Graph Neural Networks and Meta-Learning for Cryptocurrency Price Prediction

Determining future cryptocurrency prices proves difficult because the market shows strong unpredictability across many factors. This research uses GNNs and metalearning to develop better price prediction models that stay effective across different cryptocurrencies. The hybrid model uses GNNs to analyze blockchain transaction relationships and meta-learning to switch settings quickly across different cryptocurrencies with limited data information. Our model worked with different cryptocurrencies and showed better results compared to ARIMA and LSTM when testing Bitcoin, Ethereum, Litecoin, and Ripple. Experimental tests prove that the combined model beats ARIMA and LSTM in every evaluative benchmark. For Bitcoin the model produced hybrid results with an MAE of 0.032 an RMSE of 0.056 along with an $\mathbf{R}^{\mathbf{2}}$ value 0.91 which outperformed ARIMA’s MAE of 0.070 RMSE of $\mathbf{0. 0 8 9}$ and $\mathbf{R}^{\mathbf{2}}$ score of $\mathbf{0. 7 5}$ and also surpassed LSTM’s MAE of 0.061 RMSE of 0.082 and $\mathbf{R}^{\mathbf{2}}$ score of $\mathbf{0. 7 8}$. Our model generalized well to untested cryptocurrencies because it earned MAE 0.048, RMSE 0.072, and $\mathbf{R}^{\mathbf{2}}$ score $\mathbf{0. 8 6}$ for Bitcoin Cash prediction. These tests prove that the hybrid model stands apart by accurately forecasting values in new markets. Using GNNs with meta-learning lets our framework handle cryptocurrency price prediction effectively across changing markets through many kinds of digital assets.