Forecasting the PKR–USD exchange rate using recurrent neural networks: a comparative analysis of LSTM, GRU and CNN

Purpose This study investigates the forecasting of the PKR/USD exchange rate by employing three deep learning architectures: Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU) and Convolutional Neural Networks (CNN). For the first time in PKR/USD exchange rate literature, the study attempts to fills the gap by comparing the models of time-series and econometric frameworks to the emerging market exchange rate forecasting literature for Pakistan. Design/methodology/approach The dataset covers the period from January 2012 to July 2022 and contains daily observations of the USD exchange rate and macroeconomic variables, including the Commercial Bank Reserves (ComRes) and Six-Month KIBOR (SixMKib). The models are constructed using TensorFlow, and Keras, and the walk-forward validation is used to maintain temporal order in the dataset. The forecasting performance of the proposed models was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and the coefficient of determination (R2). Findings The econometric GRU model was the most effective model overall, with a predictive accuracy of R2 of 99.29%, an RMSE of 3.27 and an MAE of 1.84. The CNN models, with 98.49 R2 for the time series, also recorded high values although they had higher errors. Although the LSTMs performed adequately, they were more expensive to run, and the MAPE and SMAPE were higher, as were the LSTMs, RMSE and MAE values. Practical implications This study's findings provided a more advanced strategy to help predict foreign exchange more accurately, which can significant to central banks and other financial institutions in managing risks associated with fluctuating currencies. This study strongly empowers actively guiding the range of values in decision-making by defining and adjusting target policies to improve the balance of a portfolio. Originality/value The novel aspect of this paper is the application of three deep learning frameworks LSTM, GRU and CNN, in a comparative study to predict the PKR/USD exchange rate, fulfilling a gap in the literature. The study also improved econometric models of exchange rate prediction while incorporating macroeconomic fundamentals which provided a more complete perspective on exchange rate behavior and demonstrated the ability of GRU models to predict short-term volatility.