A differential evolution-based regression framework for forecasting Bitcoin price.

Area of Science:

• Computational Finance
• Machine Learning
• Time Series Analysis

Background:

• Accurate Bitcoin price forecasting is challenging due to market volatility.
• Existing predictive models often struggle with the complex dynamics of cryptocurrency markets.
• Advanced regression techniques are needed for reliable financial time series prediction.

Purpose of the Study:

• To develop and evaluate a differential evolution-based regression framework for one-day-ahead Bitcoin price forecasting.
• To enhance prediction accuracy by decomposing time series into linear and nonlinear components.
• To compare the proposed framework against six established machine learning algorithms.

Main Methods:

• Maximal Overlap Discrete Wavelet Transform (MODWT) for time series decomposition.
• Polynomial Regression with Interaction (PRI) and Support Vector Regression (SVR) applied to decomposed components.
• Differential Evolution (DE) for optimizing PRI coefficients and SVR hyperparameters.

Main Results:

• The proposed differential evolution-based regression framework achieved higher forecast accuracy than all six compared algorithms.
• Component-wise projections from PRI and SVR, optimized by DE, significantly improved prediction performance.
• Numerical experiments on historical Bitcoin data (2013-2019) and Monte Carlo simulations validated the framework's efficacy.

Conclusions:

• The novel regression framework integrating MODWT, PRI, SVR, and DE offers a robust approach for Bitcoin price prediction.
• This method effectively captures both linear and nonlinear patterns in financial time series.
• The findings suggest a promising direction for improving cryptocurrency forecasting accuracy.