Genetic algorithm to optimized CNN-BiLSTM model to forecast Moroccan real GDP time series

Abstract

Accurate forecasting of economic growth remains a major challenge in econometrics. In this paper, we propose a novel CNN-BiLSTM model optimized by using Genetic Algorithm (GA) for Moroccan real quarterly GDP forecasting. This approach combines the joint optimization of nine hyperparameters and dynamic lag-window selection, reinforced by a variable selection process based on Spearman correlation, Mutual Information, and XGBoost, followed by SHAP-based explainability to improve the economic interpretability of the retained predictors. The model performance is assessed under an expanding-window out-of-sample design. To quantify forecast uncertainty, Monte Carlo Dropout is employed to generate prediction intervals. Experiment results show that CNN-BiLSTM model outperforms other models, including CNN, BiLSTM, GRU, Random Forest, XGBoost, Lasso, AdaBoost, and VAR, with lower MSE = 0.014943, MAE = 0.008809, and MAPE = 0.890626% values, and a higher value of R 2 = 0.985057. Diebold-Mariano test results further confirm that the proposed model delivers statistically significant improvements over most competing models. Overall, the results suggest that CNN-BiLSTM model provides an effective framework for improving GDP forecasts for Morocco. • A GA-CNN-BiLSTM model is developed for Moroccan GDP forecasting under a realistic expanding-window strategy. • Key GDP predictors are selected using Spearman, MI, and XGBoost and SHAP analysis provides economic interpretation of the retained predictors. • Genetic Algorithm optimizes nine hyperparameters, including lag windows. • Monte Carlo Dropout provides reliable intervals and shock robustness. • Diebold-Mariano tests confirm significant gains over deep learning, machine learning, and VAR benchmarks.