In briefFinNextAssist: A Deep Research Assistant for Professional Finance
SummaryThis paper presents FinNextAssist, an end-to-end deep research framework for professional financial analysis. It splits the research process into four stages: task planning, evidence compilation, reasoning, and report assembly, and adds two lightweight sub-agents, TabAgent for cross-market financial table understanding and HeteroAgent for interpreting heterogeneous financial data across modalities. The authors report experiments on three benchmarks showing it outperforms strong proprietary and open-source deep research agents, with ablations confirming each component's contribution.
In briefMintEval: Do LLMs Build the Trading Strategy Requested? A Behaviour-Matching Benchmark for Language-to-Strategy Code
SummaryThis paper introduces MintEval, a benchmark that checks whether code generated by large language models from plain-language trading descriptions behaves like the strategy requested. Generated and reference programs are run bar by bar on identical market data and frictions, and compared by their trading actions rather than code similarity or profit. The results show that even a frontier model silently fails on a substantial share of tasks, and that an existing LLM judge accepts all of these failures.
WhyProposes bar-level behavioural-equivalence evaluation of trading strategy code, with experiments exposing silent failures that existing judges accept.
In briefExplaining Neural Calibration of Volatility Models Using Shapley Values
SummaryThis paper uses two complementary Shapley-based explainable AI methods, SHAP and νSHAP, to analyze neural network calibration of the Heston and rough Heston stochastic volatility models. It compares several architectures and finds that short maturities and smile wings dominate parameter inference, with a stable attribution pattern across architectures. The authors show that νSHAP explanations can guide a large reduction in input dimensionality for the rough Heston model while keeping calibration accuracy comparable to using the full implied volatility surface.
Sha\"in Afzali, Serena Della Corte, Antonis Papapantoleon
In briefEvent History Over Scale: Compact Transformers for Low-Latency Order Book Forecasting
SummaryThe paper introduces MBOFormer, a 7,203-parameter causal transformer, and MBOFusion, a 14,371-parameter extension with a slow temporal-context branch, for short-horizon price-trend prediction from limit order books. Both models use market-by-order (level-3) histories of order submissions, cancellations, and executions, and are compared with level-2 snapshot-based baselines. The authors report the highest mean macro-F1 in eleven of twelve settings across three markets and four horizons, with sub-millisecond median inference and state dictionaries under 70 kB.
David Schaurecker, Lasse B. Strand, Kevin O'Sullivan et al.
In briefMulti-agent AI as a nested principal-agent problem in private wealth management across Switzerland, Germany and Austria
SummaryThe paper proposes a model-independent framework that combines nested principal-agent delegation, in which a manager hands work to AI acting for the client, with constrained joint maximisation of client and manager outcomes. Legal duties, mandate requirements and evidence sufficiency determine admissibility, with Switzerland, Germany and Austria as the legal context. Simulations using constructed mandates and public-market data show how weights and omitted terms can change selected services. These are illustrative examples, and professional and field studies are needed to assess oversight and client outcomes.
Walter Kurz, Reinhard Magg, Florian Kollberg et al.
In briefAxient: A Manifest-Bound Evidence Framework for On-Chain Financial Protocols
SummaryThis paper presents an evidence architecture for hybrid on-chain financial protocols, in which a sealed execution manifest binds system identities, roles, schemas and release identity. Seven evidence layers are linked by a common correlation identity and must satisfy a conjunctive rule before a claim is promoted. The authors formalize the evidence graph, prove that copied hashes cannot establish independent derivation, and report a registered local case study showing a reviewable chain from financial assertions to a replay-bound package.
In briefAxient: A canonical protocol graph for leveraged event markets with single-authority state and recoverable execution
SummaryThis paper addresses leveraged event-market protocols, where individually correct contracts may still lack a single authoritative financial execution path. It proposes a canonical protocol graph in which each financial domain has one storage authority and composition components coordinate transitions. The authors formally prove properties such as atomic closure, loss conservation and priority, and exactly-once recovery, and evaluate the implementation through twelve Financial Interaction Assertions and a reproducible evidence chain.
In briefSame Ranking, Different Scale: Checking LLM Judges for Job-Related AI Measurement
SummaryThis paper introduces O*NET-BENCH, an audit suite built from an existing survey of 45,796 worker ratings, to test 33 LLM judge configurations on whether AI responses meet workplace requirements. Many judges agree with human raters on response rankings, yet their estimated acceptance rates range from 3.0% to 97.9%, compared with 61.1% for occupation-matched workers. The authors conclude that ranking agreement alone is insufficient for occupational measurement and that judges should be validated against the acceptance rates and aggregates they will be used to estimate.
In briefExpected Utility Regret Rule: Minimax and Bayes-Optimal Portfolio Selection
SummaryThis paper studies portfolio choice, aiming to recommend portfolios that maximize an investor's expected utility of wealth. It proposes the Expected Utility Regret (EUR) rule, which jointly selects a portfolio class and estimates its weights. In a regular parametric return model, a single EUR rule attains both minimax and Bayes lower bounds without using the prior distribution. Mean-variance and risk-parity portfolios are shown to be special cases of this framework.
WhyProposes an EUR rule attaining both minimax and Bayes regret lower bounds with leading constants without a prior, a notable theoretical advance.
In briefShared Models, Selective Trading, and Order Flow: A study of whether diverse language models still differ in the orders they submit when news framing changes in synthetic markets.
SummaryThe paper uses synthetic markets with a fixed mix of three language-model families to examine how news presentation affects the orders they submit. It reports that Qwen's share of submitted orders shifted by 48 percentage points in a financing event with little change in net order counts, and Mistral's share shifted by 40 points in a workforce-reduction event where net counts reversed sign. An analytical decomposition shows selection can improve or worsen price accuracy even when aggregate demand sensitivity is unchanged.
