Forecasting the Moroccan All-Share Index: A Comparative Study of ARIMA and LSTM Deep Learning Approaches

Abstract

This study addresses the problem of accurately forecasting financial markets in emerging economies, specifically focusing on the Moroccan All-Share Index (MASI) which lacks dedicated research. Motivated by the need to provide a clear and comprehensive comparison for local and international investors, this paper evaluates two prominent time series models: the classical ARIMA model and the deep learning-based LSTM network. The methodology involves using a time series of daily MASI closing prices from January 4, 2010, to August 8, 2025, which was chronologically split into a training sample (80%) and a test sample (20%). The models' out-of-sample performance was then rigorously evaluated using key error metrics. The results reveal a significant disparity in performance. The LSTM model delivered drastically superior accuracy with an MAE of 178.52 and an RMSE of 272.69, vastly outperforming the ARIMA model which yielded an MAE of 1917.74 and an RMSE of 2751.31. This demonstrates that the LSTM’s ability to capture complex, non-linear dependencies is far more effective for forecasting the MASI index than the linear assumptions of the ARIMA model. The study concludes that deep learning methods offer a more reliable approach for financial forecasting, with practical implications for investors who can use LSTMs to make better-informed trading decisions and for analysts who can incorporate these models for more nuanced market surveillance and risk management.