What can AI do in equity research today?
The honest short answer to can AI do equity research is that it depends which task is meant. AI compresses the distance between disclosure and understanding: scanning hundreds of pages, extracting figures, and drafting first-pass summaries at a scale no analyst sustains manually.
But capacity is not judgment. Current systems remain unreliable at deciding what matters to a specific thesis, why a change occurred, and what would prove a view wrong. This series therefore never issues a single score — only capability profiles across accuracy, coverage, reasoning, evidence, context, and speed.
1. Executive Summary Most answers come from demos or anecdotes. FinanceBench (10,000+ triplets) found 81% wrong-or-refused. Deep FinResearch Bench and FinDeepResearch (64 companies, 8 markets, 15,000+ items) find AI behind on KPIs and scenarios. Yet Chicago work shows GPT-4 beating analysts on narrow directional prediction. AI already matches humans on narrow structured tasks. AI still trails on judgment-heavy work. Only workflow-grounded tests with human baselines can answer what helps a real desk.
2. Why Equity Research Is Twelve Tasks, Not One Screening, business understanding, financial analysis, filings, earnings, competitor analysis, management assessment, thesis formation, valuation, risk, monitoring, and recommendation. An AI strong at extraction may be weak at materiality. Results from one stage must never be generalized.
3. Retrieval and Question-Answering at the Filing Level FinanceBench: 150 cases, 2,400 answers — hallucination and retrieval failure dominated. FinAgentBench: multi-step document-then-passage retrieval still open. Test: find the figure, cite the location, avoid blending nearby facts.
4. Report Generation and Forecasting Evidence Deep FinResearch Bench: AI well structured but weaker on KPI selection and scenarios. Kim, Muhn, Nikolaev (Chicago): GPT-4 beat analysts on direction from clean statements. Korean 2026 study (121,369 forecasts, 607 firms): humans +~20% on average, wider for loss-making and small caps. Cutoff discipline is non-negotiable.
5. The Methodology Brane Will Reuse Task mapped to workflow stage. Inputs specified precisely. Information cutoff with temporal-leakage controls. AI system recorded in reproducible detail. Human baseline: analyst alone and analyst-plus-AI on speed, accuracy, evidence, and usefulness.
6. Failure Taxonomy and Scorecard 13 patterns: hallucination, numerical error, retrieval failure, context failure, materiality failure, temporal leakage, causal error, false synthesis, confirmation bias, contradiction blindness, overconfidence, source-quality failure, missing-context failure. Scorecard, not a score: accuracy, evidence, coverage, materiality, reasoning, contradiction handling, temporal integrity, usefulness — reported separately.
7. Proposed Experiments and Recurring Format Read a 10-K. Earnings-call new vs. repeated language. Statement trends with verified numbers. Competitor comparison without false equivalence. Thesis construction. Thesis stress test. Historical reconstruction under cutoff. Monitoring for thesis-material events. Each edition: task, test, rules, AI output, human baseline, verdict, failures, delegation guidance, one takeaway.
8. Five Documented Weaknesses 1. Demo bias — 81% wrong-or-refused shows variance. 2. No standardized evaluation across benchmarks. 3. Accuracy conflated with usefulness. 4. Prediction obsession over full research chain. 5. Failure invisibility in marketing.
9. Five Concrete Improvements 1. Publish workflow-grounded task definitions. 2. Build expert ground-truth datasets. 3. Require evidence-linked grading with locations. 4. Enforce strict temporal controls. 5. Report genuine human baselines including analyst-plus-AI.
10. Conclusion Can AI do equity research as a single capability? No. Can it do stages under documented conditions? Sometimes yes for extraction and organization, while assumptions, scenarios, and materiality stay human. Each edition will test one task and publish failures.
Where Brane Fits: measurement infrastructure for verified retrieval, evidence chains, and thesis-aware monitoring — authority earned experiment by experiment.
FAQ Q1: Can AI do equity research end to end today? No — strong on extraction, gaps on materiality and scenarios. Stage-specific leverage only. Q2: What benchmark evidence matters most? FinanceBench 81%, Deep FinResearch / FinDeepResearch on KPIs, Chicago on narrow prediction. Q3: How will Brane avoid demo bias? Pre-registered tasks, side-by-side outputs, dimensional scorecard, full failure anatomy. Q4: Why does temporal leakage matter? Using unavailable information invalidates the test. Explicit cutoffs enforced. Q5: What should analysts delegate now? First-pass retrieval and scanning with verification. Retain judgment and the decision.
Sources: FinanceBench (2023); FinAgentBench; Deep FinResearch Bench (JPMorganChase 2026); FinDeepResearch (64/8/4); Kim, Muhn, Nikolaev (Chicago 2024); An and Lee (Korean 121,369 forecasts 2026); CFA Institute; ESMA 2025.
This article is educational and does not make a recommendation about any security. Evidence as of September 2026.