A hybrid deep learning framework for financial time-series forecasting: evidence from Asia-Pacific stock indices

This study proposes DeepCBN with Temporal Infusion, a dual-path hybrid architecture integrating convolutional neural networks, bidirectional long short-term memory, and temporal convolutional networks to capture localized volatility, bidirectional dependencies, and long-range sequential patterns. Using 22 years of daily data from eight Asia-Pacific stock indices, the framework is evaluated through sliding-window forecasting, Bayesian hyperparameter optimization, walk-forward validation, ablation analysis, and comparisons with eight benchmark models. It achieved an average mean absolute percentage error (MAPE) of 8.74% and mean absolute error (MAE) of 1321.01, compared with 9.37% and 695.85 for standalone DeepCBN and 8.76% and 781.34 for standalone TCN. Because the Hang Seng index produced an unusually large error, median results better represent typical performance. Ablation analysis confirms the complementary contributions of all three components. Although Wilcoxon tests did not establish significant superiority, the framework generally produced competitive forecasts and may be applicable to other nonlinear sequential prediction problems.