An Early Warning System for Liquidity Stress in Cryptocurrency Markets Using Trade Flow Analysis and Machine Learning

Abstract

Most research on liquidity crises in financial markets relies on expensive institutional data or coarse daily aggregates, leaving little room for reproducible, operationally deployable risk monitoring. We propose an early warning system for liquidity stress in cryptocurrency markets using freely available high-frequency trade data from major exchanges. Unlike prior work that evaluates models using static accuracy measures, we emphasise the evaluation gap: standard classification metrics ignore the temporal dimension critical for risk management, where a perfect prediction with zero lead time offers no practical value. Our contribution is fourfold. First, we introduce a regimeaware labelling method that adapts stress-event definitions to different volatility conditions using trade-based liquidity indicators, ensuring both statistical balance and economic realism. Second, we develop and benchmark a suite of machine learning models—from interpretable baselines to advanced sequence architectures—using a purged expanding walk-forward validation framework on four years of five-minute resolution data across Bitcoin, Ethereum, and Solana. Third, we provide rigorous explainability analysis using SHAP feature attribution and Integrated Gradients, with surrogate fidelity validation confirming attribution validity at a mean agreement rate of 91.8%. Fourth, we address the methodological concern that HMM stress labels are self-referential through a three-tier external validation framework and a controlled experiment comparing global versus locally-fitted HMMs across all four walk-forward folds, using genuinely independent validators— price drawdown, documented crisis timestamps, and crossasset simultaneous stress —none of which were HMM training features.