Do You Care About Your Positions? Users Under Liquidation Risk in Decentralized Lending Protocol

Abstract

Lending protocols have transformed the Decentralized Finance (DeFi) ecosystem, driving innovation while also introducing new risks. This study develops a machine learning framework to predict user behavior and assess factors influencing changes in health ratios within the Compound V2 protocol. By analyzing user historical data, position metrics, and market conditions, we propose machine learning-based models to predict whether users will adjust their positions or face liquidation. We find that Random Forest and XGBoost models excel in predicting these outcomes, with features like collateral values, historical risk exposure, and asset composition playing significant roles. Additionally, panel regression models reveal insights into health ratio dynamics over time and across asset types, as well as user sophistication. These findings offer a better understanding of user behavior, highlighting opportunities for improved risk modeling and adaptive strategies in DeFi lending.