"Evaluating the Impact of PCA-Based Dimensionality Reduction on Bitcoin Transaction Forecasting: A Comparative Study of XGBoost, LSTM, and GNN"

Abstract

Accurate forecasting of Bitcoin dynamics is essential for digital asset management. While existing literature primarily focuses on market price prediction using ensemble and deep learning, this study extends the frontier by analyzing high-dimensional on-chain transaction data. We present a comparative evaluation of XGBoost, LSTM, and Graph Neural Networks (GNN), specifically investigating the impact of Principal Component Analysis (PCA) on model stability. Our findings reveal a 'PCA Paradox': while dimensionality reduction enhances the performance of GNN and LSTM by filtering noise, it marginally reduces the precision of XGBoost. Results show that XGBoost achieves the highest numerical accuracy (RMSE: 0.018), whereas GNN and LSTM provide superior trend stability. This research provides critical insights into feature engineering for blockchain-based financial systems