Résumé:
In emerging, hydrocarbon-dependent economies like Algeria, managing banking
liquidity is a persistent challenge due to external shocks and structural breaks. The Bank of
Algeria (BoA) must accurately forecast aggregate liquidity to calibrate open market operations,
set reserve requirements, and safeguard financial stability. Traditional statistical models, such
as ARIMA, SARIMA, and SARIMAX, have long served as the backbone of central bank
liquidity forecasting due to their transparency and ability to capture seasonal patterns. However,
these linear approaches frequently struggle with the non-linear dynamics and complex
interactions characteristic of Algerian liquidity data, particularly during periods of external
shocks like the 2014 oil crisis and the COVID-19 pandemic.
This dissertation addresses this gap by comparing the performance of two traditional
models (SARIMA and SARIMAX) against three deep learning architectures (CNN, LSTM,
and GRU) in forecasting aggregate central bank liquidity for the Bank of Algeria. The study
evaluates forecasting accuracy across three operational horizons—day-ahead, week-ahead, and
month-ahead—and assesses the marginal value of incorporating exogenous variables (Brent
crude oil price, USD/DZD exchange rate, and holiday/seasonal dummies). The findings
confirm that deep learning models, particularly LSTM and GRU, significantly outperform
traditional statistical models by better capturing non-linear dynamics and structural breaks.
Furthermore, the inclusion of exogenous variables related to hydrocarbon prices and exchange
rates substantially reduces forecasting errors. Ultimately, this study provides both a
methodological contribution to the literature and practical, operational guidance for the Bank
of Algeria’s transition toward proactive, AI-driven liquidity management.
