In briefModelling Regime Shifts in Electricity Intraday Markets with State-Dependent Hawkes Processes
SummaryThis paper proposes a multivariate state-dependent Hawkes process to model order flow across liquidity regimes in the continuous intraday electricity market. Market states are identified with a Liquidity Stress Index built from bid-ask spread, order book volume, and mid-price volatility. Calibrated on Dutch XBID hourly order data from EPEX SPOT for January-March 2024, the estimates show strongly self-exciting, near-critical order flow, with liquidity stress mainly strengthening same-side transaction self-excitation.
Ayoub Jhabli, Tarek AlSkaif, Kwabena E. Bennin et al.
In briefFinVector-Market-4B: a controlled test of LoRA tuning for structured finance tasks
SummaryThis paper fine-tunes the Qwen3.5-4B model with rank-16 LoRA on 22,000 examples for structured financial tasks, then compares it with the base model on a 600-example benchmark. Supplying a JSON schema alone raised the base model's JSON validity from 0% to 91.3%. Under matched explicit prompting, several task scores also improved. The authors note that the reported macro-F1 drop reflects a changed label set, and that filing overlap and other issues limit the generalization claims.
In briefFinite bid-ask spread from replenishment placed away from the quote
SummaryThis paper gives an analytic account of a finite bid-ask spread in a two-field reaction-diffusion order book model that tracks bid and ask densities over operational time. A market-making source withholds order placement over a finite interval around the reaction boundary, representing replenishment displaced from the quote. Under a quasistationary, weak-overlap approximation, the quoted spread is the sub-threshold core of the placement-free interval. Market orders remove density at the quote, while replenishment arrives outside it and must diffuse back, producing a delayed and attenuated quote-level response.
In briefA functional representation of credit behavior for default probability modeling
SummaryThis paper proposes a framework that models probability of default using functional data analysis, representing credit variables as continuous functions. It shows that commonly used variables such as available funds, utilization rate, and overdraft can be derived from three time-observed quantities: account type, account balance, and credit limit. A relative time framework aligns recurring banking-day cycles across borrowers, and the approach is evaluated with functional logistic regression. The model reportedly matches XGBoost's predictive performance while outperforming standard logistic regression and retaining the interpretability of linear models.
Jonas Brunholm, Bjarne H{\o}jgaard, Thomas D. Nielsen et al.
In briefLearned Monotone Recurrent Features in Regulated Credit Scoring
SummaryThis paper examines whether learned monotone recurrent features add value to regulated credit scoring, testing five production-scale credit datasets. The authors extend the monotonicity guarantee with proofs to vector inputs and macro-conditioned decay gates. They find that learned feature value rises with governance strictness, and that conditioning the recurrence on macroeconomic series, rather than using them as inputs, yields crisis-cohort AUC gains.
In briefConvex Order in Higher Dimensions: Projection Tests, Counterexamples, and Gaussian Mixtures
SummaryThis paper studies convex order, a way to compare random vectors with the same mean, in dimensions greater than one. It shows that checking all one-dimensional projections is not sufficient in general and builds radial counterexamples in every dimension at least two. The authors prove that projection tests do suffice for linear transformations of a common radially distributed vector, including centred Gaussian distributions. They also settle when a centred Gaussian is dominated by a mixture of two centred Gaussians, showing this holds exactly when all one-dimensional projections are ordered.
WhyConstructs radial counterexamples in every dimension and fully settles convex order for Gaussian versus two-component mixtures, with rigorous proofs.
In briefAdversarial training for deep hedging when markets shift over time
SummaryThis paper proposes WRAP, a drift-aware adversarial training framework for deep hedging based on a two-budget distributionally robust optimization formulation. It lets an adversary reweight observed trading trajectories under a φ-divergence constraint and perturb their paths under an optimal-transport constraint, to handle mismatch between historical training data and future market conditions. Experiments on stationary and nonstationary Heston dynamics and a generalized affine diffusion (GAD) show that reweighting and transport each help, with joint adversarial training giving the largest gains under nonstationarity.
Philipp J. Schneider, Lukas Looser, Antoine Garin et al.
In briefSynthetic cultural agents built from group-level preference anchors
SummaryThe paper proposes building synthetic survey respondents by mapping the signs of six Global Preferences Survey coordinates to group-specific choice policies, each fitted as a parameter-efficient adapter with Direct Preference Optimization. On candidate World Values Survey items, the adapters recover the imposed labels and separate the two GPS-sign groups on a sixteen-country development panel. However, agreement between the adapters and human WVS responses remains unresolved.
Augusto Gonzalez-Bonorino (1, 2), Kseniia Biriukova (2 et al.
In briefError bounds for option pricing under volatility uncertainty
SummaryThe paper develops a posteriori primal-dual bounds for numerical approximations of European option prices in the Uncertain Volatility Model. A lower bound comes from a feedback control induced by the Hessian of a smooth candidate value function, while an upper bound comes from a dual representation based on a matrix-valued Gamma field. The authors also quantify discretization errors and test candidates from policy-gradient and physics-informed neural networks. The experiments indicate that second-order derivative accuracy is important for tight bounds.
Lokman A Abbas-Turki (LPSM (UMR\_8001)), Chassagneux Jean-Fran{\c c}ois (ENSAE Paris), Jean-Philippe Lemor et al.
