Deep Learning in Financial Analytics - Exchange Rate Modelling

Abstract

In finance, a major enthralling research question has been the accurate determination of future market and economic movements. A lot of researchers have tried to develop different models with different accuracies of prediction over the years. It appears that the full potential of deep learning has not been explored to study FX rates. The current study, therefore, explored the proficiency of deep neural networks in predictive modeling. I tested different models of artificial neural networks (using hyperparameters tuning like training algorithms, number of hidden layers, and hidden nodes) using neural network input-output fitting and tried to find the best fit model. The model was also validated by layered digital dynamic time series modeling using autoregression with two delays. The appraisal metrics used were regression R - value, MSE, time-series response plot, and error autocorrelation plot. It was concluded that the artificial neural network with a single hidden layer having 17 nodes and trained using the Levenberg- Marquardt algorithm gave the best performance in a minimum number of iterations. This study marks an extensive examination of ANN modeling. This model can be used by traders, investors, financiers, economists, bankers, speculators, hedgers, and governments to get insights into future forex rates and thus make profitable decisions. Various trading policies, import-export policies, and pricing of commodities in indigenous markets can be managed precisely. Future studies can use these models in simulated trading and help establish an alliance between statistical significance and economic significance.