In briefComparing learned and hand-crafted monotone temporal risk aggregation under equal admissibility
SummaryThe paper studies risk scores that must be monotone non-decreasing in every exposure input, a constraint drawn from financial regulation. The authors build a recurrent network whose state uses classical risk statistics, keeping it monotone, and compare it with hand-crafted aggregates under the same constraint. They find that learned and engineered aggregation are substitutes within the separable family, but the learned approach fails on consecutivity, time localization, and the exposure floor. Each incumbent approach has advantages on different axes, and none dominates on all three.
In briefStock-JEPA: Learning Equity Market Representations by Revising a Financial Prior
SummaryThis paper proposes Stock-JEPA, a framework for learning representations from low signal-to-noise equity market data. It anchors a latent space to fixed statistics from a simple financial model describing multi-horizon returns and risk, then learns a context-based predictor of the revision needed to reach future representations. The authors report that it outperforms 13 baselines on China and U.S. equity universes across five evaluation metrics.
In briefBuying at support: optimal entry under multi-skew Brownian price motion
SummaryThe paper studies when to buy a share that will later be sold optimally, when the price follows a geometric multi-skew Brownian motion whose skew levels represent support and resistance. The buying reward is the exit premium from a companion liquidation problem. The entry problem reduces to a finite one, and with one support and one resistance, the optimal rule is to buy exactly at the support, with an explicit value.
In briefAlphaPADI: Automatically discovering stock prediction formulas by generating formula pools with diffusion
SummaryThis paper studies formulaic alpha discovery, which seeks symbolic formulas that predict cross-sectional stock returns. The authors propose AlphaPADI, a framework with grammar-constrained buffer initialization, pool-aware hierarchical discrete diffusion, and reward-guided pool refinement that considers both predictive performance and diversity among formulas. According to the paper, experiments on Chinese and U.S. stock markets show AlphaPADI outperforms the evaluated baselines in predictive and portfolio performance.
