Forecasting major currency exchange rates using long short-term memory networks: Evidence from multi-currency time series analysis

Exchange-rate dynamics are non-linear and volatile, which challenges conventional forecasting approaches. This study evaluates a reproducible long short-term memory (LSTM) framework for daily EUR/USD, GBP/USD, USD/TRY, and USD/JPY over 1 January 2010 to 31 December 2021. The contribution is twofold: (i) a fully specified and deployment-oriented LSTM protocol (architecture, preprocessing, and leakage-safe validation) suitable for applied forecasting; and (ii) a time-series-appropriate evaluation that combines rolling-origin (walk-forward) testing with standard baselines (random walk and ARIMA) and diagnostic visualizations. Forecast performance is reported using root mean square error (RMSE), mean absolute error (MAE), Pearson correlation (R), Nash-Sutcliffe efficiency (NSE), and the RMSE-to-SD ratio (RSR), alongside distributional diagnostics (violin plots) and horizon-specific error profiles. The results quantify performance gains relative to baselines under leakage-safe evaluation, while highlighting practical implications for treasury and risk management. Limitations include the exclusion of exogenous drivers and longer-horizon tests, motivating extensions that incorporate macro-financial signals and interpretability modules.