In briefCertifying simplified and neural models for option pricing under the tempered-fractional CGMY model
SummaryThe paper checks how accurate simplified and neural approximations are for European put prices under the CGMY option pricing model, using residual error bounds from a spline Galerkin discretisation. The system matrices are built from an explicit periodised Fourier symbol, and a Lévy–Khintchine argument gives a parameter-explicit eigenvalue bound. The authors report that, at full resolution (N=1024), the certified reduced model costs about 3% of a direct solve and settles 99.8% of 10^4 queries without the full space.</summary>
In briefPenalize the Common Trade: A Collision-Aware Layer for Learned Portfolios
SummaryThe paper argues that learning systems sharing data, objectives, or representations can make nearly identical trades, creating a "signal collision." A linear-impact game shows allocators take on too much common exposure, and an overlap-dependent toll can recover the social optimum. The authors propose a simple post-model layer that penalizes loading on the recent common return mode, which lowered maximum drawdown in a panel of ten liquid cross-asset ETFs, though the drawdown-improvement confidence interval includes zero.
Eric Benhamou, Jean-Jacques Ohana, Beatrice Guez et al.
In briefSpecification LASSO and a Flexible Characteristics-Based Asset Pricing Model: a statistical methods paper on variable selection and model specification
SummaryThis paper studies a partially linear semiparametric additive model that allows interactions among variables in high-dimensional settings. The authors propose the Specification LASSO (S-LASSO), a two-step method combining LASSO and Adaptive Group LASSO to perform variable selection and model specification simultaneously. The paper establishes oracle properties, including selection consistency and asymptotic normality, and Monte Carlo simulations show it outperforms existing alternatives. Empirically, the method is applied to a characteristics-based asset pricing model to examine linear and nonlinear effects of asset characteristics on cross-sectional returns and their interactions with firm size.
In briefCan We Stop the Backtest? Anytime-Valid Promotion Tests for Financial Research Agents
SummaryThe paper introduces PROMOTE, a protocol that gives each candidate trading rule predictable testing wealth at birth and accumulates a cost-aware e-process from subsequent returns only. If lifetime allocations sum to at most one, a weighted Ville argument bounds the probability of promoting a rule whose bounded net score has always had nonpositive conditional mean. The method allows dependent candidates, adaptive births, and indefinite monitoring. The authors evaluate the implementation in a retrospective replay and a controlled experiment, comparing it with repeated unadjusted tests. The certificate controls historical false promotion under the declared null and does not show that an earlier edge remains profitable at deployment.
Eric Benhamou, Jean-Jacques Ohana, Beatrice Guez et al.
In briefEventGPT: a model pretrained to predict banking customers' next events
SummaryThe paper introduces EventGPT, a hierarchical causal transformer pretrained to predict a customer's next event, which lets it estimate future event probabilities without labels and provide reusable representations for downstream banking tasks. Trained on the public MBD-mini dataset, it scores five next-month first-transaction targets at ROC-AUC 0.79 to 0.92. For product prediction, its frozen embeddings score slightly below engineered features, but combining them with those features improves results. The code and pretrained weights are released.
