---
title: "When Valid Signals Fail — paper by Yang (2026)"
description: "Claude proposes structured edits, and each edit is scored by the Information Coefficient (IC), not by classification accuracy. 12 references."
authors: ["Zhengzhe Yang"]
year: 2026
venue: "Proceedings of the Second Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U)"
source: "https://aclanthology.org/2026.customnlp4u-1.17/"
doi: "10.18653/v1/2026.customnlp4u-1.17"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/yang-2026-when-valid-signals-fail-regime/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/yang-2026-when-valid-signals-fail-regime/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/56/) / When Valid Signals Fail: Regime Boundaries Between LLM Features and RL Trading Policies

# When Valid Signals Fail: Regime Boundaries Between LLM Features and RL Trading Policies

**Authors:** Zhengzhe Yang · **Year:** 2026 · **Venue:** [Proceedings of the Second Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U)](https://md.research.elamur.ai/free/en/venues/customnlp4u/) ([Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual 2026](https://md.research.elamur.ai/free/en/venues/customnlp4u/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.customnlp4u-1.17/) · [PDF](https://aclanthology.org/2026.customnlp4u-1.17.pdf) · [DOI 10.18653/v1/2026.customnlp4u-1.17](https://doi.org/10.18653/v1/2026.customnlp4u-1.17) · [OpenAlex](https://openalex.org/W7166819841)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper asks whether a frozen LLM can produce numerical features that make a reinforcement-learning trading agent better. The author built a pipeline in Go that gathers daily news from Benzinga and RSS feeds, plus SEC EDGAR filings, for each ticker. It stores these in a SQLite snapshot to avoid look-ahead bias. Qwen3 235B A22B Instruct 2507 turns each daily bundle into a small JSON vector covering sentiment, impact, conflicting signals, novelty and regime flags. A FinRL PPO agent consumes this vector together with price data and FRED macro indicators.

The extraction prompt is treated as a tunable hyperparameter. Claude proposes structured edits, and each edit is scored by the Information Coefficient (IC), not by classification accuracy. The tuned prompt yields features that genuinely predict returns. However, during a tariff-driven shock in H1 2025 these features hurt the agent relative to a price-only baseline. In the calmer H2 2025 period they help modestly, and macro variables remain the most dependable source of gains.

The lesson is that features that look valid on intermediate metrics can still fail under distribution shift. LLM-fed pipelines therefore need out-of-sample testing across multiple market regimes.

## Key points
- **Prompt optimization:** five mutation types were tested on January 2025 data (769 bundles, 38 tickers). The winning prompt moved IC from −0.024 to +0.104, and the abstract reports held-out IC above 0.15.
- **Agent setup:** The agent is a PPO model using a multilayer-perceptron policy, which learns from 2023–2024 data over 500k timesteps. It trades 21 tickers, and results are averaged over 5 seeds.
- **H1 2025 (volatile period):** LLM+Macro reached a Sharpe of −0.267 versus 0.010 for the baseline, and no configuration beat the baseline. This gap held across transaction costs from 0 to 50 bp.
- **H2 2025 (calmer period):** Sharpe ratios were 1.038 for LLM+Macro, 1.099 for Macro-only and 0.809 for the baseline. The LLM+Macro gain was not significant (p = 0.487), and SPY buy-and-hold beat every configuration with a Sharpe of 1.756.
- **Role of macro features:** in a gradient-boosted tree used as a proxy model, macro inputs received 58% of split importance. Separately, the predictive power of news features peaks at a 3–10 day horizon.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

📚 Shelves (6): [🎯 Decision Sciences](https://md.research.elamur.ai/free/en/shelves/decision-sciences/) · 249 works · [🎯 Management Science and Operations Research](https://md.research.elamur.ai/free/en/shelves/decision-sciences/management-science-and-operations-research/) · 134 works · [📈 Economics, Econometrics and Finance](https://md.research.elamur.ai/free/en/shelves/economics-econometrics-and-finance/) · 23 works · [💼 Business, Management and Accounting](https://md.research.elamur.ai/free/en/shelves/business-management-and-accounting/) · 40 works · [💼 Management Information Systems](https://md.research.elamur.ai/free/en/shelves/business-management-and-accounting/management-information-systems/) · 9 works · [📈 Finance](https://md.research.elamur.ai/free/en/shelves/economics-econometrics-and-finance/finance/) · 6 works

🧭 Topics (1): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,894 works

## Cite

Cite as: Zhengzhe Yang. 2026. When Valid Signals Fail: Regime Boundaries Between LLM Features and RL Trading Policies. Proceedings of the Second Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U).

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{yang2026when,
  title = {{When Valid Signals Fail: Regime Boundaries Between LLM Features and RL Trading Policies}},
  author = {Zhengzhe Yang},
  booktitle = {Proceedings of the Second Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U)},
  year = {2026},
  doi = {10.18653/v1/2026.customnlp4u-1.17},
  url = {https://doi.org/10.18653/v1/2026.customnlp4u-1.17}
}
```

- [← previous When Users Are Happy but Agents Are Wrong: Multi-Dimensional Evaluation of Tool-Augmented Dialogue](https://md.research.elamur.ai/free/en/papers/shourya-2026-when-users-are-happy-but/)
- [next → When Words Don't Mean What They Say: Figurative Understanding in Bengali Idioms](https://md.research.elamur.ai/free/en/papers/sakhawat-2026-when-words-don-t-mean/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/56/)

## References

12 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Drozdov, Brad Mullis, Chenyu Yue, Steve Ostrum, and 1 others. 2023. [Bloomberggpt: A large language model for finance](https://md.research.elamur.ai/free/en/papers/wu-2023-bloomberggpt-large-language-model-finance/). arXiv preprint arXiv:2303.17564.
2. Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. 2023. [Fingpt: Open-source financial large language models](https://md.research.elamur.ai/free/en/papers/yang-2023-fingpt-open-source-financial-large/). FinLLM Symposium at IJCAI 2023.
3. Boyu Zhang and 1 others. 2023. [Pixtral & finma: Instruct-finllm for financial domain](https://md.research.elamur.ai/free/en/papers/yang-2023-fingpt-open-source-financial-large/). arXiv preprint arXiv:2306.06031.
4. Andrew Ang and Geert Bekaert. 2002. International asset allocation with regime shifts. The Review of Financial Studies, 15(4):1137–1187. [find](https://scholar.google.com/scholar?q=Andrew+Ang+and+Geert+Bekaert%2E+2002%2E+International+asset+allocation+with+regime+shifts%2E+The+Review+of+Financial+Studies%2C+15%284%29%3A1137%E2%80%931187%2E)
5. Ben Hambly, Ruiwei Xu, and Huining Yang. 2023. Recent advances in reinforcement learning in finance. Mathematical Finance, 33(3):437–503. [find](https://scholar.google.com/scholar?q=Ben+Hambly%2C+Ruiwei+Xu%2C+and+Huining+Yang%2E+2023%2E+Recent+advances+in+reinforcement+learning+in+finance%2E+Mathematical+Finance%2C+33%283%29%3A437%E2%80%93503%2E)
6. Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T Joshi, Hanna Moazam, and 1 others. 2023. [Dspy: Compiling declarative language model calls into state-of-the-art pipelines](https://md.research.elamur.ai/free/en/papers/khattab-2023-dspy-compiling-declarative-language-model/). arXiv preprint arXiv:2310.03714.
7. Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. 2020. Conservative q-learning for offline reinforcement learning. In Advances in Neural Information Processing Systems, volume 33, pages 1179– 1191. [find](https://scholar.google.com/scholar?q=Aviral+Kumar%2C+Aurick+Zhou%2C+George+Tucker%2C+and+Sergey+Levine%2E+2020%2E+Conservative+q%2Dlearning+for+offline+reinforcement+learning%2E+In+Advances+in+Neural+Information+Processing+Systems%2C+volume+33%2C+pages+11)
8. Xiao-Yang Liu, Hongyang Yang, Qian Chen, Runjia Zhang, Linyi Yang, Bowen Xiao, and William Wang. 2020. Finrl: A deep reinforcement learning library for automated stock trading in quantitative finance. [find](https://scholar.google.com/scholar?q=Xiao%2DYang+Liu%2C+Hongyang+Yang%2C+Qian+Chen%2C+Runjia+Zhang%2C+Linyi+Yang%2C+Bowen+Xiao%2C+and+William+Wang%2E+2020%2E+Finrl%3A+A+deep+reinforcement+learning+library+for+automated+stock+trading+in+quantitative+finance%2E)
9. Deep RL Workshop, NeurIPS 2020. [find](https://scholar.google.com/scholar?q=Deep+RL+Workshop%2C+NeurIPS+2020%2E)
10. Alejandro Lopez-Lira and Yuehua Tang. 2023. Can chatgpt forecast stock price movements? return predictability and large language models. arXiv preprint arXiv:2304.07619. [find](https://scholar.google.com/scholar?q=Alejandro+Lopez%2DLira+and+Yuehua+Tang%2E+2023%2E+Can+chatgpt+forecast+stock+price+movements%3F+return+predictability+and+large+language+models%2E+arXiv+preprint+arXiv%3A2304%2E07619%2E)
11. H Nejat Seyhun. 1998. Investment intelligence from insider trading. MIT press. [find](https://scholar.google.com/scholar?q=H+Nejat+Seyhun%2E+1998%2E+Investment+intelligence+from+insider+trading%2E+MIT+press%2E)
12. Hao Sun and 1 others. 2023. Market regime aware reinforcement learning for quantitative trading. Proceedings of the ICAIF. [find](https://scholar.google.com/scholar?q=Hao+Sun+and+1+others%2E+2023%2E+Market+regime+aware+reinforcement+learning+for+quantitative+trading%2E+Proceedings+of+the+ICAIF%2E)

*Updated 2 November 2026*