Evaluating ANN Models in Predicting ETF Performance: A Focus on the U.S. and Emerging ESG ETF Markets

This study investigates the predictive performance of Artificial Neural Network (ANN) models, the Backpropagation Neural Network (BPN), Recurrent Neural Network (RNN), Radial Basis Function Neural Network (RBFNN), and Time-Delay Recurrent Neural Network (TDRNN), in forecasting the returns of U.S. and Emerging ESG and conventional Exchange-Traded Funds (ETFs). The Grey Relational Analysis results showed that the most influential variables on the ETF return were the Dollar Index (DXY), S&P 500 Index (SPX), and Volatility Index (VIX). The paper demonstrated various ANN model configurations, which included numbers of epochs, batch sizes, validation splits, and dataset sizes. Through testing Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Efficiency (CE), the BPN and RNN models outperformed the others in all configurations, having a configuration of 30 epochs and using a 70 batch size, uses 50% of the training data provides proper ways for both U.S. and Emerging ETFs. Conventional ETFs generally outperform their counterpart ESG ETFs. The Diebold-Mariano (DM) test confirms statistically significant differences in prediction performance between the RNN, RBFNN, and TDRNN models. This study emphasizes the importance of choosing the right ANN model to bring forth greater prediction efficacy toward capital markets, with a special focus on ESG and conventional ETF domains from the perspectives of policy, fund managers, and investors' decisions. JEL Codes: C45, G11, G15, Q56