Finemma English

Toutes les notes

Research digest, 12 November 2026

Revue de recherche

Cette note n'existe qu'en anglais.

This week's notes cover four preprints. Three of them are about the same question: does a model's apparent skill survive an honest test? The fourth looks at how the training environment shapes a hedging policy.

Time-series foundation models on index futures (Upadhyay, arXiv:2611.12141). Five popular pretrained forecasting models, used without fine-tuning, are tested on E-mini Nasdaq-100 and S&P 500 futures at horizons from 5 to 120 minutes, under a pre-registered protocol with a blind test set. None of 160 combinations beats a forecast of no change, and a formal test for superior predictive ability finds nothing. The paper also explains why some common metrics, such as price-level correlation and within-day rank correlation, still make these models look useful. The study is limited to two futures contracts and to frozen models.

What a quantum layer adds to equity ranking (Chen et al., arXiv:2611.10801). The authors run a careful ablation of a hybrid quantum-classical model ranking 2,000 Tokyo-listed stocks: the quantum circuit trained, frozen, removed, or replaced by a classical surrogate, with walk-forward folds, several seeds, deflated Sharpe ratios and a locked holdout. Training the circuit makes no measurable difference, and no variant earns a Sharpe ratio distinguishable from zero; US stocks show the same pattern. The protocol, including equivalence tests and a stated minimum detectable effect, is reusable for testing any new model component. Statistical power is limited by the short test window.

LLM agents and lagging web news (Feng et al., arXiv:2611.10414). Web news about listed firms often reports a price move that has already happened. Using a point-in-time multi-agent system, the authors find that LLM judgments of news correlate with returns around the event, but that the association largely vanishes when returns are measured from the first trade possible after publication (a weekly rank correlation of 0.119 falls to 0.023). They recommend that text signals be scored from publication time. The system relies on Jev, a hosted decision model whose availability the paper does not discuss, and some analyses are exploratory.

Reinforcement learning for options hedging on real data (Chen et al., arXiv:2611.10956). A risk-sensitive RL agent hedges an options market maker's book, trained on S&P 500 options from 1996 to 2017 and tested on 2018 to 2023, against delta hedging and the same agent trained in simulated markets. Training on real data lowers tail losses by about 3% overall and about 9% in falling, volatile markets, but costs more in rising markets. It also learns a short-delta tilt that disappears or weakens when the agent is trained in simulated markets. The authors note that the result rests on a single historical test path.

Articles résumés

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