RiskAwareTNet: A Multi-Task Deep Learning Framework for Stock Price Prediction and Risk Assessment

This paper introduces RiskAwareTNet, a novel deep-learning model designed for multi-task learning that addresses stock price prediction and risk evaluation. The model combines Bidirectional Long Short-Term Memory (BiLSTM) networks with Transformer encoders to analyze a 15-year dataset (2008–2024) for MasterCard (MA) and Visa (V), capturing complex temporal trends inherent in financial data. The model’s multi-task design allows it to simultaneously predict stock prices and assess critical financial risk metrics, including volatility and the Sharpe ratio, offering a more holistic view of financial performance. RiskAwareTNet is evaluated against traditional models like ARIMA and XGBoost. The results demonstrate its superior performance, achieving a Mean Absolute Error (MAE) of 3.12, a Root Mean Squared Error (RMSE) of 4.50, and a Mean Absolute Percentage Error (MAPE) of 1.95% for stock price predictions. For risk assessment, the model yields a Mean Squared Logarithmic Error (MSLE) of 0.021, a Symmetric Mean Absolute Percentage Error (SMAPE) of 4.12%, and an Explained Variance Score (EVS) of 0.95. These findings indicate an improvement in accuracy by 12.3% compared to traditional forecasting methods. RiskAwareTNet offers a robust and reliable solution for financial analytics, providing a powerful tool for both stock price forecasting and comprehensive risk evaluation, thereby enhancing the decision-making process for financial analysts and investors.