Researchers train AURA models to forecast stock returns in chronological sandbox
A new arXiv paper reports that fine-tuning plus reinforcement learning lifted two open models' forecasting scores, with the larger model beating evaluated frontier models on the authors' benchmark.
In plain words
Researchers built BETA, a test environment where a language model must use tools to look up prices, indicators, and news available at a past date, then predict how a stock's return would move. They trained their models in two stages, including reinforcement learning, where the model learns from whether its predictions matched real outcomes. The trained models, called AURA, scored much higher and fixed three weaknesses seen in frontier models.
What was announced
The authors, writing in an arXiv abstract, report that frontier models tested in BETA showed three weaknesses: their scores declined as the test period moved closer to the present while a simple momentum baseline stayed stable, they predicted gains far more often than gains actually occurred, and they predicted moves much smaller than those realized. To address this, the team first fine-tuned models on filtered, hindsight-generated tool-use trajectories for historical tasks judged predictable from evidence available at the time, then applied reinforcement learning with rewards based on realized returns.
Training raised the 8B model's score from 24.0 to 47.0 and the 30B model's from 27.4 to 54.7, with the 30B model surpassing all evaluated frontier models on the authors' benchmark and achieving its largest margin on the most recent test period. The reinforcement learning stage required two technical fixes described in the abstract: a selective compaction tool that lets the model compress portions of its interaction history when long investigations exceed the context window, and Context-Aligned Turn-Level PPO, which uses a turn-level critic and computes policy updates using each turn's actual context. The authors suggest these results indicate agentic reinforcement learning can improve performance on tasks with noisy, delayed real-world outcomes.
What it does not show
- The source is only the paper's abstract, so method details, experimental setup, and full results cannot be verified from the supplied evidence.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.