Latent Velocity Spikes as Label-Free Market Instability Alerts
Abstract
Detecting market regime shifts is challenging due to non-stationarity and the lack of reliable labels. Conventional approaches—such as regime-switching models—often rely on discrete state assumptions that break down during gradual transitions. We propose a label-free transition monitoring framework that treats the market as a continuously evolving latent state and detects instabilities through abrupt changes in its dynamics. Using BYOL-style self-supervision with asymmetric augmentations on return-volatility windows, we learn compact market representations without discrete regimes. On a period held-out test (2022–2026), spikes exhibit low overlap with a 2-state Gaussian HMM, suggesting a complementary transition signal. Importantly, these spikes are followed by worse forward drawdown outcomes: 21-day forward maximum drawdown (MDD) deteriorates after spikes, with a stronger relative worsening for S&P 500(-0.35) and a weaker shift for NASDAQ(-0.11). Finally, a downstream volatility test shows increased prediction error on spike days, supporting latent velocity as a practical instability alert. Our approach provides a practical instability alert that can inform risk-aware decisions under changing market conditions.