Last week, researchers published a study that should matter to anyone using ChatGPT for investment ideas. They tested GPT-4, Gemini, and DeepSeek on stock selection tasks. The headline: AI outperformed the market — but only when a human was checking its work.
Without supervision, every model made the same three mistakes: hallucinating financial data that did not exist, carrying errors forward across multi-step analysis, and confusing correlation with causation in market trends.
Sound familiar? If you have ever asked ChatGPT “should I buy NVDA?” and gotten a confident answer that felt slightly off, this is why.
The good news: there is a framework that makes AI stock analysis genuinely useful. Here is how it works.
The biggest failure mode is asking AI to analyze a stock from memory. It will confidently cite a PE ratio from 2023 or invent an earnings date. Instead:
I am analyzing [TICKER] for a potential [BUY/SELL] position. Here is the current data (as of [DATE]): - Price: $[X] - PE (TTM): [X] - Revenue growth (YoY): [X]% - EPS (last 4 quarters): [Q1, Q2, Q3, Q4] - Sector: [X] - Next earnings: [DATE] Based ONLY on the data I provided (do not use your training data for financial figures), analyze: 1. Is this stock cheap or expensive relative to its sector? 2. What is the biggest risk in the next 90 days? 3. What would change your mind about this position?
Why this works: The key phrase is “based ONLY on the data I provided.” This eliminates hallucinated financials — the #1 failure mode the researchers identified.
Where to get the data: Yahoo Finance (free), Finviz (free screener), or your broker’s research tab. Copy-paste the numbers. 30 seconds. Done.
Most people use AI as a confirmation machine. You think NVDA is a buy, so you ask AI to make the bull case, and it obliges. This is useless. Instead:
I believe [TICKER] will [rise/fall] over the next [timeframe] because [your thesis]. Your job is to destroy this thesis. Find the strongest counterarguments. Use the data I provided above. Be specific — cite numbers, not generalities. I want to feel uncomfortable reading your response.
The test: If the AI cannot produce a compelling counter-thesis, your position is probably strong. If it produces three counter-arguments you had not considered, you just avoided a costly mistake.
AI is excellent at analysis but terrible at timing. It does not know what happened after its training cutoff. It cannot feel market sentiment. It has no concept of “the trade is crowded.”
Rule: Use AI for the what (which stocks, which direction, which risks). Use your own judgment for the when (entry timing, position sizing, stop placement). The researchers found AI recommendations improved by 40% when humans handled the execution layer.
Tool of the Week: Finviz Screener + Claude
Finviz’s free stock screener lets you filter by fundamentals and export results as a table. The workflow:
- Go to finviz.com screener
- Set filters (PE under 15, Market Cap over $10B, Insider Buying > 0)
- Copy the results table
- Paste into Claude with: “Rank these stocks by investment quality. For each: one reason to buy, one to avoid, confidence level. Flag any uncertain data.”
3 hours of tab-switching Prioritized research list in 60 seconds
Eliminates hallucinated data — the #1 AI failure mode in finance.
Quick Prompt: The Earnings Preview
Before any earnings report, paste this:
[TICKER] reports earnings on [DATE]. Consensus EPS estimate is $[X], revenue estimate is $[X]B. In 3 bullet points: 1. What result would surprise the market positively? 2. What result would cause a sell-off? 3. What should I watch in the conference call that analysts might miss?
15 seconds to paste. A framework for processing the earnings release in real-time instead of reacting to the headline.
The Automated Research Pipeline: How I use AI to monitor 500+ stocks daily without reading a single analyst report. The exact system, from data collection to actionable signals.