Abstract
The main objectives of this study are threefold. Firstly, to develop a machine learning model that can timely predict multinomial de-pegging events (i.e., 24 hours before they occur). Secondly, this study also seeks to explain the drivers of both discount and premium de-pegging events. The third objective is to examine when a de-pegging event occurs and how long it takes for the stablecoin to return to its peg.
Four stablecoins are chosen to investigate these objectives, which include USDT, USDC, TUSD, and FRAX for three years (i.e., from January 2022 to December 2024). This observation period was chosen due to its recency and relevance to the implementation of major cryptocurrency regulatory frameworks in Europe, such as MiCA. The features used to develop the machine learning models are derived from stablecoin trading price and information, market cap, trading supply, and volatility data. Then, three machine learning models are developed, namely Logistic Regression, Random Forest, and XGBoost, that use the feature data as input, and a dynamic de-pegging target variable is derived based on work by Carey (2023) and Lee et al. (2025). The created target variable is a multinomial target instead of the binary target found in the literature. The Synthetic Minority Over-sampling Technique (SMOTE) is used to account for the class imbalance. The findings are that non-linear models (i.e., Random Forest and XGBoost) consistently outperform classic linear models (i.e., Logistic Regression). These findings are consistent with Lee et al. (2025), who found the same results while modelling de-pegging events as a binary target.
The findings from the co-integration analysis show that the premium de-pegging is consistently faster than discount de-pegging events for the selected stablecoins. Secondly, the findings also show that the speed of adjustment varies depending on the type of stablecoin. For example, if there is a discount de-pegging for USDT, 18.6% of the adjustment happens each day. In contrast, the same discount de-pegging for USDC results in 66.7% of the disequilibrium being corrected daily, indicating a faster adjustment, like FRAX, with a 55% daily speed of adjustment. Meanwhile, for TUSD, the adjustment is slow, with only 3.45% corrected daily during discount de-pegging.
These findings will help traders and investors prepare for possible de-pegging events in the stablecoin secondary market and manage their risk. That is, since we...