Empirical Analysis of EIP-3675: Miner Dynamics, Transaction Fees, and Transaction Time

TL;DR

This paper empirically analyzes the impact of Ethereum's transition to Proof of Stake (EIP-3675) on miner participation, network decentralization, and transaction fee dynamics, revealing increased miner participation but reduced randomness and decentralization.

Contribution

It provides the first comprehensive empirical study on EIP-3675's effects on miner behavior, network decentralization, and fee prediction using machine learning models.

Findings

Miner participation increased by approximately 50 times compared to PoW.

Miner distribution shifted towards more small miners, about 60 times more than PoW.

Transition reduced miner selection randomness, negatively impacting decentralization.

Abstract

The Ethereum Improvement Proposal 3675 (EIP-3675) marks a significant shift, transitioning from a Proof of Work (PoW) to a Proof of Stake (PoS) consensus mechanism. This transition resulted in a staggering 99.95% decrease in energy consumption. However, the transition prompts two critical questions: (1). How does EIP-3675 affect miners' dynamics? and (2). How do users determine priority fees, considering that paying too little may cause delays or non-inclusion, yet paying too much wastes money with little to no benefits? To address the first question, we present a comprehensive empirical study examining EIP-3675's effect on miner dynamics (i.e., miner participation, distribution, and the degree of randomness in miner selection). Our findings reveal that the transition has encouraged broader participation of miners in block append operation, resulting in a larger pool of unique miners ($\approx50\times$ PoW), and the change in miner distribution with the increased number of unique small category miners ($\approx60\times$ PoW). However, there is an unintended consequence: a reduction in the miner selection randomness, which signifies the negative impact of the transition to PoS-Ethereum on network decentralization. Regarding the second question, we employed regression-based machine learning models; the Gradient Boosting Regressor performed best in predicting priority fees, while the K-Neighbours Regressor was worst.