When Quantustik shows a 90% confidence band on a forecast, that number means one specific, testable thing — and a few things it does not mean, which are the ones people usually assume. Read this before you read a band width as a difficulty score.
A 90% confidence interval promises coverage: if you ran this forecast many times, on many tickers, over many periods, about 90% of the real, realized future prices would land inside the published band. The other roughly 10% will land outside it — by design, not by mistake. A forecaster whose 90% band is never missed over any meaningful sample is not more accurate; either the band is too wide to be useful, or the reported coverage is not being measured honestly.
One note for the statistically careful: a band around a future price is, strictly speaking, a prediction interval — "confidence interval" is used here in the everyday sense the industry uses it. The testable promise is the same either way: stated coverage, checked against realized outcomes.
This is the number tracked on the live calibration page, continuously, across the committed backtest set. It is published as one row per quarterly forecast start date, not as a single figure: measured against the model as served, coverage ran 78–95% at the 3-month horizon, 71–92% at 6 months and 86–96% at 1 year. The spread comes from nothing but when the forecast started, which is why an average of those rows would mislead rather than summarise. Individual tickers vary more still. Check the live page for the current numbers; this article will not repeat a snapshot that drifts.
Coverage is not the same as being right. A forecast "hits" in the coverage sense whenever the real price falls anywhere inside the band — including near an edge, including when the band is wide enough to cover almost any plausible outcome. That's a much weaker claim than "the model correctly called what happened", and treating a covered outcome as a correct prediction is the most common misreading of a confidence interval.
It also is not a promise about direction. A 90% band can fully contain the real price while still being centered above or below it — the model's own point estimate can be on the wrong side of where the price ends up even when the band as a whole covers the outcome.
A calibration figure that only shows wins is not a calibration figure — it is a highlight reel. The 10% of cases the interval is expected to miss are as important to publish as the 90% it covers, because that miss rate is the actual test of whether the number "90%" was honest in the first place.
Some individual tickers miss far more than 10% of the time — a structural break is one cause, but large, stable names score poorly too, and they do so even while the aggregate figure looks well-calibrated. That per-ticker detail is on the calibration page and is the subject of the where-our-model-fails article in this series.
A 90% confidence interval is a statement about coverage, not about whether a trade based on the forecast would have made money. A perfectly calibrated 90% band can still be centered on the wrong direction, or wide enough that it covers both a profitable and an unprofitable outcome. Calibration and directional accuracy are separate questions — see the calibration-vs-directional-accuracy article for why we do not publish a win-rate.
Educational research only — not investment advice.