Exchange Rate Prediction Under Data Volatility: A Comparison Of Deep Learning Techniques

This Accurate prediction of exchange rates is of critical importance for economic policy related decision making. This study employs a deep learning based framework to model exchange rate dynamics by explicitly accounting for the inherently nonlinear, and regime sensitive nature of financial time series. Unlike traditional artificial neural network approaches that overlook temporal dependencies, recurrent neural network architectures are capable of directly modeling long term time dependencies. Accordingly, this study applies Long Short Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Gated Recurrent Unit (GRU) models to predict the USD/TRY and EUR/TRY exchange rate series. The empirical analysis is conducted using data from the Turkish economy across three distinct subperiods, representing a relatively stable regime and two high-volatility regimes. Hyperparameters are independently optimized for each model and regime with grid search. Predictive performance is evaluated using Mean Absolute Error, Root Mean Squared Error, and Mean Absolute Percentage Error, with Diebold Mariano tests confirming statistical significance and formal benchmark comparisons against ARIMA and random walk models. Results indicate that no single architecture dominates across all conditions: GRU consistently outperforms competitors during the Crisis period for both exchange rates, while LSTM and BiLSTM perform best during COVID19 for USD/TRY, and GRU maintains its advantage for EUR/TRY. Diebold Mariano tests confirm that 89% of pairwise performance differences are statistically significant. Robustness analysis using lag-3 specifications supports that regime-dependent patterns are not artifacts of input structure. By incorporating volatility regimes, statistical testing, and econometric benchmarks, this study demonstrates the practical importance of regime-aware model selection for exchange rate forecasting in emerging market economies.