MeebleAI has now graded 171 live stock forecasts made by AI models, each one written down with a date, a price, a direction and a target before anyone knew how it would turn out. That's a small sample, and this post is as much about what it can't tell us as what it can. The numbers are as of 13 November 2026; the track record has the current ones.
The headline number
Across the 169 forecasts with a direction to grade, the AI called it right 39% of the time (66 of 169).
That sounds bad until you remember there are three answers, not two: up, flat or down. Guessing at random would land around 33%. So the AIs are a little better than a dice roll. Not by much, and on 169 calls that gap could easily be luck.
Price is graded separately: whether the stock landed within a tolerance band around the target. That happened 36% of the time (62 of 171).
By model
| Model maker | Direction right | Graded |
|---|---|---|
| Google (Gemini) | 46% | 59 |
| OpenAI | 39% | 31 |
| Anthropic (Claude) | 35% | 77 |
| xAI (Grok), Mistral | not shown yet | 1 each |
Don't read this as a ranking. The models didn't all forecast the same stocks on the same days, and a few dozen calls per model is nowhere near enough to separate skill from luck. The site doesn't show a rate at all until a model has five graded forecasts, which is why Grok and Mistral have none yet.
Pattern 1: they lean bullish
On average, the real outcome landed 3.1 percentage points below what the AI forecast. They expect stocks to go up more than stocks do.
It shows in the calls themselves. Of the 23 graded predictions in the 12 most recently graded forecasts, 21 said “up”. When the market falls, nearly every forecast is wrong at once.
The clearest example: on 6 October, Claude, GPT and Gemini all looked at nVent (NVT) at about $167 and all said it would rise over the next month, by 3% to 5%. A month later it closed at $98.38, down 41%. Three models, the same evidence, the same miss.
Pattern 2: the rare “down” calls have held up
Because the AIs so rarely say “down”, the times they do stand out, and so far they've mostly been right. On 4 November, with Micron (MU) at $855.10, Claude and ChatGPT both said it would fall over the next week. It closed at $811.30, down 5.1%. Gemini, Grok and Mistral said flat and missed (the full comparison).
Two or three examples aren't a finding. But it's the kind of thing a public record exists to test: if the pattern is real, it will still be there after 500 forecasts. If it isn't, that'll be visible too.
What this record can't tell you yet
- Whether any of these models is actually good at forecasting stocks. 171 calls over about two months, on roughly two dozen stocks, is a start, not an answer.
- Whether one model is better than another. See above.
- Anything about the stocks themselves. A wrong forecast says more about the forecaster than about the company.
Why publish it anyway
Because the alternative is what usually happens: an AI gives a confident answer about a stock, nobody checks it, and the misses quietly disappear. Here the misses stay up next to the hits, and the numbers above will keep changing as more forecasts are graded, in whichever direction they go.