Explainable machine learning-based prediction of deviations from triangular arbitrage parity across stablecoin and forex markets: addressing nonstationarity

I propose an explainable machine learning-based framework for non-linear classification of deviations from triangular arbitrage parity (TAP) and exploration of the importance of wavelet-based time-frequency features and rolling statistics of the deviations in prediction of stablecoin arbitrage opportunities, taking TAP deviations across Tether and eighteen fiat currencies since 2025 as a case study. Two conclusions emerge: (1) persistence of TAP deviations across Tether and rupee or Argentine peso currency pairs in the sampling period seems more like capital-control effects than arbitrage; and (2) potential arbitrage opportunities are predictable via machine learning under nonstationary parity deviations and limits to arbitrage. The article provides insights into prediction of arbitrage opportunities from parity deviations and examination of market efficiency in the digital world.