Large language models went from research curiosity to desk tool in about three years, and trading has not been an exception: banks summarize research with them, funds parse filings with them, and retail traders ask them for strategy ideas. What an LLM actually is, though — a probabilistic next-token predictor trained on text — determines very precisely what it can and cannot do for a trader. This article draws that line.
Where LLMs genuinely earn their keep
Where they fail, predictably
A sane architecture
The pattern that works in practice: LLMs as feature extractors and research assistants, classical models as decision makers. Text goes in; structured, validated fields come out (sentiment score, event type, surprise direction); those fields join price-based features in a statistical or ML model that is backtested with the usual discipline — walk-forward, after costs, against boring baselines. The LLM never touches an order, and every extraction is logged with its source so failures can be audited.
The bottom line
Treat an LLM like an exceptionally fast junior analyst with unlimited reading capacity, an occasional tendency to invent facts, and zero trading judgment. Let it read, extract, summarize and draft code; validate everything it produces; and keep the actual predictions and position decisions in models whose assumptions you can test. Traders who use LLMs this way get a real edge in research throughput — those who ask them for trade calls get fluent noise.
Where LLMs genuinely earn their keep
- Reading at scale. The core superpower. An LLM can summarize hundreds of headlines, earnings-call transcripts, central-bank statements or regulatory filings into structured takeaways in seconds. Classic NLP pipelines needed a custom model per task; an LLM does passable zero-shot extraction from a well-written prompt.
- Sentiment and event tagging without training data. Traditional sentiment models must be trained on labelled financial text. An LLM can label "hawkish / dovish", "guidance raised / cut", or "supply disruption" out of the box, and its errors are usually explainable when you ask it to justify the label. For low-frequency signals (daily/weekly), this is a real alternative to building a bespoke classifier.
- Code generation for research. Writing the boilerplate of a backtest, translating an indicator from one platform's language to another, vectorizing a slow loop, explaining an unfamiliar codebase — LLMs cut hours to minutes. The output still needs review, but the productivity gain in strategy research is the least controversial LLM win in trading.
- Structured extraction. "Return JSON with ticker, event type, direction and date for every corporate event in this text" works remarkably well and feeds directly into a systematic pipeline.
Where they fail, predictably
- They do not predict prices. An LLM has no market model, no live data and no notion of probability calibrated to returns. Asking one "will EUR/USD go up?" produces fluent, confident, worthless text. Any real predictive power comes from the features an LLM extracts (sentiment, events) being fed into a proper statistical model — not from the LLM's own opinion.
- Hallucination with conviction. LLMs fabricate numbers, tickers, dates and citations while sounding authoritative. Never let un-verified LLM output flow into an execution path. Extraction tasks must be validated against the source text, ideally with a second pass that checks each claimed fact.
- Training-data lookahead. A model trained on text through last year "knows" what happened to markets in its training window. Backtesting an LLM-derived signal inside that window silently leaks the future — the model may label a 2022 headline bearish because it knows what came next. Honest evaluation uses data after the model's cutoff, or point-in-time models.
- Non-determinism and drift. The same prompt can return different answers across runs and model versions. A signal that changes when the provider updates the model is a dependency risk classic quant infrastructure never had. Pin versions, log everything, and re-validate on every upgrade.
- Latency and cost. Hundreds of milliseconds to seconds per call rules LLMs out of any latency-sensitive loop. They belong in the research and signal-preparation layer, not inside the order path.
A sane architecture
The pattern that works in practice: LLMs as feature extractors and research assistants, classical models as decision makers. Text goes in; structured, validated fields come out (sentiment score, event type, surprise direction); those fields join price-based features in a statistical or ML model that is backtested with the usual discipline — walk-forward, after costs, against boring baselines. The LLM never touches an order, and every extraction is logged with its source so failures can be audited.
The bottom line
Treat an LLM like an exceptionally fast junior analyst with unlimited reading capacity, an occasional tendency to invent facts, and zero trading judgment. Let it read, extract, summarize and draft code; validate everything it produces; and keep the actual predictions and position decisions in models whose assumptions you can test. Traders who use LLMs this way get a real edge in research throughput — those who ask them for trade calls get fluent noise.