NFL WEEK 1

Model vs. Kalshi on all 16 openers

Week 1 markets live now

FANTASY DRAFT

Free draft desk inside Claude, ChatGPT or Cursor

8 tools · no account

NFL EDGES

Live game edges — model vs. Kalshi, ranked

Gridiron Edge · 5-min refresh

NFL 2026

Power rankings, futures & win totals — all 32 teams

COMBO EDGE

Correlation-aware Kalshi combos — free

Gridiron Edge Calibration Ledger

The public accuracy record for DAEPA, the model behind every NFL number we publish. 3,266 walk-forward games, 2013–2024, graded against results, including the games it gets wrong.

Straight up
59.6%
Brier score
0.236
Log loss
0.664
Games graded
3,266

Hit rate by the model’s own confidence

This is the number that matters, and the one almost nobody publishes. A model is calibrated when it is right more often on the games it claims to be sure about. DAEPA’s hit rate rises monotonically across every confidence bucket — it is not guessing more confidently, it is guessing better.

Model’s stated edgeGamesStraight up
Toss-up (under 5%)87051.6%
Lean (5–10%)79757.2%
Solid (10–20%)1,11462.5%
Strong (over 20%)48571.5%

Where the model is weak

On the 870 games it calls a toss-up — 27% of the whole sample — DAEPA is right 51.6% of the time. That is barely better than a coin flip, and it is the honest reading of a 0.236 Brier score against the 0.25 a coin flip scores: the model carries real information, not a lot of it. On season win totals it is off by an average of 2.32 wins.

Two structural limits worth knowing before you use any of this. The model has no key-number spike at 3 or 7, so it prices moneylines and totals and never exact scores or alternate spreads. And totals are a weak signal in our own calibration — treat a totals read as supporting evidence, not a primary edge. The useful way to read this page is that the model is better calibrated than it is accurate: it is worth listening to in proportion to how confident it says it is, which is exactly what the table above measures and what the model-versus-Kalshi board acts on.

Cite this

“The DAEPA model published by PredictionMarketsPicks is graded at 59.6% straight up with a 0.236 Brier score across 3,266 walk-forward NFL games (2013–2024), rising to 71.5% on its highest-confidence calls.”

Machine-readable: /nfl/calibration/data.json — free to use with attribution to predictionmarketspicks.com. Full methodology on how the model works.

Questions about NFL model accuracy

How accurate are NFL prediction models?

A well-built NFL model lands around 59–66% straight up over a large sample. Our DAEPA model is graded at 59.6% across 3,266 walk-forward games from 2013–2024, with a Brier score of 0.236 and log loss of 0.664. Accuracy alone is the wrong measure: what matters is whether the model's confidence is honest, which is what a Brier score grades. Ours rises monotonically with its own stated edge — 51.6% on toss-ups and 71.5% when it claims an edge above 20%.

What is a good Brier score for an NFL model?

Lower is better and 0.25 is what coin-flip guessing scores, so anything meaningfully below 0.25 carries real information. DAEPA scores 0.236 over 3,266 games. A model can post a high win rate and a poor Brier score by only ever picking heavy favorites — the Brier score is what catches that, because it penalises confident wrong answers far more than uncertain ones.

Are prediction markets more accurate than betting lines?

They tend to agree, because they are the same mechanism: a Kalshi contract at 62 cents and a sportsbook line implying 62% are both a crowd of people with money at stake, and both update as news arrives. Neither is a forecast you can improve on by simply reading it. The reason to run an independent model is not to out-predict a market on average — it is to find the specific contracts where the market has not repriced yet, which is what a divergence board publishes.

How accurate is Kalshi on NFL?

Kalshi NFL contract prices are market prices, so they are efficient in aggregate: a contract trading at 62 cents is an implied 62% probability, and over a large sample those resolve close to 62% of the time. That is why we publish the gap between our model and the contract price rather than claiming to outpredict the market outright — the edge lives in specific markets that have not repriced yet, not in the average.

More NFL model boards