Market-Conditioned JEPA learns asset signals from realized market shocks
Original title: Market-Conditioned JEPA: Learning from Unpredictable Market Shocks
On 11 years of minute-level Chinese A-share data, it beat per-asset JEPA on returns, exposure, risk and liquidity.
The paper introduces Market-Conditioned JEPA, which conditions each asset's prediction on a leave-one-out market factor built from other assets' future embeddings. On Chinese A-share minute data it reports gains over per-asset JEPA across nine targets, strongest for exposure, risk and liquidity, and the best RankIC on liquidity and on return horizons from two hours to five days. Ablations say return and exposure gains require the condition to match realized market movement.
Unconfirmed: Performance gains are reported by the paper and are not independently confirmed here.
- Factor loadings come from each asset's own representation.
- After restoring volatility scale, loadings rank market beta nearly as well as a supervised probe.
- A synthetic factor-model study reports better recovery of exposure and the asset-specific component.
- The paper also reports the best portfolio result in a historical backtest.
From the original article
1 min read at huggingface.co
Much of the variation in financial time series is driven by market-wide shocks that cannot be predicted from an asset's own history. Per-asset self-supervised learning treats this variation as irreducible noise.
By Yuqi Wang, Fengyuan Liu, Mukai Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.