Finemma Français

All notes

Research digest, 13 November 2026

Research digest · Discuss on LinkedIn

This week's notes cover four preprints. Each asks what is left of an apparent signal once real-world conditions are added: adversarial inputs, transaction costs, portfolio constraints, or an independent replication.

Fake news and LLM forecasters (Zhang and Stadie, arXiv:2611.15102). The authors show that an attacker who only knows the forecasting question can steer retrieval-based LLM forecasters by publishing one invented article. Once the article is in context, 74% to 82% of opposing forecasts flip, across five frontier models, on stock questions and resolved prediction-market questions. Because the forecaster ends up right whenever the attacker is right, average accuracy barely changes, so the problem is invisible to the usual metrics. Two published defences can be bypassed with small changes to the attack. The experiments use archives built by the authors and small question sets, and code is promised on publication.

Seasonality as a factor in global futures (Butler et al., arXiv:2611.15657). Five calendar effects are modelled market by market with nonlinear regressions and tested walk-forward from 1995 to 2026, net of costs. The fast patterns, such as day of week, are statistically visible but turn negative after costs because of turnover. Slower patterns, such as day of year, remain positive and are largely uncorrelated with carry, momentum and skewness; in the authors' tests, adding them raises a factor portfolio's Sharpe ratio by about 30%. Readers should keep in mind that calendar effects are prone to data mining and that we did not see a formal correction for the number of features tested.

Return scenarios versus trading rules in ETF portfolios (Mensah et al., arXiv:2611.16540). Using 30 US ETFs over 946 trading days, the authors compare historical return inputs with conditional ARMA-GARCH and Student-t copula scenarios in matched mean-variance and CVaR portfolios. No difference is significant after multiple-testing control, and the conditional scenarios do not consistently reduce realised downside risk. Limits on how far target weights may move each day affect results more than the scenario settings. The window is short and costs are a simple flat rate, but the main point, that constraints and trading needs belong in the evaluation of any return model, is broadly useful.

Replication-based selection of feature interactions (Gao and Sun, arXiv:2611.13765). LENS fits the same model on two disjoint data halves and keeps an interaction only if the second model reproduces the pattern found in the first at a meaningful strength, with a single multiple-testing correction. It needs only two fits and works with any model that can be queried for predictions. On insurance data it recovers interactions documented by actuaries, and on two credit datasets it selects a short list without losing accuracy. For teams that need interpretable scorecards, it offers a documented cutoff for which interactions to include. The real-data evidence is limited to three public datasets, and code is available.

Papers covered

These notes summarise research for information only; they are not investment advice.