AI price predictions in the media, graded

Finance sites publish articles in which ChatGPT or another AI model names a price for a stock on a date. Once the date passes, nobody goes back to check. We do: each prediction below is graded against the stock's actual close on its target date, beside the forecast you would make by doing nothing (the close on the day the article ran). Predictions we could not grade fairly are listed with the reason, and they stay in the count. StockTools makes no predictions; this page only grades ones already published.

By publisher, and by the model named

Every graded prediction

Miss is the actual against the predicted price (a range's midpoint), as a percent of the prediction. For a prediction of an average price over a period, the actual is the average close over that period. Doing nothing is the miss of a forecast that the price stays at the close on the day the article ran.

Abstained, with the reason

Published before our price history begins, so a split in between cannot be ruled out. 4; The averaging window began before the article ran, so part of the outcome was already known. 7; The article does not pin the claim down well enough to grade. 1; It names an intraday high or a take-profit level from a trade plan, not a price on a date. 4.

• Finbold, 2024-05-16: TSLA by 2024-12-31. Published before our price history begins, so a split in between cannot be ruled out. (history from 2024-07-01)
• Finbold, 2024-05-21: NVDA by 2024-12-31. Published before our price history begins, so a split in between cannot be ruled out. (history from 2024-07-01)
• Finbold, 2024-01-11: AAPL by 2024-12-31. Published before our price history begins, so a split in between cannot be ruled out. (history from 2024-07-01)
• Finbold, 2024-01-26: BUD by 2024-12-31. Published before our price history begins, so a split in between cannot be ruled out. (history from 2024-07-01)
• Benzinga, 2025-12-17: AMD by 2026-01-06. The averaging window began before the article ran, so part of the outcome was already known. (window from 2025-12-16, published 2025-12-17)
• Benzinga, 2025-12-17: DDOG by 2025-12-31. The averaging window began before the article ran, so part of the outcome was already known. (window from 2025-12-12, published 2025-12-17)
• Benzinga, 2025-12-17: AVGO by 2026-01-05. The averaging window began before the article ran, so part of the outcome was already known. (window from 2025-12-16, published 2025-12-17)
• Benzinga, 2025-12-15: SNOW by 2025-12-31. The averaging window began before the article ran, so part of the outcome was already known. (window from 2025-12-11, published 2025-12-15)
• Benzinga, 2025-12-17: PLTR by 2026-02-10. The averaging window began before the article ran, so part of the outcome was already known. (window from 2025-12-16, published 2025-12-17)
• Benzinga, 2025-12-22: SMCI by 2025-12-31. The averaging window began before the article ran, so part of the outcome was already known. (window from 2025-12-17, published 2025-12-22)
• Benzinga, 2025-12-22: AI by 2026-02-15. The averaging window began before the article ran, so part of the outcome was already known. (window from 2025-12-17, published 2025-12-22)
• Benzinga, 2026-02-09: META by 2026-04-10. The averaging window ends only "late April"; the article gives no end date.
• Benzinga, 2026-09-25: MU by 2026-10-02. It names an intraday high or a take-profit level from a trade plan, not a price on a date.
• Benzinga, 2026-09-25: MU by 2026-10-02. It names an intraday high or a take-profit level from a trade plan, not a price on a date.
• Benzinga, 2026-09-25: MU by 2026-10-02. It names an intraday high or a take-profit level from a trade plan, not a price on a date.
• Watcher.Guru, 2026-04-26: GOOG by 2026-04-30. It names an intraday high or a take-profit level from a trade plan, not a price on a date.

How each prediction was graded

• One rule, the settlement referee's: graded only when the prediction was published before its target date, names a price or a range and a date, and we hold the close that settles it. Otherwise it is abstained, listed above and counted.
• The actual is the closing price on the target date, or the last close before it when the date fell on a weekend or holiday (no further back than 5 days).
• Some articles predict an average price over a period rather than a close. Those are graded against the average close over the same period. When the article gives only the end of the period, it runs from the first session after publication; a period that began before the article ran is abstained, because part of its outcome was already known.
• An intraday high or a take-profit level from a trade plan is not a price on a date, so it is abstained rather than graded on a different question.
• Prices are our own split-adjusted daily closes, held from 2024-07-01. A price stated before a later split is divided by that split, so it compares like for like. A prediction published before our history begins is abstained, because a split in between cannot be ruled out.
• A median prints from 10 graded predictions and a rate from 20; counts always print.
• This is a retrospective of other publishers' claims. It does not say what any stock will do, and nothing here is advice.

Where the predictions came from

Each article was read one at a time and its claim recorded with a link, including where a publisher's terms limit automated collection. Corpus collected 2026-10-07.

Related: the AI answer ledger, which asks models a fixed question and scores the answers 30 days later, and our own AI prompts, which forbid predictions.