Utilization of Various Optimization Algorithms in Neural Network Models for Forecasting Exchange Rates

In the current global economy, predicting exchange rates precisely is a crucial task for investment decisions. Presently, Sri Lanka is in a challenging economic environment due to the unforeseen depreciation of the Sri Lankan Rupee (LKR). This study investigates the viability of employing artificial neural networks with several optimization algorithms in the prediction of exchange rates. The data were sourced from the Yahoo finance database, spanning from January 1
^st
, 2008, to February 28
^t
h, 2022. The investigation used historical lagged data with moving average technical indicators as the inputs for Feedforward Neural Network (FFNN), Long Short-Term Memory (LSTM), and Generalized Regression Neural Network (GRNN) models. This study applied neural networks of FFNN with optimization algorithms of adaptive moment estimation (ADAM), stochastic gradient descent (SGD) and root mean square propagation (RMSprop), Time Delay Neural Network (TDNN) with SGD and RMSprop optimization algorithms, GRNN, and LSTM model implementing ADAM and RMSprop optimization algorithms in forecasting exchange rates involving major currencies of AUD, CAD, CHF, EURO, GBP, JPY, SGD and USD against LKR. The assessment of the predictive capacity of each fitted model was conducted by evaluating error metrics of mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean squared error (RMSE). The empirical findings demonstrated that the FFNNs, TDNNs and LSTM models exhibited superior performance in predicting exchange rates compared to GRNNs. Specifically, the FFNN with the RMSprop algorithm demonstrated lower error values for AUD/LKR, CHF/LKR, EURO/LKR, and JPY/LKR. Additionally, for GBP/LKR, the FFNN with the ADAM algorithm performed well, while the TDNN with RMSprop showed superior performance for CAD/LKR. Furthermore, the LSTM with ADAM algorithm proved to be most effective for predicting SGD/LKR and for USD/LKR it was LSTM with RMSprop algorithm.