Predicting Bitcoin's price is close to a coin flip, and this paper wisely doesn't try. Instead it aims to explain market mood by pulling together three signals: on-chain data (the public record of transactions), historical price, and a daily read of Twitter sentiment, all stitched into one normalized dataset so you can see how the mood tracks with activity on the chain.

They tried several machine learning models with cross-validation, and gradient boosting (XGBoost) came out most reliable, landing an average F1 around 0.84 for classifying sentiment. To keep it from being a black box, they used SHAP to measure how much each on-chain feature moved the predictions, which is a sensible touch. The authors frame this as a starting point for data-driven crypto analysis, with deep learning as a next step.

This is a modest result read off the abstract, so check the paper for how sentiment was labeled and how it holds up out of sample.