Paper 2023/1281

Leveraging Machine Learning for Bidding Strategies in Miner Extractable Value (MEV) Auctions

Abstract

The emergence of blockchain technologies as central components of financial frameworks has amplified the extraction of market inefficiencies, such as arbitrage, through Miner Extractable Value (MEV) from Decentralized Finance smart contracts. Exploiting these opportunities often requires fee payment to miners and validators, colloquially termed as bribes. The recent development of centralized MEV relayers has led to these payments shifting from the public transaction pool to private channels, with the objective of mitigating information leakage and curtailing execution risk. This transition instigates highly competitive first-price auctions for MEV. However, effective bidding strategies for these auctions remain unclear. This paper examines the bidding behavior of MEV bots using Flashbots' private channels, shedding light on the opaque dynamics of these auctions. We gather and analyze transaction data for the entire operational period of Flashbots, providing an extensive view of the current Ethereum MEV extraction landscape. Additionally, we engineer machine learning models that forecast winning bids whilst increasing profitability, capitalizing on our comprehensive transaction data analysis. Given our unique status as an adaptive entity, the findings reveal that our machine learning models can secure victory in more than 50% of Flashbots auctions, consequently yielding superior returns in comparison to current bidding strategies in arbitrage MEV auctions. Furthermore, the study highlights the relative advantages of adaptive constant bidding strategies in sandwich MEV auctions.