Analyzing the performance of geometric mean optimization-based artificial neural networks for cryptocurrency forecasting

In this study, we utilize a recently proposed non-parametric metaheuristic algorithm known as geometric mean optimization (GMO) to adjust the hidden layer input weights and bias of six ANN variants, namely PSNN, SPNN, JPSNN, FLANN, RBFN, and MLP, thereby creating six hybrid models. Later on, we engage all these hybrid models to predict the closing price of four widely used cryptocurrencies. For comparison purposes, we also engage the widely popular traditional GD algorithm to fix the hidden layer input weights and biases of the same set of ANNs, creating six alternate models and employing those models for the same prediction task. In order to assess the effectiveness of each model, we employ the mean squared error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE) statistics. The experimental result shows the superiority of GMO-based PSNN over others. We then compare the GMO-based PSNN's outcomes with some of the existing hybrid models in the literature.