Résumé:
This research investigates the forecasting of exchange rate volatility for USD/DZD and
EUR/DZD, comparing classical time series models (SARIMA, SARIMAX) with Deep
Learning models (CNN, LSTM) using daily data from January 5, 1999, to February 28, 2025
(6,824 observations). In Algeria s hydrocarbon reliant economy, precise volatility forecasts are
essential for economic decision maki ng, as exchange rate fluctuations significantly affect trade,
investment decisions, and monetary policies. The study seeks to determine which approach
classical or deep learning better captures the intricate dynamics of exchange rate volatility, a
critical measure of market risk in an emerging market context. The analysis reveals that, when
predicting in an all at once perspective, classical model failed to capture long term patterns and
trends. While using the rolling forecast on the classical models, ARIM AX consistently
outperforms other models. This method demonstrates greater adaptability to shifting market
conditions, effectively capturing trends and fluctuations in exchange rates, resulting in lower
error metrics compared to other models the rolling fo recast techniques allows the models to
enhance its predictive accuracy for both currency pairs. Among deep learning models, LSTM
surpasses CNN, showing a stronger ability to model long term dependencies and complex
patterns in sequential data. Despite this , LSTM still falls short closely behind of the classical
ARIMAX rolling forecast approach, indicating that adaptive classical techniques offer more
reliability for this specific dataset. Finally, the volatility is visualized by directly deriving it
from US D/DZD and EUR/DZD prices, highlighting the chosen model performance and
flexibility in achieving accurate forecasts. The findings emphasize that while Deep Learning
models excel in capturing nonlinear dynamics, classical models with adaptive forecasting
me thods provide a more robust solution in this context.
