Forecasting Bitcoin closing price series using linear regression and neural networks models

Area of Science:

• Financial econometrics
• Computational finance
• Machine learning applications

Background:

• Cryptocurrency price behavior remains largely unexplored, offering opportunities to compare with traditional financial markets.
• Existing research on cryptocurrency forecasting and stock market predictions provides benchmarks for evaluating new models.

Purpose of the Study:

• To forecast daily closing prices of Bitcoin, Litecoin, and Ethereum.
• To compare the performance of statistical and machine learning models in cryptocurrency price prediction.
• To investigate the impact of price "regimes" on forecasting accuracy.

Main Methods:

• Implemented Simple Linear Regression (SLR) and Multiple Linear Regression (MLR) models.
• Utilized artificial neural networks: Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM).
• Analyzed time series data including prices and trading volumes, partitioning into distinct price regimes.

Main Results:

• While the entire time series resembled a random walk, partitioning data into shorter sequences representing different price regimes enabled precise forecasting.
• Models achieved high accuracy, evaluated by Mean Absolute Percentage Error (MAPE) and relative Root Mean Square Error (relativeRMSE).
• The best results were obtained using multiple previous price points, indicating the existence of predictable time regimes.

Conclusions:

• Cryptocurrency price series can be accurately forecasted by identifying and analyzing distinct price regimes.
• Machine learning models, particularly LSTMs and MLPs, show strong performance in capturing these regime-specific dynamics.
• The developed models offer improved forecasting accuracy and time complexity compared to existing benchmark studies, advancing the state-of-the-art in cryptocurrency price prediction.