Market-Conditioned JEPA: Learning from Unpredictable Market Shocks

Market-Conditioned JEPA. Learning from Unpredictable Market Shocks. By Yuqi Wang and five others, from the University of Hong Kong and Grace Investment Machine. Published on the tenth of November.

Much of the variation in a financial time series comes from market-wide shocks, and those cannot be predicted from an asset's own history. Nothing in the past of one stock tells you what the whole market is about to do. So when self-supervised learning is run per asset, that variation has nowhere to go and is treated as irreducible noise. It is not noise, though. It is a real component of the series that the learner has simply been given no way to account for.

The authors work under a latent factor model and show what happens if you condition the prediction on the realized market factor instead of trying to forecast it. Two consequences follow. The asset's exposure to the market becomes necessary for optimal prediction, so the model has a reason to learn it rather than averaging it away. And the noise against which the asset-specific component is learned is reduced, because the market-wide part of the movement is now given rather than guessed. Market-Conditioned JEPA, or MC-JEPA, is the joint-embedding predictive architecture built on that insight.

During training, each asset gets a leave-one-out market condition. It is constructed from the future target embeddings of the other assets, which keeps the asset's own future out of its own condition. That condition then enters through factor loadings, and those loadings are derived from the asset's own representation. The market says what happened; the asset's representation says how much of it that asset should absorb.

The data is minute-level Chinese A-share, covering eleven years. Against per-asset JEPA, MC-JEPA improves across nine targets spanning returns, market exposure, risk and liquidity, with the strongest and most statistically significant gains in exposure, risk and liquidity. It achieves the best RankIC on all the liquidity targets, and on return horizons from two hours out to five days, which puts it ahead of time-series foundation models and of self-supervised baselines. In a historical backtest it delivers the best portfolio performance.

Three further results probe the mechanism. Ablations show the return and exposure gains require the condition to be aligned with the realized market movement. Once the volatility scale that normalization removed is restored, the predictor's factor loadings rank stocks by market beta nearly as well as a supervised probe does, with no labels involved. And a synthetic factor-model study with known parameters confirms that conditioning improves recovery of both the market exposure and the asset-specific component.

Treat the unpredictable part as noise and it contaminates everything learned around it. Hand it to the model as a condition and it stops being an obstacle and starts being the thing that makes an asset's own signal legible.