Authors: Dr. Swathi Pothala
Abstract: The crypto currency markets are highly volatile with high levels of noise and non-stationary properties that pose challenges for classical forecasting techniques. In this paper, we intro-duce a hybrid approach which is based on time-series decomposition and deep learning to forecast the volatility of cryptocurrencies. The proposed method consists of using the Com-plete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) for de-composition of the signal, Bidirectional Gated Recurrent Unit (Bi-GRU) along with multi-head self-attention for extracting temporal features, and a Particle Swarm Optimization (PSO) for tuning the hyperparameters. The performance evaluation of the introduced ap-proach was done using data for five major cryptocurrencies (Bitcoin, Ethereum, Binance Coin, Cardano, Solana) in the period from 2019 to 2025. The obtained MAE is equal to 0.0058, RMSE – 0.0082, and MAPE is equal to 4.32%. We managed to outperform the popular GARCH family of models (GARCH – 0.0182 MAE, EGARCH – 0.0156 MAE).
DOI: http://doi.org/10.5281/zenodo.20672600
