Comparative Performance Analysis of Deep Learning Models for Cryptocurrency Price Forecasting

Abstract

Purpose: A cryptocurrency's high volatility and nonlinear market dynamics make it extremely difficult to predict its price with any degree of accuracy. This study aims to evaluate and contrast the predictive capabilities of five Deep Learning architectures for the same reason: LSTM, GRU, BiLSTM, Transformer, and Performer, to identify the best model capable of predicting the price of cryptocurrencies. It is aimed at providing an empirical base for making such predictions with high reliability in such volatile financial markets.

Methods: The dataset used in this study, namely the price per minute data for BTC, ETH, BNB, and XRP, was obtained from Kaggle. Data processing includes normalization using MinMaxScaler and sequence generation through the Sliding Window technique. An 80:20 data split is used to train and validate each deep learning model, and four metrics consisting of MAE, MSE, RMSE, and MAPE are used for evaluation. Standardized experimental protocols were guaranteed by Python-based frameworks. 

Result: The Transformer model created the best results for the lowest MAPE value across all datasets, the smallest being BTC and ETH at 0.20%, BNB at 0.29%, and XRP at 0.36% demonstrating high accuracy and generalization. The BiLSTM was ranking second since it captured effectively the bidirectional temporal dependencies; the GRU was moderate but stable in its performance. The data showed that the accuracy of LSTM and Performer varied.

Novelty: This research provides a comprehensive comparison between various models, highlighting the Transformer's self-attention mechanism as the most superior in capturing long-term temporal dependencies and nonlinear market behavior compared to other deep learning methods. These findings provide valuable insights for the development of advanced AI-based forecasting frameworks in financial analysis.