Every few months a new tool claims it can "predict" a stock price. The honest answer is more useful than the hype: a bare point number is nearly impossible to hold anyone accountable for, but a published range with a stated confidence level is — and that's the difference this page is about.
A point prediction — "AAPL will hit $250 by Friday" — sounds precise but is almost never wrong in a way anyone can pin down. If the stock lands at $248, was the prediction right? $260? Nobody defined "right" in advance, so nobody has to admit "wrong" after the fact. That asymmetry is what makes point predictions such good marketing: they sound confident and are never falsified, because no one wrote down the rule that would falsify them.
A testable forecast has to commit to that rule before the outcome is known: "90% of the time, the real price will land inside this range." Now there is something to check. If the true rate is 60%, the forecaster was wrong, on the record, in public.
Instead of one number, a testable forecast publishes a range and a confidence level: a 90% interval that should contain the real future price about 90% of the time, across many tickers and many forecasts — not every single one. Some misses are expected and healthy; a forecaster whose 90% band is never wrong over any meaningful sample is not more accurate — either the band is too wide to be useful, or the misses are not being counted.
Quantustik publishes this number continuously, including the misses, on the live calibration page — coverage per horizon, per ticker, updated as the backtest window rolls forward. That page is the actual evidence for this article, not a screenshot.
Before trusting any "AI predicts stocks" claim — ours or anyone else's — ask four questions: (1) Is the forecast a range with a stated confidence level, or a bare number? (2) Is the coverage rate published somewhere you can check, including the misses, not just the hits? (3) Is it measured out-of-sample, on prices the model could not have seen while it was built? (4) Does the forecaster separate "the band was calibrated" from "the direction was right" — those are different claims and a page that conflates them is hiding the weaker one behind the stronger one.
The next two articles in this series go deeper on each half of that: what a 90% confidence interval actually promises, and why calibration and directional accuracy are not the same thing. Read them from the AI Lab hub.
This page is educational: it explains why interval forecasts are the testable unit, not a claim that any interval forecast — including Quantustik's — is profitable to trade on. A well-calibrated band can still surround a price that moves the wrong direction for your position. See the calibration-vs-directional-accuracy article for why those are two different questions.
Educational research only — not investment advice.