How Attention-Based Models Outperform Traditional Bitcoin Fee Estimators

The article presents FENN, a neural network framework designed to enhance Bitcoin transaction fee estimation accuracy. Unlike traditional models like BCore and BtcFlow, FENN integrates transaction features, mempool states, and network data using LSTM and attention mechanisms. Experimental results across multiple datasets show that FENN, particularly the Adv and LSTMadv variants, consistently achieve lower RMSE and MAPE scores, confirming superior performance. Moreover, FENN adapts quickly to new blockchain data, demonstrating both efficiency and scalability for real-world use.