PerkinsProps
Results · updated after every slate

The Scorecard

A model that is never checked against what happened is an opinion with decimal places. Every call the model makes gets graded off the real box score, and the grades live here whether they are good or not.

Backtest: Week 1

614 real props · graded against what happened
The honest result: the model does not beat the betting line. We re-ran the exact model this site uses against every Week 1 prop, using only what was known before kickoff, then checked each one against the final box score. Simply trusting the line — calling every prop a coin flip — scores 0.2500 (lower is better). Our raw model scores 0.2640, which is worse. Leaning on it lightly, the way the site does, scores 0.2502 — a tie with the line, inside the noise. So the leans on this site are not a demonstrated edge, and the page will say so until the record here shows otherwise.
Trust the line
0.2500
Every prop a coin flip.
Raw model
0.2640
Worse than the line.
Site (light lean)
0.2502
A tie, inside the noise.
Overs hit
51%
Lines really are set near the middle.

When the raw model sounded sure, it wasn't

perfectly calibratedclaimed 30%, hit 48% over 148 callsclaimed 43%, hit 47% over 105 callsclaimed 47%, hit 50% over 99 callsclaimed 52%, hit 57% over 94 callsclaimed 57%, hit 55% over 67 callsclaimed 68%, hit 54% over 101 calls0%claimed 100% →hit rate

Dots far from the diagonal are the problem. Props the raw model rated around 30% actually went over nearly half the time, and its 65% calls landed barely better than a coin. That overconfidence is why the site anchors to the market and only leans a tenth of the way toward its own view.

What the backtest did fix

perfectly sizedlow end of forecasthigh end →

Grey is before, green is after. Real results kept landing in the middle of our forecasts and rarely in the tails, which meant our spreads were too wide. Tightening them by a fifth evened that out. It doesn't change the headline chance on the posted line — the market pins that — but it makes the Lab's teaser and alternate-line numbers much more trustworthy, because those live in the tails.

Nothing graded yet

Results appear once this week's games finish — grading runs automatically each morning. Until then the model's calls are on the board and nobody, including us, knows how they will do.

The family

0 graded picks

No graded picks yet. Post one on the board and it will be settled automatically once the game finishes.