Claude 一句話股票篩選的三種潛在誤導

The Equibles MCP server has a ScreenStocks tool, so with it connected to Claude or ChatGPT a one-sentence description of the stocks you want becomes a set of filters and a table. I ran one screen through it on November 3, 2026, and it showed three ways a screen can quietly mislead you. Every figure below is from those tool calls.

The sentence and the filters

Find US stocks worth more than $2 billion that are profitable, grew revenue more than 20% in their latest quarter, have gross margins above 50% and trade below 30 times earnings. Skip anything that trades less than $10 million a day.

That sentence maps to six arguments:

{
  "minMarketCap": 2000000000,
  "minNetIncome": 0,
  "minRevenueGrowth": 20,
  "minGrossMargin": 50,
  "maxPe": 30,
  "minDollarVolume": 10000000
}

32 stocks matched. The reply also dated every column: prices from the November 2 close, short interest from FINRA's October 15 settlement, and 13F holder counts from the quarter ended June 30. That's the first habit worth keeping: ask the assistant to show its arguments, and check that every number in your sentence became exactly one filter.

Trap 1: a missing number removes the stock

I then ran the same screen on a watchlist (the tool takes a tickers list): Apple, Microsoft, Amazon, Reddit, Palantir, SoFi, Robinhood and Uber. Only Reddit passed, which I expected. What I didn't expect was that Robinhood and Uber didn't fail any filter. They vanished, because the screen excludes a stock missing any bounded metric, and neither has a gross margin on file.

If you build or use a screener, "no data" and "fails the test" should look different. Ask the assistant which filter removed each stock.

Trap 2: the same growth number can mean different things

Sorted by revenue growth, the top ten were Micron (379%), Sandisk (378%), Royal Gold (115%), Nvidia (106%), Chord Energy (84%), TG Therapeutics (70%), Franco-Nevada (57%), Western Digital (56%), Reddit (53%) and Hinge Health (53%). Memory makers, gold royalty companies and an oil producer sell commodities, so the prices of what they sell move their revenue. That's a different kind of growth from Reddit's or Hinge Health's, and the filter can't tell them apart.

Trap 3: cheap is a question, not an answer

Reddit passed at 18.7 times trailing earnings. GetValuationMultiplesHistory showed it at 53.4 times in early February. Earnings rose over that time (Q3 net income $283.3M against $162.7M a year earlier), so the multiple fell because the price fell: $129.60 on October 29, $91.73 on October 30, down 29.2% the day after the report.

GetEarningsBrief gave the reason in the CEO's own words: search referrals fell by more than a million a week in October, and if that continued, daily active uniques "could fall in Q4 by a low double-digit percentage." A screen finds what's cheap; one more prompt tells you why.

The one-prompt version

Screen US stocks worth more than $2 billion that are profitable, grew revenue more than 20% last quarter, have gross margins above 50%, trade below 30 times earnings and over $10 million a day. Sort by revenue growth. For the three most familiar names, compare the valuation with peers and with the company's own history, explain any big price drop since the last report from the earnings call, and list insider sales and short interest. Give the date of every number.

The full walkthrough, with six prompts, every answer and how to check each one, is here:

https://equibles.com/blog/how-to-screen-for-stocks-with-chatgpt-or-claude

How do you handle missing data in your own screens: exclude, flag, or fill?

Data: SEC filings, FINRA short interest, 13F filings and market prices, processed by Equibles.