استخدام نماذج التعلم العميق من نوع LSTM المحسّنة بخوارزمية ADAM في التنبؤ بسعر الصرف في سورية

Abstract

This study aims to develop an accurate predictive model for the weekly exchange rate in Syria using Long Short-Term Memory (LSTM) neural networks optimized with the ADAM algorithm, based on weekly data spanning from 2010 to May 2025. The research responds to a critical academic and practical need, as traditional models have failed to provide accurate forecasts in an economy characterized by volatility and monetary instability. A quantitative analytical approach was adopted, starting with preprocessing the exchange rate time series and performing statistical analysis through descriptive indicators and autocorrelation functions (ACF and PACF). The LSTM model was then trained on 80% of the data and tested on the remaining 20%. The results showed high predictive accuracy, with the coefficient of determination (R²) reaching 0.9693 on the test set and a mean absolute percentage error (MAPE) of approximately 5.74%, confirming the model’s ability to capture temporal patterns and achieve stable training. Additionally, the model provided future forecasts for the final four months of 2025, incorporating realistic uncertainty intervals ranging between ±1,300 and ±1,500 SYP around the central forecast. The study concludes that the LSTM model with ADAM optimization is an effective forecasting tool in complex economic environments. It recommends that the Central Bank of Syria adopt this model to support monetary policy decisions. Future research is advised to compare the proposed model with alternatives such as GRU and Transformer networks to further validate its performance.