مجلة الاقتصاد وادارة الاعمال
Volume 10, Numéro 1, Pages 1-8
2026-06-10
Comparative Analysis Of Multi-layer Perceptron Architectures For Usd/dzd Exchange Rate Forecasting
Authors : Benlaria Ahmed .
Abstract
Exchange rate forecasting is a cornerstone of international finance; however, it remains a significant challenge due to the nonstationarity and nonlinearity of financial markets. This study investigates the optimisation of deep neural network (DNN) architectures for predicting the USD/DZD exchange rate, a currency pair that is less frequently studied in emerging markets. Utilising daily data from the International Monetary Fund (IMF) spanning May 2010 to May 2024, this study employs systematic feature engineering, including lagged values and moving averages, to capture short- and medium-term market dynamics. We evaluated various multilayer perceptron (MLP) configurations, ranging from single-hidden-layer models to deeper, multi-layer structures. The results indicate that while all tested architectures achieved high predictive accuracy, a modest two-layer configuration—specifically ANN ([10, 5])—yielded the lowest error metrics, with a Mean Squared Error (MSE) of 3.4452\times{10}^{-6} and a Root Mean Squared Error (RMSE) of 0.0018561. These findings suggest that increasing network complexity beyond a certain threshold offers diminishing returns when effective feature engineering is applied. This study underscores the practical utility of optimised DNNs for financial stakeholders and policymakers seeking to mitigate currency risk and enhance capital allocation in volatile economic environments.
Keywords
Exchange Rate Forecasting ; Deep Neural Networks (DNN) ; USD/DZD ; Multi-Layer Perceptron (MLP)
Les articles similaires
Semane Wassila
.
pages 469-491.
Ouakkal Noureddine
.
pages 305-323.
Aissaoui Soufyane
.
Bourouis Mohammed El Amin
.
Zadjaoui Abdeldjalil
.
Chalabi Youssouf
.
Mimoune Arezki El Haj
.
pages 267-280.
مداني هاجر
.
بن علقمة مليكة
.
ص 348-365.
Tawati Ahlaam
.
pages 571-580.
