Evaluating the Performance of Metaheuristic Based Artificial Neural Networks for Cryptocurrency Forecasting

The irregular movement of cryptocurrency market makes effective price forecasting a challenging task. Price fluctuations in cryptocurrencies often appear to be arbitrary that has been a hot topic. Though various statistical and econometric forecasting models exists, still there is lack of advanced artificial intelligence models to explain behaviour of such price fluctuations. Artificial neural networks (ANNs) are data-driven models and can effectively handle complex nonlinear functions in presence of abundant data. However, optimal parameter tuning of such models with conventional back propagation-based learning entail domain expertise, higher computational cost, and yield inferior accuracy thus, makes its use tough. Contrast to this, metaheuristic-based ANN training has been emerging as an efficient learning paradigm. This article constructs few optimal ANNs through three efficient metaheuristics with less control parameters such as fireworks algorithm (FWA), chemical reaction optimization (CRO), and the teaching–learning based optimization (TLBO) separately. The role of a metaheuristic is to investigate the near-optimal weights and thresholds of an ANN of solitary hidden layer and thereby ensuring a higher degree of accuracy. The hybrid models are then used to simulate and predict the behaviour of four fast growing cryptocurrencies such as Bitcoin, Litecoin, Ethereum, and Ripple. Various experiments are carried out using real time cryptocurrency data and hybrid ANNs through four performance measures. We undertake a comparative performance analysis of forecasting models and Friedman tests to demonstrate the superiority and statistical significance. In particular, ANN trained with CRO, TLBO, and FWA obtained an average rank of 1, 2, and 2.75 respectively.