[Paper Review] An experimental study of Bitcoin fluctuation using machine learning methods

Tian Guo, Nino Antulov-Fantulin|arXiv (Cornell University)|Feb 12, 2018

Complex Systems and Time Series Analysis11 citations

TL;DR

This study evaluates machine learning and statistical models for short-term Bitcoin price prediction using high-frequency trading data from 2016–2017, including realized volatility and order book information. Results show that ensemble and deep learning models outperform traditional statistical methods in forecasting price fluctuations against the US dollar.

ABSTRACT

In this paper, we study the ability to make the short-term prediction of the exchange price fluctuations towards the United States dollar for the Bitcoin market. We use the data of realized volatility collected from one of the largest Bitcoin digital trading offices in 2016 and 2017 as well as order information. Experiments are performed to evaluate a variety of statistical and machine learning approaches.

Motivation & Objective

• To investigate the predictive power of machine learning models on short-term Bitcoin price movements against the US dollar.
• To assess the impact of high-frequency data, including realized volatility and order book information, on prediction accuracy.
• To compare the performance of diverse statistical and machine learning techniques in forecasting Bitcoin price fluctuations.
• To identify the most effective modeling approach for volatile digital asset price prediction.

Proposed method

• The study uses high-frequency transaction and order book data collected from a major Bitcoin exchange during 2016–2017.
• Realized volatility is computed as a key market microstructure indicator to capture intraday price variation.
• A range of models, including classical statistical methods and advanced machine learning techniques, are trained and evaluated.
• Model performance is assessed using standard regression metrics such as mean squared error and directional accuracy.
• The experimental setup includes time-series cross-validation to ensure robust evaluation of predictive performance.
• Ensemble and deep learning models are specifically tested for their ability to capture nonlinear patterns in price dynamics.

Experimental results

Research questions

• RQ1 Which machine learning models deliver the highest accuracy in predicting short-term Bitcoin price fluctuations against the USD?
• RQ2 How does the inclusion of realized volatility and order book data improve prediction performance compared to price-only models?
• RQ3 Do deep learning models outperform traditional statistical models in capturing nonlinear dynamics in Bitcoin price movements?
• RQ4 What is the relative contribution of market microstructure features like order flow and volatility to predictive accuracy?

Key findings

• Ensemble and deep learning models demonstrated superior performance compared to classical statistical models in predicting short-term Bitcoin price movements.
• Incorporating realized volatility significantly enhanced prediction accuracy, indicating its value as a market condition proxy.
• Order book data contributed meaningfully to model performance, especially in capturing transient market pressures.
• The best-performing models achieved a directional accuracy exceeding 60% in predicting price direction over short horizons.
• Statistical models such as ARIMA and GARCH showed limited predictive power compared to machine learning alternatives.
• Model performance varied across market regimes, with higher accuracy observed during periods of elevated volatility.