Helformer: an attention-based deep learning model for cryptocurrency price forecasting

Abstract

Cryptocurrencies have become a significant asset class, attracting considerable attention from investors and researchers due to their potential for high returns despite inherent price volatility. Traditional forecasting methods often fail to accurately predict price movements as they do not account for the non-linear and non-stationary nature of cryptocurrency data. In response to these challenges, this study introduces the Helformer model, a novel deep learning approach that integrates Holt-Winters exponential smoothing with Transformer-based deep learning architecture. This integration allows for a robust decomposition of time series data into level, trend, and seasonality components, enhancing the model’s ability to capture complex patterns in cryptocurrency markets. To optimize the model’s performance, Bayesian hyperparameter tuning via Optuna, including a pruner callback, was utilized to efficiently find optimal model parameters while reducing training time by early termination of suboptimal training runs. Empirical results from testing the Helformer model against other advanced deep learning models across various cryptocurrencies demonstrate its superior predictive accuracy and robustness. The model not only achieves lower prediction errors but also shows remarkable generalization capabilities across different types of cryptocurrencies. Additionally, the practical applicability of the Helformer model is validated through a trading strategy that significantly outperforms traditional strategies, confirming its potential to provide actionable insights for traders and financial analysts. The findings of this study are particularly beneficial for investors, policymakers, and researchers, offering a reliable tool for navigating the complexities of cryptocurrency markets and making informed decisions.