LLM agents vs quantitative algorithms.

The honest position on LLMs in share trading, mid-2026: they are transformational for extracting structured information out of unstructured financial text, and they are not a substitute for a quantitative decision engine. The papers that claim otherwise do not survive careful re-runs of their own experiments.

Section IWhere LLMs earn their keep

The reliable pattern is LLM as feature engineer, deterministic model as decision-maker.

Lopez-Lira and Tang (arXiv 2304.07619) showed GPT headline sentiment produces a long-short signal, strongest in small caps. Kirtac and Germano (arXiv 2412.19245) report Sharpe 3.05 net of ten basis points transaction costs using OPT over 965,000 news items. Cao et al. (arXiv 2510.03195) run GPT-4o-mini on transcripts after its training cutoff and still find abnormal-return signal — a rare, honest robustness check that rules out lookahead leakage. BlackRock’s LoRA-tuned Llama and Mistral news-embedding work (arXiv 2407.18103) treats LLM output as features for a downstream model, not as signals in themselves. On hawk-versus-dove classification of central-bank speech, BIS Working Paper 1215 shows fine-tuned domain language models beat GPT-class general models.

Structured extraction and retrieval-augmented generation over filings is the other confirmed win. Anthropic’s contextual retrieval pattern — prepending chunk context before embedding, adding BM25, then reranking — reduces retrieval errors by roughly 49% on the SEC-filings worked example. Reference stacks include LlamaIndex’s sec-insights for 10-K/Q questions, LlamaExtract for schema-driven extraction, and LangChain’s EDGAR retriever. JPMorgan’s DocLLM (arXiv 2401.00908) beats GPT-4 on form understanding — the layout-aware answer for messy PDF filings.

Section IIWhere LLMs fail

• Full analyst / numeric reasoning over filings. FinanceBench and FAITH already told you why. FinSheet-Bench adds that Gemini 2.5 Pro drops from 82.4% to 48.6% as spreadsheet complexity rises. Glasserman and Lin (SSRN 4586726) show anonymising tickers materially shrinks GPT “alpha” — proof of look-ahead leakage.
• HFT and execution. Nasdaq/NYSE round-trip is measured in microseconds; NVIDIA’s capital-markets inference target is single-digit microseconds. LLM inference is 100 milliseconds or more — two to five orders of magnitude off. A recent survey states it plainly: “unsuitable for high-frequency trading environments.”
• Position sizing and P&L arithmetic. Every numeric-reasoning benchmark shows LLMs collapsing on multi-step calculations. Use a deterministic calculator; make the LLM emit structured intents, not numbers.
• Red-teaming quant strategies. Embryonic literature only. TradeTrap (arXiv 2512.02261) perturbs intelligence and execution to break LLM trading agents; arXiv 2601.13082 uses adversarial headlines. Useful for stress-testing LLM-in-the-loop systems, not yet a mature adversary for pure quant.

Section IIIThe hybrid pattern

The design pattern that keeps recurring in the credible literature and in production deployments looks like this:

1. LLM as feature engineer. Read news, tag filings, extract structured facts, classify sentiment. Output goes into a database as features with timestamps and source lineage.
2. Deterministic model as decision-maker. Classical quant or ML model consuming those features alongside numerical price / volume / fundamentals to output a signal.
3. Deterministic risk engine. Position sizing, exposure limits, kill switches. Never inside an LLM.
4. LLM as challenger. A separate LLM red-team agent tries to disprove the thesis before an order is proposed.
5. Restricted execution service. Signed, validated, schema-checked orders. No language model inside the broker-connection path.

This is precisely the shape of the eleven-agent architecture set out in the companion blueprint at agentteam.co.uk/share-trading.html. It also matches the pattern the more careful multi-agent trading papers converge on (analyst-debate-trader-risk-manager) once you strip away their headline P&L claims.

Section IVCost economics, mid-2026

Batch APIs cut roughly 50% off list price; prompt caching cuts input around 10×. Tagging a 30,000-token filing with Haiku 4.5 costs about $0.03; on flagship tier around $0.30. That order-of-magnitude gap decides whether an LLM-derived feature is a viable factor or a science project.

Open-source alternatives

Llama 4 Scout offers a 10-million-token context on a community licence (700M monthly-active-user cap; no training-on-outputs) — a genuinely useful window for tick-history analysis. DeepSeek V3.2 / R1 are MIT-licensed and R1 is currently the strongest open reasoner. Qwen 3.5 (Apache 2.0, 262K native context) is the multilingual all-rounder. Mistral Large 3 and Small 4 are now Apache 2.0 — the cleanest compliance story for MiFID II shops keeping data in-region.

Section VThe one-page verdict

---

Next: Multi-agent architecture →