The Lab · Reading the market · updated 2026-10-22
Player props — the price knows his role before his stats do
We built a stats model for NFL player props and tested it on 592 FanDuel lines before trusting it. It lost to the price. Here's why, what a prop really costs, and how much a hot streak is worth.
A player prop is a bet on one player's numbers. Props carry a bigger fee than game bets, and the price already knows the player's current role. Our own stats model lost to it, so we rank props by the market's probability and show the stats beside it as context.
A player prop is a bet on one player's numbers instead of the game: Kelce over 5.5 catches, Ovechkin over 0.5 points, Jokić over 11.5 rebounds. They're the fastest-growing bet there is, partly because they feel like something a fan can know more about than a bookmaker. You watched the last three games. You know he's been getting the ball.
So did the price. Here's what we found when we tested that.
Same math, bigger fee
A prop is priced exactly like a moneyline: "over 5.5 at −120, under at +100" means the book thinks the over is a little more likely, and the two prices' implied chances add up past 100% — the book's fee. (How to read that.)
The difference is the size of the fee. Across 5,163 two-sided NFL props from the 2025 season, eight books' typical prop added up to between 6.0 and 7.5 points past 100% — 6.8 at the middle book, and 6.0 at FanDuel, the cheapest of the eight. A standard −110/−110 game line adds up to 4.8. Every prop starts with a bigger hole to climb out of than a game bet does — which is why the vig matters more here, and why shopping for the price matters more too.
We built a model. Then we tested it before trusting it.
Before Prop Cards went live we built a stats model for NFL props, the way most prop models work: each player's per-game average this season, leaned toward last season's while the sample is small; how much the opponent allows to his position; and the game script the betting lines imply. Then we wrote down one test, in advance, and ran it once: 592 FanDuel receptions props from Weeks 1–4 of 2025, priced 90 minutes before kickoff, using only what was known before each game.
| Our stats model | FanDuel's price, fee removed | |
|---|---|---|
| Brier score (lower is better) | 0.2653 | 0.2399 |
| Betting its side at FanDuel's price | −7.5% | |
| Betting only where it saw an edge | −11.8% | |
| Its 100 most confident calls | won 56.0%; the price needed 56.8% |
The market was the better forecast by more than four standard errors — not close, and not luck. Betting the model would have lost money every way we sliced it, and it lost most where it was most sure it had an edge.
It was also overconfident in both directions. When it gave a prop under 35%, those props came in 43% of the time. When it gave one over 59%, they came in 47.5%. A model that's right about direction but too sure of itself will find "edges" everywhere, and they'll be the model's mistakes.
Why it lost: the price knows his role
Reading the props one by one, the cause was plain. A season average is a record of who a player was. The price is set on who he is this week: the depth chart after Tuesday's injury report, the snap share since the trade, the backup who's starting Sunday. Averages catch up to a new role a few games late. The market prices it the day it changes.
So Prop Cards don't rank by our model. They rank by the market's own probability — every sportsbook on FanDuel's line, fee removed — and the stats show beside it as context. The model is still graded on every NFL prop it prices, forward since Week 6, in case it learns something the price doesn't. We'll say so if it does.
Hot streaks: worth a look, not a bet
Prop Cards mark a 🔥 when the bet's side has cleared tonight's line in 7 or more of the player's last 10 games. It's the first thing most bettors look at, so it's worth knowing what it's worth.
Ten games is a small sample. A prop that's a true coin flip clears its line 7+ times in 10 about 17% of the time — one in six, by luck alone. A true 60% shot does it about 38% of the time. Every night, across hundreds of props, plenty of streaks are nothing but that. And the line has usually moved already: a player who's been clearing 4.5 catches tends to be facing 5.5 by now.
Whether a streak means anything beyond that is a question with an answer, so we're getting it. On October 16, before any of its bets were graded, we registered a test: a streak of 7+ in 10 on a side the market gives under 60% — history saying more than the price does. A top-10 list made only of those props runs beside the normal one, and it replaces the normal one only if its picks beat FanDuel's price by two standard errors over at least 100 decided bets. Until then, it's context.
One-sided props: an estimated fee
"Anytime touchdown" and "anytime goal" are usually quoted one way only — yes, at a price, with no "no" to bet. With no other side, there's nothing to measure the fee against. We estimate it at 6.8 points, the middle book's typical fee on two-sided props, take that out, and mark these props one-sided. Then we check: if they win less often than the estimate says, the true fee is bigger than 6.8, and we'll have measured by how much.
Where the price is better
Because props carry a bigger fee, a better price is worth more. When we compared 74 identical props on October 14 — same player, same line — the Novig exchange paid more than FanDuel on 53 of them. Shopping is the one edge in props that requires no forecast at all.
What isn't an edge: one book's price beating the rest. In the 2025 history, FanDuel's price beat the other books' consensus on only 48 of 1,174 props, and betting those 48 lost 5.8%. A soft price usually means a slow book, not a smart one. We still flag those props and grade the flag forward; so far, it's a flag, not a reason.
For the numbers people: the test, in full
Pre-registered, one look. The model and the test were written down before the history was
scored: 2025 Weeks 1–4, player_receptions only, FanDuel's two-way price snapshotted about
90 minutes before kickoff, 64 games. Model inputs used only games before each one. No tuning
after the look beyond bug fixes.
The model. The player's catches per game this season, shrunk toward last season's with weight n/(n+4) for his n games; what the opponent allowed to his position group against the league average, shrunk toward 1.0 with weight g/(g+4); and his team's implied points from the spread and total against the slate's average, scaling the count by the square root of that ratio. Receptions as a negative binomial, its spread by position fit on 2024 only; a push on an integer line counts as neither side.
The comparison. Brier is paired prop by prop, model against FanDuel's de-vigged chance on the same side, so the z of +4.31 is the model's extra miss divided by its standard error across 592 pairs. The market itself was calibrated on the same props — within a tenth of a point overall (50.2% against 50.1%).
What the history can't say. One market, one season's first month, one sportsbook. The NBA, NHL and MLB props on Prop Cards have never had a stats model tested against them; they're ranked by the market alone.
What this doesn't tell you
That props can't be beaten. It tells you that this way of modeling them — season averages against the price — loses, and why. Someone who knows about the role change before the market does has something real. By the time it's on a stats page, the price has it.
How the forward record will land. Every Prop Cards call is frozen before its game and graded after: the market's favored side against what FanDuel's price needed, the stats model against the market, the 🔥 test, the one-sided estimate and the soft-price flag, each kept apart. They're young, and young records mean very little.
Anything about parlays of props. Props from one game move together, and the book prices that. That's its own article.
Model output is informational and entertainment content, not betting or financial advice. If you bet, bet what you can afford to lose.
Log what you actually bet — units and dollars — and see it graded beside a model that publishes every call it makes, wins and losses alike. Free, no card.
Start a free tracker