Résumé:
This study evaluates the forecasting performance of traditional time series models (ARIMA, VAR) alongside advanced deep learning methods (LSTM, CNN) in predicting key determinants of banking liquidity in Algeria. The analysis is based on daily data from 2015 to 2023, provided by the Bank of Algeria, focusing on three critical variables: Net Foreign Assets, Currency in Circulation, and Government Net Position. We first performed forecasting using the all-at-once method on the statistical models, then implemented a Rolling Forecast approach to enhance their predictive accuracy and adaptability over time. Following this, we proceeded to forecast using deep learning models to compare their performance with the statistical methods. Our results demonstrate that LSTM outperforms classical models in capturing complex nonlinear patterns and temporal dependencies in Net Foreign Assets and Government Net Position, while traditional models still provide reliable forecasts for Currency in Circulation, particularly when enhanced by Rolling Forecasts. This comparative analysis provides valuable insights for developing more effective forecasting tools to assist monetary authorities in improving liquidity management and maintaining the financial stability of Algeria
