BLOG / STATISTICS

A good 80% interval misses 20% of the time

The confidence band is the most ignored part of a forecast response — and the easiest to fake. Here is how we keep ours honest.

LR Lasse Rafn FOUNDER
18 JUN 2026
3 MIN READ

The decorative band

Every forecast you have ever seen comes with a band around the line. Almost nobody checks it. That is a shame, because the band is where a forecast admits what it does not know — and it is also where forecasting products cheat, usually without meaning to.

There are two ways to fake a confidence interval. Make it too narrow and it looks impressively precise while reality lands outside it week after week. Make it too wide and it covers everything — a band that contains 100% of outcomes at a claimed 80% sounds safe, but it is telling you nothing at all. Both failures produce a chart that looks fine.

An interval that never misses is not cautious. It is decorative.

Coverage cuts both ways

The first honest number is coverage: across many forecasts, what fraction of actual outcomes landed inside the band? For an 80% interval the right answer is 0.80. Not higher — higher means the band is wider than its label claims, and every decision sized to it carries hidden slack.

But coverage alone can be gamed by width, so it needs a partner that cannot: we score intervals with a rule that charges for width and fines every miss by how far outside the band it landed. Under that rule the winning strategy is the narrowest band that still covers at the stated rate — which is exactly what you want from one.

How we check ourselves

The method is backtesting. Take a real series, cut off the tail, forecast the held-out window blind, and compare — the band against what actually happened, at the level that was requested. Then repeat across a suite of series with very different personalities: smooth, seasonal, intermittent, bounded.

We run this on every change to the forecasting engines. Not because charts are ugly, but because point accuracy and interval quality move independently: a change can leave every point forecast untouched and quietly transform the band, or the reverse. Eyeballing a chart catches neither.

What to do with the band

The band is not chart furniture — it is the operational half of the response.

ONE PERIOD OF A RESPONSE · CONFIDENCE 0.80
{ "period": 1, "date": "2026-07-01", "forecast": 4210, "lower": 3780, "upper": 4650 }

If reality lands above 4,650 more than one month in ten, that is not bad luck — that is a bug on our side, and it is exactly what our backtests exist to catch.

MORE FROM THE BLOG ALL POSTS →

See who wins on your data.

Start with 1,000 free API calls.
No credit card required.