AI stock analysis: how it works and where it breaks
By MacrosLM Team · Reviewed by Damira Baigozha, ex-PwC Valuation & M&A Advisory Expert
AI stock analysis is the use of large language models and machine learning tools to screen securities, summarize filings and earnings calls, build comparable-company sets, and gauge market sentiment faster than a manual process allows. It does not predict future prices, and it does not replace analyst judgment — it compresses the research workflow so a human can spend more time on the parts that require judgment: interpreting ambiguity, weighing qualitative risk, forming a view.
That distinction matters because the marketing around AI investing has run well ahead of what the technology reliably does. This post covers what AI stock analysis is genuinely good at, where it breaks down, and how to use it without fooling yourself.
What AI stock analysis actually does well
The strongest use cases share a trait: processing large volumes of structured or semi-structured text and data faster than a person could, with output that's easy to spot-check against a source. That means screening thousands of tickers against valuation, growth, or quality criteria in seconds; summarizing filings and earnings calls to pull guidance changes, new risk factors, or tone shifts against the prior period; suggesting comparable-company sets from business descriptions or reported segments; aggregating sentiment across news, transcripts, and filings to flag directional shifts; and turning a multi-hour filing read into a shorter review of AI-flagged sections.
Task-by-task: is AI good at it?
| Task | Is AI good at it? | Why |
|---|---|---|
| Screening on quantitative criteria | Yes | Filtering structured data is a rules-and-math problem AI handles reliably and fast. |
| Summarizing 10-Ks and earnings calls | Yes, with review | Strong at extraction; can miss nuance or hedged language, so a source check is needed. |
| Building comps sets | Partially | Good at a candidate list from business descriptions; weak at judging true comparability. |
| Sentiment analysis | Yes, directionally | Reliable at detecting tone shifts across large text volumes; not a standalone trading signal. |
| Forecasting a stock price | No | Prices reflect forward-looking information that doesn't exist yet. No model reads the future. |
| Replacing analyst judgment on a thesis | No | Weighing qualitative risk and competitive dynamics requires judgment AI doesn't have. |
| Flagging anomalies vs. prior filings | Yes | Pattern-matching between periods is exactly what these tools are built for. |
3
Reliable on their own
2
Useful with a source check
2
AI can't do
The green rows above share a trait: high-volume text or data work that's easy to spot-check. Prediction and judgment are the tasks AI can't do — and the ones that stay yours.
Where the hype outruns the technology
A wave of "AI investing" apps market themselves as if they can identify winning stocks or time entries and exits. That claim doesn't hold up, for a structural reason rooted in the efficient-market hypothesis (Eugene Fama, Nobel Prize 2013): public markets price in available information quickly, and a model trained on historical patterns has no privileged access to what happens next. An AI tool can tell you sentiment on a name has turned negative. It cannot tell you what the stock will do tomorrow.
Regulators have started to act on the gap between claim and capability. In 2024 the SEC charged two investment advisers for "AI washing" — making false or misleading statements about their use of AI — its first enforcement actions of the kind. The more defensible framing is that AI is a research accelerant, not a forecasting engine. This mirrors a broader pattern in AI across banking and finance: staying-power applications augment a defined workflow — document review, reconciliation, screening — rather than promise to out-predict the market. The same holds for machine learning in finance: strong at pattern detection, weak at genuine prediction of novel events.
How to evaluate a stock with AI, step by step
Most people use AI for stocks the wrong way — they ask for a conclusion ("is this a buy?") when they should use it as a research assistant that summarizes, structures, and stress-tests. Here's a workflow that holds up, with a copy-paste prompt for each step and where MacrosLM does the same step natively.
Understand how the company makes money before you look at a single figure.
Prompt
Explain [Company]'s business model as if I were buying the entire company. How does it make money, what are the reporting segments and their revenue share, and who are the customers? Base it on the latest 10-K and cite where each point comes from.
What good looks like: named segments with a revenue mix and customer types, not a Wikipedia paragraph.
In MacrosLM: the Company Equity Deep Dive (Overview tab) pulls this from the latest filing through the SEC EDGAR connector, every claim linked to its source.
Find the moat, and be specific about it.
Prompt
What is [Company]'s durable competitive advantage? Name the specific moat (switching costs, network effects, scale, brand) with evidence. What would a competitor need to do to take share over the next 5-10 years?
What good looks like: a specific, evidenced moat. Push back on generic “strong brand.”
In MacrosLM: the Deep Dive structures the peer set and competitive notes alongside the fundamentals.
Don't trust a general chatbot for numbers — they hallucinate figures and confuse periods. Pull the data from the filing and make the model cite it.
Prompt
From the attached 10-K, extract 5 years of revenue, gross/operating/net margin, ROE, net debt, and free cash flow into a table. Give the exact page or statement reference for every figure. Do not estimate any number that isn't in the document.
What good looks like: a table where every number carries a page reference you can check.
In MacrosLM: the Comprehensive Financial Ratio & DuPont Analyzer computes the full ratio set from the filing, each figure clickable back to source — so there's no invented number to catch.
This is where AI helps most — if you force it to be specific.
Prompt
Give the strongest, most specific bull case for [Company] over 3-5 years — name catalysts, not “tailwinds.” Then the strongest bear case: what specifically would have to go wrong for this to be a bad investment?
What good looks like: named catalysts and named risks. Reject “macro headwinds.”
In MacrosLM: the Deep Dive's bull/bear tab drafts both sides from the financials and filings, each point traceable.
Valuation needs current market data and your view; verify multiples against a data source.
Prompt
Compare [Company]'s current P/E, EV/EBITDA, and P/S to its own 5-year history and to three named peers. Flag where it looks rich or cheap and what would have to be true to justify the gap.
What good looks like: a comparison against the company's own history and named peers — and the assumptions stay yours.
In MacrosLM: dedicated DCF, relative-valuation, and WACC agents build the valuation with a live WACC × terminal-growth sensitivity table you can flex.
A thesis that only exists in your head isn't a thesis.
Prompt
Draft a one-page investment thesis for [Company]: company type (grower / stalwart / turnaround / cyclical), the specific bull case, the specific bear case, and explicit sell criteria I commit to before buying.
What good looks like: a written thesis with sell triggers you'd actually honour when the stock drops.
In MacrosLM: the Deep Dive assembles the five-tab worksheet (overview, financials, valuation, peers, bull/bear) in your house style, every figure source-traced — the thesis is yours to sign.
Or skip the six prompts — press run
MacrosLM runs all six steps in one workspace: it pulls the filings, extracts and cites every figure, builds the valuation with a live sensitivity table, and drafts the bull/bear and thesis — each number clickable back to its source, saved to re-open, and re-runnable on a schedule. What it won't do is predict a price or hand you conviction: you verify the read and own the call.
One rule sits over all six steps: verify every AI-surfaced number against the primary filing before it goes into a model or memo. AI summarization tools occasionally misstate a figure or miss a footnote qualifier — the failure mode is usually subtle rather than absurd, which is why it needs checking rather than trusting. For the tools built specifically for this workflow, see our roundup of AI tools for equity research.
Why a workspace beats ad-hoc prompting
Even done well, the six-prompt approach has limits no clever prompting fixes. You rebuild the wheel every time — five stocks means five separate chats and five slightly different frameworks, so you can't compare them cleanly. The data problem is real: general models pull figures from aggregators, confuse trailing-twelve-months with annual, or invent a number when they're unsure. And nothing is saved — the analysis you built on a Sunday is buried in a chat history two months later, with no way to check whether the thesis still holds. A purpose-built workspace fixes all three: one consistent, source-traced framework, real figures linked to filings, and a saved deliverable you can re-run on a schedule.
Where MacrosLM fits
Disclosure: MacrosLM is our own product. It's built for the workflow above: agents like the Earnings Analyzer, Company Equity Deep Dive (5-Tab), and Comprehensive Financial Ratio & DuPont Analyzer handle screening, filing and transcript summarization, and comps construction — pulling from the SEC EDGAR connector with every figure linked to its filing, so outputs are designed to be checked rather than blindly trusted. It does not forecast prices or give investment advice, and it flags when a figure needs a human look rather than presenting a guess as fact.
Bottom line
AI stock analysis is genuinely useful for screening, summarizing filings and calls, drafting comps sets, and tracking sentiment — the parts of research that are high-volume and checkable. It is not useful, and shouldn't be marketed as useful, for predicting stock prices or substituting for analyst judgment. The tools that hold up are the ones that make verification faster, not the ones that ask you to skip it.
Sources
- Nobel Prize — Eugene Fama (2013) Efficient-market hypothesis — the structural reason a model trained on historical patterns can't forecast prices.
- SEC — Charges against two investment advisers for “AI washing” (2024) First SEC enforcement actions over false or misleading statements about the use of AI.
This post is educational, not investment advice, and nothing here recommends buying, selling, or holding any security. Whatever an AI tool surfaces should be checked against the primary filing or transcript before it informs a real decision.
Frequently asked questions
- Can AI predict which stocks will go up?
- No. AI models are trained on historical patterns and can't access information that doesn't yet exist, which is what future stock moves depend on. Tools marketed this way are describing a probability or sentiment score, not a forecast — treat the output as a prompt to investigate, not a signal to act on.
- Is AI stock analysis reliable for reading earnings calls?
- It's reliable for extraction and summarization — flagging guidance changes, new risks, or tone shifts — but it can miss nuance in hedged management language. Treat the summary as a starting point and verify anything material against the transcript.
- What's the difference between AI stock analysis and algorithmic trading?
- AI stock analysis supports human research — screening, summarizing, comps — with a person making the final call. Algorithmic trading executes rules or models directly in the market, often without a human in the loop per trade. This post is about the former.
- Can AI build a reliable comparable-company set?
- It can generate a strong candidate list from business description and reported segments, but comparability also depends on capital structure, accounting policy, and growth stage — factors that still need a manual check before the comps go into a valuation.
- Should I use AI output directly in a valuation model?
- Only after verifying the underlying figures against the primary filing. AI-generated summaries occasionally misstate or conflate numbers; the error is usually subtle, which is why every figure feeding a model should be checked at the source.
Reviewed by Damira Baigozha
ex-PwC Valuation & M&A Advisory Expert. Written by the MacrosLM editorial team.
View profile →
