How DeuceLab measures its own accuracy
Anyone can claim a model is accurate. Here is exactly how we test ours — and how you can verify it yourself.
Walk-forward, no hindsight
Every prediction is graded using only information available before the match. We never re-score old matches with today's smarter ratings — that would bake in hindsight and inflate the numbers. The accuracy you see is what the model actually achieved at the time.
The metrics we report
Pick accuracy (how often the named favourite wins), Brier score (how close probabilities are to outcomes), log loss, and a calibration grade (does a stated 60% really happen 60% of the time?). No single number tells the whole story, so we show all of them.
Stability over time
A model that is good for a month but drifts is not trustworthy. So we track the monthly Brier across the whole history — a flat line means no drift. That month-by-month record, across every league, is on the track-record page for anyone to check.
What we don't claim
We tested our model against real closing market odds and found no reliable edge. So we never show a value-bet badge. What we offer is a transparent, well-measured read on who is likely to win — nothing more, nothing dressed up.