An Optimized Deep Learning Model for Stock Price Prediction Incorporating Investor Sentiment

Abstract

The Multimodal Semi- Supervised Attention-based Long Short-Term Memory (MS-SSA-LSTM) model introduces an improvement in stock price forecasting by combining sentiment analysis, swarm intelligence algorithms, deep learning methods, and multi-source data. This model builds a dictionary of sentiments and calculates a sentiment index based on aggregation from East Money forum postings for sentiment analysis. As a result, this kind of study is insightful in providing analyses on how changes in market sentiment influence stock prices.In order to optimize the prediction accuracy, the LSTM hyperparameters are tuned using the Sparrow Search Algorithm (SSA). Empirical findings indicate that the MS-SSA-LSTM model outperforms the other models evaluated. This methodology is an effective approach for generating accurate stock price forecasts. Excellent, the program, which was specially designed for China's erratic financial market, excels in the prediction of short-term stock prices and provides useful information to help investors make an informed decision. A hybrid model of LSTM+GRU has also been introduced for the classification of stock emotion. The approach also utilized a strong ensemble, which consisted of a voting regression for stock price prediction: LinearRegression + RandomForestRegressor + KNeighborsRegressor, and a voting classifier for sentiment analysis: AdaBoost + RandomForest. All these ensembles were combined to enhance the overall predictive performance by being easily integrated with the existing models (MLP, CNN, LSTM, MS-LSTM, and MS-SSA-LSTM). A user-friendly Flask frame work with SQ Lite support was developed to streamline the sign up, sign in,  and model assessment procedures and enable user engagement and testing.