Calibration
How accurate has Horizon actually been?
Most forecasting tools ask you to take accuracy on faith. Horizon grades itself every time you confirm a real balance, and this page is where the grades go — including the ones we'd rather not print.
Across everyone
We have not published a fleet accuracy figure, because we do not yet have one. Nobody has been running Horizon long enough to grade honestly.
Refresh histories live with each person, not in a pool we quietly mine. A fleet figure will only ever be built from members who opt in, and it will carry its sample size in the same breath as its number.
What counts as a grade
When you confirm your balances, Horizon takes the balance you confirmed last time, applies every scheduled checking item in between, and compares that projection to what you actually have. The gap is the grade. Your very first confirmation has nothing to compare against, so it isn't counted — it isn't scored as a perfect one either.
What we deliberately don't count
Demo forecasts and sample lives never enter the numbers, no matter how good they look. Savings isn't graded, because most savings movement is a transfer Horizon can't verify from one side. We'd rather be right about one balance than vague about two.
Two rules we hold ourselves to
Dollar errors round up: if we missed by $47.10, we print $48. Accuracy shares round down: if four of six refreshes landed inside $100, we print 66%, not 67%. Every rule here bends against us, on purpose.
We publish the worst miss too
A median is easy to like. The largest single miss is the number that tells you what a bad week looks like, so it sits next to the median rather than in a footnote.
Why the sample size is always on screen
Three good grades is an anecdote. Horizon won't state an accuracy figure for you until it has 5, and won't publish one across everyone until it has 100. An accuracy claim without a sample size is marketing.
The engine behind these numbers is written up in how Horizon forecasts.