Forecasting Cryptocurrency Prices: An Inter-Exchange Feature Engineering Approach Using Ensemble Regression Models

In this paper, we tackle the challenge of short-term price prediction in the notoriously volatile cryptocurrency market. Our approach leverages inter-exchange signals within a machine-learning framework. Focusing on Ethereum (ETH), we analyze two complementary daily series: (i) high-granularity trading activity from Binance and (ii) a broad market index used as a benchmark for comparison. Our approach uses extensive feature engineering, including traditional technical indicators, volatility metrics, and new inter-exchange features that are specifically designed to capture possible lead-lag signals between markets, because these features are crucial for predicting the next day's closing price. We frame the task as supervised regression to predict the next day's closing price and compare three models - Linear Regression, Random Forest Regressor, and Gradient Boosting Regressor - with time-series split to mimic real-world deployment. Somewhat surprisingly, Linear Regression was the best-performing model, with an R 2 of 0.9894 and an RMSE of 55.49, indicating that the engineered features relate to the target in a largely linear fashion; therefore, a feature importance analysis was conducted, which further demonstrates that the general market price is the strongest predictor, confirming that inter-exchange information improves forecasting. Our results show that careful feature engineering, combined with simple linear and ensemble models, can produce highly accurate short-term cryptocurrency price predictions.