Joint-Embedding Predictive Learning of Latent Market States in U.S. Equities

We investigate whether Joint-Embedding Predictive Architectures (JEPA) can learn useful representations of U.S. equity markets. We jointly train a permutation-invariant tokenizer that maps each trading day's unordered per-asset features to a fixed set of learned factor tokens, together with a temporal JEPA using masked prediction to obtain a compact daily market-state embedding. Our evaluations show that these embeddings are strongly associated with second-moment market structure (realized volatility, correlation concentration, effective factor dimensionality) and weakly associated with market direction. The embedding helps predict gradual recovery dynamics but not sudden stress onsets. Without any text supervision, latent regimes show statistically significant alignment with news-topic shifts.

Financial markets move in two ways: direction, whether prices rise or fall tomorrow, driven by surprise news and nearly unpredictable; and the market's "risk weather", how turbulent trading is and how tightly stocks move together, which changes slowly. We asked whether a computer could learn to recognize this weather on its own, from daily prices of about 500 U.S. stocks. We used a method that teaches itself by hiding parts of the data and predicting what is missing, so it locks onto slow, underlying predictable patterns and ignores the noise (direction). The resulting representation captured turbulence and how stocks move together, reacted sharply to crashes but not routine announcements, and grouped days that shared similar news without ever reading a headline. It is not a trading system and we make no claims about predicting prices.