Predictive modeling for the Moroccan financial market: a nonlinear time series and deep learning approach

Abstract This work presents a comparative study of predictive modeling methods in the context of financial markets, focusing on the Moroccan market. The objective of this study is to evaluate the performance of three different approaches: the autoregressive integrated moving average (ARIMA), bilinear models and long short-term memory (LSTM) networks. Predictive modeling is a vital skill for understanding and anticipating market behavior and facilitating informed decisions for investors, policymakers and financial institutions. This research uses historical financial data dating back to the post-COVID-19 pandemic period to train and evaluate these models. The bilinear model, a new relativistic approach to time series financial forecasting, is compared to the traditional linear ARIMA model and an advanced LSTM neural network. Our methodology includes data preprocessing, including normalization, to ensure compatibility with each important technique. Then, we train the models using residuals from historical market data and evaluate their performance using metrics such as mean absolute error, root-mean-square error and trend forecast accuracy. The results provide insights into the effectiveness of our recently developed hybrid models, which combine linear and nonlinear elements, in monitoring the dynamics of the Moroccan financial market. We thoroughly discuss the strengths and weaknesses of each model, focusing on areas where hybrid techniques outperform others. This research aims to delve deeper into and open up new avenues for understanding these models in financial markets and also to provide valuable insights for practitioners and researchers. The findings have implications for portfolio management, risk assessment and decision support systems within the Moroccan financial ecosystem.