Feature selection method to improve time series forecasting of Moroccan GDP using deep learning

Abstract

The objective of this work is to identify relevant indicators to improve the performance of the CNN-LSTM hybrid model for forecasting Morocco’s gross domestic product. Before training, we utilized a robust variable selection step by combining Mutual Information, TSFresh, and the RFECV method with XGBoost, which enabled us to select important features that are closely related to GDP, while Bayesian Optimization was used to determine optimal hyperparameters. The introduced model was tested on quarterly indicators and GDP time series for Morocco from the Haut-Commissariat au Plan, covering a 45 -year period from 1980 to 2025, consisting of 183 observations and 16 features. To ensure a realistic assessment of predictive performance, the proposed model was evaluated using an out-of-sample forecasting strategy, which simulates real-world forecasting conditions and tests the model’s ability to generalize to unseen data. Experimental results showed that the hybrid CNN-LSTM model with feature selection outperformed the hybrid model without feature selection, the individual models—namely CNN, LSTM, and TCN—as well as the traditional econometric Vector Autoregressive (VAR) model, with lower MSE, MAE, and MAPE values, and a higher value of $\mathbf{R}^{\mathbf{2}}$.