Comparative Analysis of ARDL, LSTM, and XGBoost Models For Forecasting The Moroccan Stock Market During The COVID-19 Pandemic

Abstract

This study evaluates and compares the forecasting performances of the ARDL (AutoRegressive Distributed Lag), LSTM (Long Short-Term Memory), and XGBOOST (Extreme Gradient Boosting) models on the MASI (Moroccan All Shares Index). The analysis incorporates daily new COVID-19 cases into the ARDL approach to investigate short-term and long-term relationships with MASI. Cointegration and causality tests are conducted on daily time series data. In terms of accuracy, the ARDL model, especially when including trend and seasonality variables, outperforms LSTM and XGBOOST models. The ARDL model with lags, trend, and seasonality variables achieves the lowest Mean Absolute Percentage Error (MAPE) of 26.7%, with a processing time of 1 second. In comparison, the LSTM and XGBOOST models have MAPE values of 30.5% and 32%, respectively, while requiring significantly longer processing times. These findings suggest that the ARDL model is more efficient and accurate in predicting future values of MASI under pandemic conditions.