Comprehensive Modeling Approaches for Forecasting Bitcoin Transaction Fees: A Comparative Study

TL;DR

This study tackles Bitcoin transaction fee forecasting over a 24-hour horizon using six approaches, including SARIMAX, Prophet, Time2Vec (with and without attention), TFT, and a SARIMAX-Gradient Boosting hybrid, on a 91-day, 11,809-record dataset with 23 features. Through expanding-window 5-fold cross-validation and a fixed 144-block test set, traditional statistical models (notably SARIMAX) outperform deep learning architectures, with SARIMAX achieving the best cross-validation and test metrics ($MAE$ and $RMSE$) and deep models suffering from data limitations. The hybrid model offers robustness by combining linear trend capture with non-linear adjustments, while smoothing effects in long-horizon forecasts provide practical guidance for fee-sensitive planning. The results highlight the current practical advantage of classical methods in this domain and establish a baseline for future enhancements involving hybridization, transfer learning, and interpretable deep architectures for cryptocurrency analytics.

Abstract

Transaction fee prediction in Bitcoin's ecosystem represents a crucial challenge affecting both user costs and miner revenue optimization. This study presents a systematic evaluation of six predictive models for forecasting Bitcoin transaction fees across a 24-hour horizon (144 blocks): SARIMAX, Prophet, Time2Vec, Time2Vec with Attention, a Hybrid model combining SARIMAX with Gradient Boosting, and the Temporal Fusion Transformer (TFT). Our approach integrates comprehensive feature engineering spanning mempool metrics, network parameters, and historical fee patterns to capture the multifaceted dynamics of fee behavior. Through rigorous 5-fold cross-validation and independent testing, our analysis reveals that traditional statistical approaches outperform more complex deep learning architectures. The SARIMAX model achieves superior accuracy on the independent test set, while Prophet demonstrates strong performance during cross-validation. Notably, sophisticated deep learning models like Time2Vec and TFT show comparatively lower predictive power despite their architectural complexity. This performance disparity likely stems from the relatively constrained training dataset of 91 days, suggesting that deep learning models may achieve enhanced results with extended historical data. These findings offer significant practical implications for cryptocurrency stakeholders, providing empirically-validated guidance for fee-sensitive decision making while illuminating critical considerations in model selection based on data constraints. The study establishes a foundation for advanced fee prediction while highlighting the current advantages of traditional statistical methods in this domain.

Figures (5)

• Figure 1: Correlation Heatmap of Features.
• Figure 2: Hybrid Fee Rate Prediction Model Algorithm.
• Figure 3: Temporal dynamics of Bitcoin transaction activity, network state, and fee dynamics during a two-day observation period (2024-10-01 to 2024-10-02). The figure highlights significant variability in transaction counts, block intervals, mempool metrics, hash rate, and Bitcoin price, as well as the volatility of the block median fee rate. These patterns underscore the dynamic nature of the Bitcoin network and the interconnected relationships between congestion, fee-setting behaviors, and network security.
• Figure 4: Comparison of Cross-Validation Results Across Multiple Time Series Models: SARIMAX, Prophet, Hybrid, Temporal Fusion Transformer (TFT), Time2Vec, and Time2Vec + Attention. Blue lines represent actual values, while orange lines show predicted fee rates across different validation folds.
• Figure 5: Comparison of Test Results Across Multiple Time Series Models: SARIMAX, Prophet, Hybrid, Temporal Fusion Transformer (TFT), Time2Vec, and Time2Vec + Attention. Results show model predictions over the 144-block test period, with blue lines indicating actual fee rates and orange lines showing predictions.