Decoding Market Emotion from Blockchain Activity: A Data-Driven Sentiment Classifier

Summary§

This paper presents a data-driven approach to classify Bitcoin market sentiment into bullish, bearish, or neutral categories by fusing on-chain metrics (e.g., transaction volume, active addresses), price data (returns, volatility), and daily Twitter sentiment scores. The authors benchmark several classifiers under cross-validation; Gradient Boosting (XGBoost) yields the best average F1-score of 0.84. Model interpretability is provided via SHAP values, quantifying the impact of each on-chain feature on predictions.

Methodology§

Dataset Construction§

• On-chain data: Blockchain transaction counts, average transaction value, miner revenue, hash rate, etc.
• Financial data: Daily close price, returns, volatility, trading volume.
• Twitter sentiment: Precomputed daily sentiment polarity scores (positive/negative/neutral percentages).

All features are normalized (z-score) and aligned by date. The target variable is discretized sentiment from Twitter (3 classes).

Model Training§

Six classifiers are compared: logistic regression, random forest, SVM, k-NN, AdaBoost, and XGBoost. Stratified 5-fold cross-validation is used. XGBoost hyperparameters are tuned via grid search: number of estimators (100–500), max depth (3–10), learning rate (0.01–0.3), subsample ratio, etc.

XGBoost Objective§

XGBoost minimizes the regularized objective: $$\mathcal{L} = \sum_{i=1}^{n} l(y_i, \hat{y}_i) + \sum_{k=1}^{K} \Omega(f_k)$$ where $l$ is the softmax cross-entropy loss for multi-class classification, and $\Omega(f_k) = \gamma T + \frac{1}{2}\lambda \|w\|^2$ penalizes tree complexity.

SHAP for Interpretability§

SHAP values are computed for each feature $j$ as: $$\phi_j = \sum_{S \subseteq F \setminus \{j\}} \frac{|S|!(|F|-|S|-1)!}{|F|!} [f_{S \cup \{j\}}(x_{S \cup \{j\}}) - f_S(x_S)]$$ where $F$ is the set of all features, and $f_S$ is the model trained only on features in $S$. This provides additive feature importance scores for each prediction.

Code Snippet (PyTorch-like for reference)§

import xgboost as xgb
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import f1_score
import numpy as np

# Assume X (features), y (labels) are prepared
params = {
    'objective': 'multi:softprob',
    'num_class': 3,
    'max_depth': 6,
    'eta': 0.1,
    'subsample': 0.8,
    'colsample_bytree': 0.8,
    'eval_metric': 'mlogloss'
}

skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
f1_scores = []

for train_idx, val_idx in skf.split(X, y):
    X_train, X_val = X[train_idx], X[val_idx]
    y_train, y_val = y[train_idx], y[val_idx]
    dtrain = xgb.DMatrix(X_train, label=y_train)
    dval = xgb.DMatrix(X_val, label=y_val)
    model = xgb.train(params, dtrain, num_boost_round=100, early_stopping_rounds=10, evals=[(dval, 'val')], verbose_eval=False)
    preds = np.argmax(model.predict(dval), axis=1)
    f1_scores.append(f1_score(y_val, preds, average='weighted'))

print('Mean F1:', np.mean(f1_scores))

# SHAP explanation
import shap
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_val)
shap.summary_plot(shap_values, X_val, feature_names=feature_names)

Results§

XGBoost achieves the highest average weighted F1-score of 0.84 across folds, outperforming other models by ~5–10%. SHAP analysis reveals that the most important features are Twitter sentiment polarity, transaction volume change, and Bitcoin returns from the previous day.

Conclusion§

The combination of on-chain, price, and social media data yields robust sentiment classification. Future work may incorporate deep learning models (e.g., LSTM) to capture temporal dependencies.