In briefMeasuring Trade Direction on a Prediction Market: Settlement-Based Validation and Trading-Cost Measurement on Polymarket
SummaryThis paper validates the trade-direction labels of Polymarket's public trade prints against the taker leg of on-chain settlement records. It compares Lee-Ready, tick-test and retrospective bulk volume classification, and shows how sign errors shift measured effective spread, realised spread and midpoint impact. The authors caution that these delivered-quote accounting measures do not identify causal impact or private information.
In briefExpert-verified AI study materials and who gains from them in a university course
SummaryThis paper studies a compulsory first-year economics course in which one cohort received AI-generated podcasts, FAQs and quiz-based study guides checked by a graduate teaching assistant. A difference-in-differences analysis of 340 examination marks links access to a 2.34-mark average advantage on a 50-mark component. Most of the average effect originates in the bottom quintile, and the authors argue that evaluations reporting only mean effects cannot show whether the intended students benefit.
In briefComparing run-off triangles and regression for estimating credit card loss given default
SummaryThe paper benchmarks run-off triangles, a common banking method, against a two-stage regression model for estimating credit card loss given default (LGD). The regression approach can reproduce the typical U-shaped LGD distribution, which run-off triangles cannot. Aggregated regression estimates also track empirical loss rates over time much more closely. The authors suggest the regression approach is likely better under IFRS 9, which prioritises accuracy.
In briefWho Keeps the Gains from AI Assistants? Seller Responses and Non-Users in a Simulated Rental Market
SummaryThis paper uses an agent-based rental market simulation with language models to study how sellers' responses shape the gains from personal AI assistants and their effects on consumers without one. With frozen sellers, executing assistants cut adopters' spending by 13.7 USD per renter-day, and adaptation recovered about a third. The predicted burden on non-users did not materialise, with a mean spending change near zero. The authors argue assistants should be evaluated at the market level.
WhyAgent-based language-model market simulation shows seller adaptation claws back about a third of assistant gains, and the unassisted burden forecast fails.
In briefA Dirichlet mixed-membership model for exact multivariate credibility in insurance pricing
SummaryThis paper proposes a Dirichlet mixed-membership model (DMMM) that represents policyholder risk as a stable composition over latent risk classes. Baseline characteristics set the a priori assessment, and repeated multivariate experience updates the a posteriori assessment. The posterior predictive distribution is written exactly as a convex combination of portfolio and experience components, extending classical credibility beyond the conditional mean. The framework is studied through a simulation and a vehicle telematics application.
In briefContinuous Latent Stress Spectrum in Limit Order Book Dynamics
SummaryThe paper analyses high-frequency limit order book data for EURO STOXX 50 index futures over 987 trading days from 2022 to 2025, testing whether market states form distinct regimes or a continuum discretised into apparent regimes. It finds an effective state dimension of about 1.4, with one dominant latent coordinate explaining 83.66% to 85.27% of covariance-state variation and tracking VSTOXX without regime labels. Fine-scale continuous information remains predictive, with forecasting peaking near ten minutes and a maximum pooled out-of-sample R2 of 0.7483.
WhyRiemannian geometry reframes limit-order-book states as a continuum rather than discrete regimes; 987 trading days and a 17-method comparison give fairly solid evidence, useful within the subfield.
