Forecasting and trading cryptocurrencies with machine learning under changing market conditions

Area of Science:

• Quantitative Finance
• Computational Economics
• Data Science

Background:

• Cryptocurrency markets exhibit high volatility and complex dynamics.
• Predicting cryptocurrency price movements is challenging due to market inefficiencies and external factors.
• Machine learning offers potential tools for analyzing and forecasting financial time series data.

Purpose of the Study:

• To assess the predictability of Bitcoin, Ethereum, and Litecoin prices.
• To evaluate the profitability of trading strategies based on machine learning models.
• To test model performance across different market conditions, including turmoil and bear markets.

Main Methods:

• Utilized machine learning techniques including linear models, random forests, and support vector machines.
• Employed classification and regression methods using trading and network activity data.
• Developed ensemble trading strategies, notably 'Ensemble 5', for enhanced signal generation.

Main Results:

• Five out of 18 individual models showed success rates below 50% during the test period.
• The 'Ensemble 5' strategy achieved the highest performance for Ethereum and Litecoin.
• Annualized Sharpe ratios reached 80.17% for Ethereum and 91.35% for Litecoin, with positive returns after costs.

Conclusions:

• Machine learning techniques demonstrate robustness in predicting cryptocurrency price movements.
• Profitable trading strategies can be devised using machine learning, even in adverse market conditions.
• The study supports the utility of advanced computational methods for cryptocurrency market analysis.