MC-JEPA Conditions Financial Time-Series Learning on Market Shocks
Summary:MC-JEPA reformulates financial time-series learning by explicitly conditioning each asset's prediction on a market-wide state. During training, a leave-one-out market condition is constructed from future target embeddings of other assets, while factor loadings are inferred from the asset's own representation. Across eleven years of minute-level Chinese A-share data, the approach improves nine targets spanning returns, market exposure, risk, and liquidity over per-asset JEPA. It reports the stron
Core Idea: Separating Market Shocks from Asset-Specific Noise
MC-JEPA addresses a structural limitation of per-asset self-supervised learning. Much of the variation in an individual financial time series comes from market-wide shocks that cannot be inferred from that asset's own history. Conventional models often treat this variation as irreducible noise, obscuring both stable market exposure and asset-specific predictive structure.
Under a latent factor model, the paper argues that conditioning prediction on the realized market factor makes an asset's factor loading necessary for optimal prediction. At the same time, it reduces the variance that the asset-specific component must learn. This is therefore a structural change to the learning problem rather than the addition of a conventional index feature.
Training Mechanism: Leave-One-Out Market Conditions
For a target asset (i), MC-JEPA constructs a market condition from the future target embeddings of other assets while explicitly excluding those of asset (i). The leave-one-out design prevents the condition from directly containing the answer being predicted.
Factor loadings are inferred from the asset's own representation and used to inject the shared market state into its prediction. Conceptually:
[ \hat{y}i = f\theta(x_i) + \beta_\theta(x_i) \cdot c_{-i}, ]
where (f_\theta(x_i)) captures the asset-specific component, (c_{-i}) is the leave-one-out market condition, and (\beta_\theta(x_i)) represents the asset's exposure to the shared factor. Ablation results show that the gains in returns and exposure require the condition to align with realized market movement; an arbitrary or misaligned condition does not reproduce the same improvements.
Why the JEPA Objective Matters
Instead of directly generating future prices, the model predicts the latent embedding of a future target and minimizes the discrepancy between predicted and target embeddings. Relative to pointwise price reconstruction, this objective can filter out less-predictable noise and retain structures that transfer across time.
MC-JEPA adds explicit cross-sectional conditioning to this joint-embedding objective. The synthetic factor experiment, where the true parameters are known, shows that conditioning improves recovery of both market exposure and the asset-specific component. This supports the interpretation that the gains arise from modeling the intended statistical structure rather than from stock-specific data artifacts.
Experimental Results and Evaluation Boundaries
The study uses eleven years of minute-level Chinese A-share data and evaluates nine targets spanning returns, market exposure, risk, and liquidity. MC-JEPA improves over per-asset JEPA, with the strongest and most statistically significant gains appearing in exposure, risk, and liquidity. The paper reports the best RankIC on every liquidity target and on return horizons from two hours to five days, outperforming the evaluated time-series foundation models and self-supervised baselines. It also reports the strongest portfolio performance in a historical backtest.
These findings support the claim that market conditioning improves cross-sectional ranking. They do not, by themselves, establish deployable alpha. RankIC measures rank correlation between predictions and future outcomes; it is not the same as return after transaction costs. Portfolio results also depend on position formation, turnover, market impact, price limits, suspensions, and out-of-sample tuning. Those details are not available in the supplied abstract and therefore require verification from the full paper and implementation.
The study also finds that normalization removes the original volatility scale. After restoring that scale, the model's factor loadings rank stocks by market beta almost as well as a supervised probe, without using labels. This suggests that the representation captures meaningful market-exposure structure rather than merely exploiting externally supplied beta targets.
Engineering Trade-offs
1. Leakage and point-in-time correctness
Training uses future target embeddings from other assets to construct each condition. Offline, this can be understood as a reorganization of historical cross-sectional targets. During inference, however, the visible asset set and information timestamp must be strictly enforced. Any look-ahead window, synchronized leakage, or post-decision information could materially inflate backtest performance.
2. Cross-sectional coupling
Unlike per-asset JEPA, MC-JEPA must maintain target embeddings and market conditions across the asset universe. This adds cache, batching, and data-pipeline complexity, while also creating operational issues around suspensions, listings, delistings, and uneven asset histories.
3. Market-specific generalization
Results on Chinese equities cannot automatically be transferred to U.S. equities, futures, or digital assets. The learned condition may depend on A-share market size, institutional structure, trading rules, industry composition, and factor behavior. Cross-market evaluation should retrain or recalibrate the system rather than reuse loadings unchanged.
4. Normalization and tail risk
A model trained in a normalized latent space may rank assets well while producing poorly calibrated risk magnitudes. Production use therefore requires scale restoration, volatility calibration, and stress testing under extreme market regimes. Better RankIC can still be overwhelmed by losses concentrated in rare events.
Deployment Guidance
- Build a strict point-in-time pipeline covering raw quotes, suspension states, delisted securities, and minute-level information timestamps.
- Enforce a genuine leave-one-out condition for every target asset; never allow its future target embedding into the condition.
- Use walk-forward validation with separate tuning, validation, and final test windows.
- Report ICIR, turnover, drawdown, quantile portfolio returns, and transaction-cost-adjusted returns in addition to RankIC.
- Incorporate A-share execution constraints, including price limits, suspensions, T+1 settlement, minimum lot size, and market impact.
- Orthogonalize or attribute market conditions against macro and style factors to determine whether gains merely repackage known risk exposures.
- Evaluate bull, bear, range-bound, and extreme-volatility periods separately rather than relying only on an aggregate result.
Conclusion
MC-JEPA's main contribution is not a more elaborate financial Transformer. It changes the conditioning structure of self-supervised time-series learning: shared market shocks should not automatically be discarded as noise. Leave-one-out cross-sectional conditions and asset-specific factor loadings provide a more consistent way to model returns, exposure, risk, and liquidity.
Deployability still depends on strict point-in-time construction, transaction-cost-aware backtesting, and independent validation across market regimes. If the full results and implementation replicate, the broader lesson is useful beyond finance: before modeling an entity in isolation, identify which shared states can be inferred from other entities, then focus the entity-specific model on the variation that remains.
— END —Source: Hugging Face Daily Papers (2026-11-10)
Tags: #MC-JEPA #Financial Time Series #JEPA #Cross-Sectional Factor Modeling #A-Share Quant Research
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments