Predictive Multivariate Financial Time series in Nigerian Stock Market Using Hybrid Runge Kutta Optimization Dual Attention Long Short-Term Memory (RUN-DA-LSTM)

This study proposes a two-stage hybrid optimization framework, RUN-DA-LSTM, which integrates the Runge-Kutta Optimization (RUN) algorithm with a dual-attention Long Short-Term Memory (DA-LSTM) network to improve the accuracy of financial time series forecasting. The RUN algorithm is employed for hyperparameter optimization, leveraging its mathematically grounded population-based search to identify optimal LSTM units, learning rate, batch size, temporal window size, and dropout rate. The optimized parameters are applied to a DA-LSTM architecture combining input attention for feature-level weighting and temporal attention for sequence-level importance modeling. Using Nigerian equity market data for DANGSUGAR, FCMB, and OANDO, the model was evaluated against GRU and baseline LSTM architectures using RMSE, MAE, MSE, and R² metrics. RUN-DA-LSTM achieved the lowest RMSE across all assets (10.4553 for DANGSUGAR, 0.2440 for FCMB, and 1.6134 for OANDO) and demonstrated improved explanatory power, particularly with volatile and structurally complex price series. While directional accuracy and trading performance showed more mixed results, these metrics serve as secondary indicators. SHAP-based feature importance analysis revealed that currency-related metrics (Selling Rate, Buying Rate, and Central Rate) were the dominant predictors across all assets, exceeding the influence of commodity prices such as crude oil. Overall, the findings validate the effectiveness of combining mathematically grounded optimization with dual-attention mechanisms for financial forecasting, highlighting the importance of asset-specific hyperparameter calibration.