A Hybrid CNN–LSTM Model for Transaction Fee Forecasting in Post EIP-1559 Ethereum

Abstract

Over the past decade, cryptocurrencies have experienced a significant surge in popularity. Several factors have contributed to their rise. First, the decentralized nature of cryptocurrencies, enabled by blockchain technology, has appealed to individuals seeking financial autonomy and freedom from traditional banking systems. Additionally, the potential for substantial financial gains, as demonstrated by the surge in the value of Bitcoin and Ethereum. Cryptocurrency transactions require the sender to include a transaction fee before initiating. Concerning the Ethereum protocol, the transaction fee calculation before the London upgrade, i.e., Ethereum Improvement Proposals (EIP-1559), led to delayed transaction confirmation and increased congestion in the Ethereum network. Ever since this upgrade, the transaction fee has become dynamic and user-friendly such that transactions get confirmed within a reasonable time. For such a scenario, the need of the hour for an effective forecasting technique can prove critical from the user’s point of view. After the EIP-1559 upgrade, there is a lack of literature that efficiently utilizes cryptocurrency transaction data’s time-series nature. To solve these issues, this paper proposes a hybrid deep learning model to predict total transaction fees in post EIP-1559 Ethereum precisely. The proposed convolutional neural network (CNN)-long short term memory (LSTM) leverages the advantages of convolutional layers and is followed by effective learning of time-series dependencies between the data by LSTM layers. The experimentation and comparison with state-of-the-art suggest significant improvement when CNN–LSTM is leveraged for this type of forecasting. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2024.