How accurate points projections are
We publish both the mean error of a projection and the top-3 hit count — always with the sample they were measured on, and measured so the model never sees the future.
The most common way to lie about accuracy is to test a model on the same races it learned from. That test always looks excellent and says nothing about the next round.
The honest test is called walk-forward: at every round the model only knows what was known before it, and projects the next one. That is how we measure.
On a typical race we are off by about 11 points per driver. Put simply: if two drivers land within 11 points of each other in the projection, treat them as equal — the difference is inside our error.
Over the last 7 races (rounds 5–11) the projected top three drivers matched the actual top three 3 times. The season's first 4 rounds are not in there: a model that predicts a race is fitted on the races before it, and at round two there is nothing to fit on.
Eleven points of error sounds like a lot, and it is an honest number: the spread of race outcomes is enormous, because a retirement, rain or a team call outweighs any amount of form. That is exactly why we show a range of outcomes rather than one figure.
We deliberately do not publish a share of "correct" lineups. A tool that looks more confident than it is does harm before a deadline: it pushes you to bet everything on one option where the decision should account for risk.
Common questions
What is walk-forward validation?
It is a test where, at every round, the model uses only the data available before it. It is fairer than testing on the whole season at once, and always produces more modest figures.
Why is the headline figure an error, not a hit rate?
Because a hit rate depends on what counts as a hit, and on a short sample it looks rounder than it is: adding one round moved ours from 50% to 43%. So we state hits as a count next to the sample they came from, and keep the mean error in points as the headline figure — it leaves no such room.