Finemma English

Toutes les notes

Research digest, 16 November 2026

Revue de recherche · En discuter sur LinkedIn

Cette note n'existe qu'en anglais.

This week's notes cover four preprints about evaluating AI and decision systems in finance beyond a single headline number. They look at compliance controls for LLM trading agents, run-to-run noise in a financial research agent, forecasts designed around a battery trading decision, and a credit collections benchmark with a known optimum.

Compliance controls for trading agents (Nandi et al., arXiv:2611.18493). TradeGauntlet is a replayable order-book simulator that measures separately what pre-trade gates detect, what they block and what they cost in legitimate trading. Across 1,680 sessions with two LLMs, every alert was a self-match, which simple gates removed; one gate setting also reduced the larger model's two-sided fills by about 15%. A targeted search found split-order schedules that evade the rules, and a learned detector that caught them all also flagged 15% of a legitimate market maker's sessions. The results are simulation-based and depend on a price response fitted to five Nasdaq stocks, but the code is public and the framing, measuring both protection and cost, is directly useful.

How stable is a finance agent's score? (Galeone et al., arXiv:2611.17939). The authors rerun the reference agent of a public finance-agent benchmark seven times on the same 18 questions and see its score range from 105 to 123 of 149 rubric criteria, with cost varying by up to 17%. Most failures trace back to the agent's instructions rather than to the model or data. An automated tuning loop delivered a real cost saving (0.34 times baseline on held-out questions), but its 14-point accuracy gain on development questions shrank to 1.5 points on held-out ones, within noise. It is a small case study, one agent and nine test questions, but the method of measuring noise before claiming an improvement applies widely.

Forecasts shaped by the trading decision (Hirsch et al., arXiv:2611.17170). For battery storage trading on the continuous German intraday power market, the authors forecast price quantiles derived from the order book and feed them into a standard rolling trading policy as resting orders. In a simulation on 2023 and 2024 EPEX SPOT data with fees and degradation costs, profits improve by up to 27% for 2-hour and 42% for 4-hour batteries, while 1-hour batteries do worse. More accurate forecasts also tended to change more between updates, a trade-off that matters wherever forecast changes trigger trades. The study covers one market and relies on replayed order books and proprietary forecast data.

Testing what limits a credit decision policy (Silva Prado et al., arXiv:2611.18367). DECODE provides synthetic consumer-credit collections environments with known dynamics, so a solver's score can be compared with the exact optimum and its inputs changed one at a time. Longer planning adds 15 to 17 points of the optimum to two gradient-boosted solvers, and the best remains 15.6 points short; swapping in the true repayment model recovers much of that gap. The authors are careful to note that the effects depend on the order of changes, and the environments are synthetic, but code is released and the approach suits teams who need to decide where to invest in a decision system.

Articles résumés

Ces notes résument des travaux de recherche à titre d'information ; ce ne sont pas des conseils en investissement.