Prompting AI for investment decisions: a short playbook

    Five prompt patterns that get useful answers out of LLMs for equity research — and three patterns that produce confident garbage.

    6 min readUpdated 2026-05-20AI for investors

    An LLM's answer is only as good as the prompt and the data it sees. These patterns are the difference between a useful research draft and a confidently wrong one.

    Patterns that work

    All of these share one feature: they ground the model in source data instead of asking it to recall facts.

    • Constrained summarization: 'Summarize this 10-K in 10 bullets, citing the page.'
    • Structured extraction: 'Extract guidance for FY revenue and EPS from this transcript.'
    • Comparative tables: 'Compare these five peers on gross margin, ROIC and FCF yield.'
    • Devil's advocate: 'Write the strongest bear case against this thesis using only this filing.'
    • Calibration: 'Rate your confidence 1–5 and explain the weakest link.'

    Patterns that produce hallucinations

    Avoid open-ended fact retrieval ('What was AAPL's Q3 2024 gross margin?') without giving the model the source — it will guess. Avoid asking for predictions ('Will NVDA beat next quarter?') — you'll get a vibe, not analysis. And never ask for a recommendation without specifying the time horizon and risk tolerance.

    Keep a verification habit

    Every number the model produces should be traceable to a source you can open in one click. Edge Trader Terminal enforces this by feeding the model structured FMP data and labeling outputs with their data source and timestamp.

    Apply this in Edge Trader Terminal

    Run AI-powered stock analysis on 10,000+ tickers — fundamentals, technicals and fair value in one report.