A Machine Learning Approach For Bitcoin Forecasting

TL;DR

Bitcoin price forecasting is challenging when relying on closing price alone. The authors assemble a rich feature set including 13 time-series signals from on-chain, market, sentiment, and a Variational Mode Decomposition (VMD) of the price, and evaluate several ML models and an ensemble. They find that Open, High, and Low prices carry most predictive power, while non-price signals contribute negligibly, and a GRU-based ensemble with a simple return/baseline framework achieves strong directional accuracy, comparable to the LMH-BiLSTM baseline. The work demonstrates that combining short-term price action signals with a GRU ensemble yields robust 1-step-ahead forecasts in cryptocurrency markets and highlights the practical value of directional-forecast performance for trading.

Abstract

Bitcoin is one of the cryptocurrencies that is gaining more popularity in recent years. Previous studies have shown that closing price alone is not enough to forecast stock market series. We introduce a new set of time series and demonstrate that a subset is necessary to improve directional accuracy based on a machine learning ensemble. In our experiments, we study which time series and machine learning algorithms deliver the best results. We found that the most relevant time series that contribute to improving directional accuracy are Open, High and Low, with the largest contribution of Low in combination with an ensemble of Gated Recurrent Unit network and a baseline forecast. The relevance of other Bitcoin-related features that are not price-related is negligible. The proposed method delivers similar performance to the state-of-the-art when observing directional accuracy.

Figures (3)

• Figure 1: Visual summary of the method.
• Figure 2: Variational Mode Decomposition (VMD) decomposition of Bitcoin's closing price from 7 October 2013 to 6 November 2022.
• Figure 3: Feature importance found by LightGBM. Most important features are Low, High and Open.