How accurate are the sims?
Tracking since Nov 15, 2026 · 11,756 props settled
NFL sim accuracy at a glance
-
Props settled
11,756
Graded over or under from the box score. Pushes and players who sat out are left out.
-
Brier score, sim
0.201
Lower is better. Always guessing 50% scores 0.250.
-
Brier score, market
0.192
On the 6,441 props with a market no-vig price. The sim scored 0.198 on the same props.
Predicted vs actual
Props grouped by the sim's chance of the over. A well-calibrated sim's overs land at about the rate it predicted.
| Sim chance of the over | Props | Sim predicted | Over won |
|---|---|---|---|
| 0–10% | 2,489 | 4% | 10% |
| 10–20% | 1,343 | 15% | 24% |
| 20–30% | 1,392 | 25% | 35% |
| 30–40% | 1,619 | 35% | 41% |
| 40–50% | 1,701 | 45% | 48% |
| 50–60% | 1,391 | 54% | 53% |
| 60–70% | 771 | 64% | 61% |
| 70–80% | 525 | 75% | 72% |
| 80–90% | 362 | 85% | 78% |
| 90–100% | 163 | Collecting, 163 so far | |
A bucket shows its rates once it holds 200 settled props.
How to read this
Take every prop where the sim gave the over a 60–70% chance. If the sim is honest, about 65% of those overs should have won. The table puts the two numbers side by side for each bucket, so you can see where the sim runs hot (it predicted more overs than happened) or cold (fewer).
The Brier score rolls that into one number: the average squared gap between the sim's probability and what happened (1 for an over, 0 for an under). Lower is better. Saying 50% on every prop scores 0.250, so a lower score means the sim knows more than a coin flip. Log loss is the same idea but punishes a confident miss much harder.
The market column scores the no-vig price the betting market itself implied at the same moment. That is the bar to beat. Where the sim scores lower than the market on the same props, it is adding something the prices did not already know; where it scores higher, the market was sharper.
Method and limits
Every 10 minutes, for each game starting within the next half hour, the sim's latest probability for each prop a venue is quoting is saved with the best price each way and the market's no-vig probability. The last save before the game starts is the one graded. After the game, the player's stat comes from the same box scores the hit rates use, and the over is graded against the line.
Each prop counts once, graded on the over. A push (the stat lands on the line) and a player who did not play are left out. The sim probability is the sim's chance of the over with a push refunded. Log loss so far: 0.592 for the sim, 0.560 for the market on its 6,441 props (sim 0.584 on the same props).
Limits: the sample is small early on and one week of games can swing it, so nothing is scored until 200 props settle and no bucket is shown until it has 200 of its own. Props from the same game move together, so the effective sample is smaller than the count. Late news after the last sim run is in the price but not the sim. This is a record of a model, not a promise about the next game.
Sim accuracy FAQ
How accurate are BeatVig's prop sims?
Across 11,756 settled NFL props since November 15, 2026, the sim's Brier score is 0.201, where 0.250 is a coin flip and lower is better.
What is a Brier score?
The average squared gap between a predicted probability and the result, counting an over as 1 and an under as 0. Lower is better, and guessing 50% on every prop scores 0.250.
Which props are counted?
Every NFL and NBA prop the sim priced in the half hour before its game started, graded on the over from the box score. Pushes and players who did not play are left out.
Why are some buckets hidden?
A bucket shows its predicted and actual rates only once it holds 200 settled props. Below that, a few results can swing the rate far enough to mislead, so only the count is shown.