features of stock data to enhance the deep feature extraction capability in the stock data, wherein the stock data are multiple groups of related tense data of basic attributes, technical indexes and fused market emotion comment texts; determining a one-dimensional convolutional neural network type according to the stock data dimension in the step 1, wherein three stock data types are included in the step 1, so three convolutional neural networks are adopted, a Gaussian kernel function is adopted as a convolution kernel, and pooling operation is performed by adopting a dimension with a pooling layer size of 4; for the basic network LSTM, selecting an attention mechanism based on locality sensitive hashing to adjust the weight of neurons, so the proportion of neurons with large influence is larger. According to the method, the attention mechanism is used for LSTM hidden layer neuron state weight adjustment, the influence degree of effective features on final output is increased, and the prediction accuracy is improved.

Description

Technical field [0001] The present invention belongs to the field of LSTM stock prices, and more particularly to a LSTM share price prediction optimization method based on 3D-CNN and hash precautionary mechanisms. Background technique [0002] Stock is an important part of the financial market, and the accurate prediction of stock prices can increase the investment income of small and medium investors from microscopic, from macro contributing to the government's abnormal movement of the financial market in advance. However, stock data is high noise, nonlinear, uncomfortable characteristics, making traditional models in stock data. Based on the above background, this paper has been made through the limitations of traditional models in stock forecasts, and the optimization model of the stock data characteristics is proposed to improve stock forecast accuracy. [0003] The LSTM model is difficult to learn the problem of depth characteristics in multi-portable data, this paper is bas...

Images