MLBGK: A Novel Feature Fusion Model for Forecasting Stocks Prices

Stock price prediction is a formidable undertaking owing to the inherent challenges posed by prolonged dependencies, high-frequency oscillations, and nonlinear attributes within stock price data. In response to these challenges, only one algorithm is typically used to capture the time series characteristics of stocks in previous studies. Although these methods each have the ability to achieve a certain level of predictive accuracy, they all have certain limitations due to the fluctuation of stock prices. In this paper, a novel predictive model termed Mask-LSTM-BiLSTM-GRU-KNN (MLBGK) is designed to augment the precision of stock price predictions by using three different algorithms to extract time series features of stock data. It begins by employing improved Long Short-Term Memory (Mask-LSTM), improved Bidirectional Long Short-Term Memory (Mask-BiLSTM), and improved Gated Recurrent Unit (Mask-GRU) to extract temporal features from sequential data. Subsequently, the extracted features are integrated, and ultimately, the K-Nearest Neighbors (KNN) algorithm is applied to forecast the closing price of stocks. The mask layer added to the traditional neural network is to reduce the influence of noise on the prediction effect. The improved neural network architectures combined with KNN algorithm are to improve the prediction ability of the model in the complex field of stock price prediction. The experimental results on datasets such as GOOGL, AMZN, AAPL stocks, and Bitcoin prices(BTCUSD) have shown that, across a spectrum of scenarios, the proposed model consistently demonstrates superior accuracy in stock price prediction when juxtaposed with four alternative methods, namely LSTM, BiLSTM, GRU, and LSTM-BiLSTM-GRU-KNN.