Stock Price Prediction Model: Assessing the Performance of a Hybrid Deep Learning Model Employing Multi-Stream Data

Abstract

Purpose - The study investigates the effectiveness of the ConvLSTM model in the next-day closing price prediction for stocks using a novel combination of input features. These features include past prices, prices of related stocks, technical indicators of the target stock, mutation point impact on closing price, stock market sentiment, stock market index, interest rate, and dollar exchange rate

Study Design/Methodology/Approach - Sentiment analysis of financial news related to the Pakistan Stock Exchange (PSX) was performed using the pre-trained FinBERT model. Relevant stocks were identified through historical prices using Random Forests and XGBoost algorithms. A comparative analysis was conducted involving ConvLSTM1D and ConvLSTM2D models, as well as LSTM, CNN, and CNN+LSTM models, all having the same network structure.

Findings - The study found that there is not a universal best model for every stock. However, the ConvLSTM1D model exhibited the best average performance for Mean Squared Error (MSE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE) across the selected stocks.

Practical Implications- The findings suggest that while no single model is optimal for all stocks, the ConvLSTM1D model may be particularly effective for improving prediction accuracy in stock closing prices. This can aid investors and analysts in making more informed decisions based on the model's performance metrics.

Originality/Novelty- This study provides a novel approach by integrating a diverse set of input features and comparing multiple models in a comprehensive manner. The use of sentiment analysis and advanced machine learning techniques such as ConvLSTM adds significant value to the field of stock price prediction.