Report card — how our numbers actually did

Every projection and every prop side on this site is published before kickoff and graded afterwards against what actually happened. Nothing here is a backtest. When the numbers are bad, they stay up.

0projection correlation
0 players ≥6 pts · MAE null

+0.16¢closing-line value, anchored
4135 sides tracked to kickoff

Every graded week

Week	players	r	MAE	sides	hit	our Brier	market Brier	ROI anchored	ROI model-only
2026 wk 9	0	0	null	–	–	–	–	–	–
2026 wk 8	27	0.499	4.95	701	50.9%	0.193	0.1951	0.3%	-7.1%
2026 wk 7	86	0.346	6.28	583	51.1%	0.2329	0.2146	-2.8%	-4.7%
2026 wk 6	17	0.512	5.65	787	50.6%	0.1811	0.1813	-1.3%	-13.4%
2026 wk 5	0	0	null	734	51.9%	0.1859	0.2136	6.5%	6.4%
2026 wk 4	97	0.245	6.48	653	50.5%	0.2122	0.2284	-3.9%	-6.5%
2026 wk 3	0	0	null	667	50.1%	0.2418	0.2475	-4.2%	-6.1%
2026 wk 2	122	0.509	6.13	162	51.2%	0.2419	0.2465	10.7%	-52.6%

A Brier score below the market's means our probabilities were sharper than the prices they were quoted against. One week is an anecdote; the table is here so you can watch whether it holds.

Closing-line value

Outcome grading needs hundreds of settled contracts before it means anything. Closing-line value needs far fewer: if the sides we flag get more expensive before kickoff, we were finding real mispricing — whether or not any single contract hit. Measured on the mid, because Kalshi's books tighten as a game approaches (5.5¢ spread at open, 2.1¢ near kickoff) and comparing executable prices would manufacture a loss on every side, both directions.

Population	sides	beat the close	mean CLV
Pinnacle-anchored	4135	46.4%	+0.16¢
Model alone, no anchor	6075	39.2%	-0.25¢
Half-spread paid to cross	—	—	+2.62¢

Read it plainly: the anchored signal is real but small, and it is smaller than the cost of crossing the spread. That is why the props page leads with a resting-order price instead of telling you to buy at the ask.

What we'd tell you not to trust yet

Model-only rows — the ones with no Pinnacle line to anchor against — have graded badly on both measures. They are shown because hiding them would make this page a sales pitch instead of a record, but the anchored filter is on by default for a reason. Sample sizes here are still small enough that a good week and a real edge look identical. Check back.

dfsroto is independent software and data. Not affiliated with, endorsed by, or sponsored by DraftKings, Kalshi, Pinnacle, ESPN or the NFL; their names identify the data only. Probabilities and projections are information, not advice. 18+. Terms · Privacy · Data policy