In briefMarket Impact Estimation via Causal Foundation Models: A causal simulator trained on synthetic markets estimates how large trades causally move prices.
SummaryThe paper presents a proof-of-concept Causal Market Simulator, a continuous-time causal PFN trained on a synthetic market where hidden informational flow confounds flow and price. On this prior, the naive estimator's bias follows a closed form (up to 68% overestimation), and the model tracks the unbiased oracle at high confounding. The authors report zero-shot responses to metaorders in an ABIDES order book and to flow bursts in crypto perpetuals and Nasdaq equities, consistent with but not proof of causal transfer. A recurrent state-space variant better captures impact accruing over execution duration.
In briefMulti-period mean-expectile portfolio optimization under Wasserstein ambiguity, with reformulation and ground-metric effects
SummaryThis paper studies multi-period portfolio optimization using expectiles as the risk measure under Wasserstein ambiguity sets. The authors derive an envelope theorem that characterizes the worst-case expectile as the unique root of a worst-case expectation, enabling standard Wasserstein duality and a reformulation into four parametric linear programs. Experiments on 90 FTSE constituents over 3,341 out-of-sample trading days show the expectile model outperforms a matched CVaR model in all nine parameter cells.
In briefWeighted selection from elliptical distributions: a stochastic representation for portfolio separation
SummaryThe paper represents weighted-selection elliptical distributions as an affine combination of q selection variables plus an elliptical term whose direction alone is independent. This representation suffices for q+2 fund separation via first-order stochastic dominance. It relaxes the three-fund assumptions of Simaan (1993).
In briefDiffusive Market Impact: A Consistent Microfoundation
SummaryThis paper provides a microfoundation for diffusive market impact, which explains the square-root law of market impact and its shift to a linear regime at low trading rates. In a model with one trader and many arbitrageurs who progressively eliminate predictable returns, impact returns become white. The authors also discuss the economic meaning of the impact phase and find that under a biexponential trade flow, impact decline falls within a 50%–60% band.
In briefForecasting Resale Prices of Used Phones and Laptops: A Benchmark of Statistical and Deep Learning Models
SummaryThis paper presents the first multi-horizon benchmark for predicting resale prices of used smartphones and laptops. Using daily listing data from Polish online marketplaces (January 2022 to March 2025), it compares eleven statistical and deep learning models across horizons from 1 to 365 days. N-BEATS achieves the lowest error beyond 30 days, with a MAPE of 8.51% at 365 days versus 14.94% for the best statistical baseline. At 1 to 7 days, all models perform similarly and a naive baseline remains competitive.
Mateusz Buczy\'nski, Micha{\l} Wo\'zniak, Konrad Kaczy\'nski et al.
In briefA Hawkes-based microfoundation for a multitype inverse-Gaussian random clock
SummaryThe paper derives a multitype inverse-Gaussian subordinator as the limit of event counts and integrated intensities of nearly critical multivariate linear Hawkes processes. The limit is a multivariate pure-jump subordinator whose dependence comes from microscopic cross-excitation. In the diagonal case it reduces to independent classical inverse-Gaussian subordinators.
WhyProves nearly critical multivariate Hawkes counts converge to a multitype inverse-Gaussian subordinator, with a cluster mechanism and extensions of prior scaling results.
In briefValidity Without Ground Truth: Applying Stated-Preference Economics to Evaluating Language Models
SummaryThis paper argues that the validity framework from stated-preference economics, covering content, construct, and criterion validity, reliability, incentive compatibility, and consequentiality, can evaluate language models when no correct answer exists. The authors test six models on a published water-quality stated-preference valuation survey (Vossler et al. 2023), using economic-theory predictions such as downward-sloping demand and willingness to pay responding to scope and income. Two older models fail the most basic test, while the two newest pass every scorable theoretical validity test but diverge on convergent validity. The authors stress that passing validity tests shows coherence, not correctness.
Daniel Robert Kling Alexander, Catherine Louise Kling
In briefLinking Stochastic Differential Equations to a Hierarchy of Generalized Beta Distributions
SummaryThe paper introduces a mean-reverting stochastic differential equation with a three-component noise term that generates a hierarchy of stationary distributions. The top level is a modified-Beta distribution, and its one- and two-parameter reductions yield compact-support, power-law, and exponential-type families. The authors build two generalizations, applying a power transformation either before or after obtaining the stationary hierarchy, and show these routes agree on some lower branches but generally differ at the top level. They derive density and distribution functions for both routes and compare them with the traditional Generalized Beta framework.
In briefSOTA: Option-Implied Return Distribution-Guided Stock Options Trading Agents
SummaryThe paper presents SOTA (Stock Options Trading Agents), an agentic framework that selects structured option strategies by abstracting the large contract universe into strategy-level decisions, with deterministic resolvers handling implementation. It post-trains Qwen3.8-27B using supervised fine-tuning followed by reinforcement learning. Evaluated on options for nine large-cap U.S. equities and SPY against rule-based and machine-learning selectors, SOTA reports an 18.3% six-month out-of-sample return, a Sharpe ratio of 1.60, and an 8.96% maximum drawdown. The authors also find news plays an asymmetric role: it improves frontier-teacher trajectories, but retaining news during reinforcement learning lowers out-of-sample return to -2.7%.
In briefOOM-RL II: Reality as Oracle, Not Debugger — Provenance-Constrained Diagnosis in Evolving Agent-Engineered Systems
SummaryThe paper examines a human-directed, agent-engineered quantitative trading system to show what external outcome feedback can and cannot establish. Over one year the account gained and beat a broad market index, yet annual alpha was not statistically distinguishable from zero under the main retrospective specification, and engineering records show changes during the period with incomplete recommendation-to-runtime binding. The authors introduce an outcome–diagnosis gap, distinguish frozen instances, pre-specified adaptive procedures, and ad-hoc development, and propose a prospective production-binding protocol that has not yet been prospectively validated.
