Predicting Crypto Sentiment Using Blockchain Data

Summary & Methodology Analysis

The research architecture centers on a multi-source data pipeline. The team ingested daily Bitcoin on-chain metrics alongside historical pricing from Blockchair. They processed raw Twitter sentiment from a Kaggle dataset, applying feature engineering to condense daily tweets into a majority-class label. To prepare the input vectors, the team merged the data on registration dates and applied MinMax Scaler normalization across 30-record intervals to standardize the feature scales before feeding them into the model pipeline. The study evaluated multiple architectures, including LSTM, RNN, CNN, SARIMA, SAM-LSTM, and jSO-LSTM, to establish a baseline for classification performance. Ultimately, the authors selected Gradient Boosting, or XGBoost (a decision-tree-based ensemble method), as the primary model. This choice was deliberate, favoring the structural transparency and interpretability of tree-based approaches over the high-dimensional complexity of deep learning models. The authors then utilized SHAP (a game-theoretic approach to explain the output of machine learning models) to map the influence of on-chain features back to the predicted classification. This allows developers to audit why the model arrived at a specific sentiment state. While the model achieved a stable F1-score of 0.84, the paper explicitly acknowledges a limitation: the shift away from deep learning reflects a trade-off where the authors sacrificed potential predictive gains to ensure that the model remains explainable and audit-friendly in production environments.