A novel hybrid deep learning framework with volatility-weighted loss for high-fidelity stock price prediction

Historical stock price data and technical indicators, such as close, volume, and open interest, are widely applied to improve forecasting and market understanding. These indicators capture essential information on trends, volatility, and potential reversals, offering actionable insights for traders and analysts. This study investigates predictive modeling of Visa, Mastercard, and leading cryptocurrencies (Bitcoin, Ethereum, and Litecoin) using four proposed Hybrid Temporal Feature Integration Network (HTFIN) models. Each model integrates positional encoding, transformer blocks, convolutional neural networks (CNN), long short-term memory (LSTM), and gated recurrent units (GRU) to capture both local dependencies and long-range temporal patterns in financial time series. Among them, HTFIN4 demonstrates the strongest predictive capability, with notable gains in accuracy. For example, the [Formula: see text] value for Visa improves from 0.9902 to 0.9978, and for Mastercard from 0.9932 to 0.9986, when trained with the proposed Volatility-Weighted Directional Loss (VWDL) instead of Mean Squared Error (MSE). VWDL improves performance by minimizing errors, incorporating volatility, and emphasizing directional correctness. A hybrid PCA-autoencoder further enhances feature selection, reducing redundancy and noise. Statistical validation using paired t-tests confirms the improvements are highly significant ([Formula: see text]). Overall, the HTFIN4 with VWDL adapts effectively to both stock and cryptocurrency markets, supporting practical forecasting applications.