Finemma Français

All notes

Research digest, 11 November 2026

Research digest

This week's notes cover four preprints. Two are about evaluation hygiene, and one is about uncertainty in risk forecasts.

Point-in-time encoders for earnings calls (Brini, arXiv:2611.14325). Embedding models used in backtests are usually trained on today's text, which leaks future information into past signals. The author trains a family of small encoders on about 412,000 earnings calls, each limited to text up to a yearly cutoff from 2010 to 2024. These date-safe encoders retrieve answers from held-out calls better than the general retrievers tested. They also improve a simple return-ranking baseline, but the paper finds that the point-in-time discipline matters much more for retrieval than for return forecasts. The code is public.

FinCutBench (Li et al., arXiv:2611.11588). An LLM can know about events after a backtest date even if its inputs are filtered by date. FinCutBench uses 9,000 dated multiple-choice questions on US and Chinese markets, updated every month, to estimate where a model's financial knowledge drops off. This gives teams a measured boundary to use when choosing evaluation windows. Multiple-choice recall is only a proxy, and the downstream effects differ by market and task.

Fine-tuning an order-flow foundation model (Shabashev, arXiv:2611.14117). A 300M-parameter generative model is pretrained on order-by-order data from two crypto exchanges, then fine-tuned to forecast sub-second microprice returns and volatility. It is reported to beat gradient boosting on hand-built features and to transfer to unseen assets. The author states that the predicted moves are smaller than spread plus fees, which points to execution and market-making uses. The study covers crypto venues only, and the abstract describes no cost-aware simulation.

Prediction intervals for hybrid volatility models (Parrella et al., arXiv:2611.13212). The paper adds an ARMA mean to GARCH-LSTM models and builds multi-step prediction intervals for realized volatility with a forward bootstrap. Two remedies are tested for intervals that are too narrow: fat-tailed residual resampling and conformal-style widening. Combining both brings coverage closest to nominal on simulations and on four US indices from 2020 to 2025. The sample is short and the benchmark set is narrow.

Papers covered

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