A Comparative Study of Prophet and SARIMAX for Transaction Fee Forecasting in the Bitcoin Network

This study compares three time series forecasting approaches Prophet, SARIMAX, and LSTM for predicting the daily average transaction fee rate in the Bitcoin network. Data is processed through a consistent pipeline, including timestamp normalization, daily resampling, and handling of extreme values to maintain integrity and fairness across models. Experiments were conducted using an 80/20 chronological train-test split and evaluated using MAE, RMSE, and sMAPE metrics. Results show that the extreme spike in early May 2023 was the dominant source of error for all models: in the absence of event indicators or leading covariates related to network congestion, no approach was able to anticipate the spike ex ante. Outside the extreme period, Prophet was the most stable in tracking trend and weekly seasonality, LSTM tended to over-forecast following the spike, while SARIMAX showed strong potential after proper alignment of timestamps and exogenous regressors. This study highlights the importance of integrating intervention features (pulse/level-shift), causal covariates, and segmented regime-based evaluation to enhance model robustness. These findings provide a methodological foundation for developing more accurate, interpretable, and adaptive forecasting systems for Bitcoin transaction fees in response to evolving network dynamics.