Time-series forecasting of Bitcoin prices using high-dimensional features: a machine learning approach

Area of Science:

• * Computational Finance
• * Data Science
• * Cryptography

Background:

• * Bitcoin is a decentralized digital asset utilizing blockchain technology for peer-to-peer transactions.
• * Decentralized cryptocurrencies like Bitcoin are susceptible to significant price volatility and non-stationary behavior.
• * Understanding and predicting Bitcoin's price dynamics is crucial due to its market instability.

Purpose of the Study:

• * To develop and evaluate high-performance machine learning models for predicting Bitcoin price movements.
• * To extend prediction horizons beyond one day, analyzing one, seven, thirty, and ninety-day forecasts.
• * To compare the efficacy of machine learning classification and regression models against existing literature.

Main Methods:

• * Implementation of machine learning-based classification and regression models.
• * Application of models to predict Bitcoin prices over multiple short and medium-term horizons (1, 7, 30, 90 days).
• * Evaluation of model performance using accuracy metrics for classification and error percentages for regression.

Main Results:

• * Classification models achieved up to 65% accuracy for next-day Bitcoin price forecasts.
• * For longer horizons (7-90 days), classification accuracy ranged from 62% to 64%.
• * Daily Bitcoin price forecast error was as low as 1.44%, with 7-90 day forecast errors between 2.88% and 4.10%.

Conclusions:

• * The developed machine learning models demonstrate high performance and feasibility for Bitcoin price prediction.
• * These models significantly outperform existing methods in the literature for both short and medium-term forecasting.
• * The study highlights the potential of machine learning in addressing cryptocurrency price volatility and non-stationarity.