Sports
Research NBA player props across every book
Basketball props are minutes props. Everything else is a rate applied to a minutes estimate, and the minutes estimate is the part that changes without warning, usually an hour before tip-off.
Analytics only. quant+ never places bets or holds funds.
- cover line. Points, rebounds, assists or the combined number, as posted.
- alt lines. The rungs above and below it, which is where tail pricing lives.
- best price. Ranked against every venue, not against one book's own menu.
- consensus line. The anchor. A projection far from it needs an explanation.
- fair price. Post-margin probability, so the comparison is like for like.
- projection. Built from projected minutes and a shrunk per-minute rate.
- movement. Timed changes, because basketball news lands close to tip.
- book count. Market breadth. Two books is a different problem from twenty.
Field names and definitions only. No odds, projections or hit rates are hardcoded anywhere on this page.
Project the rotation before the player
In basketball, the per-minute rate is a reasonably stable property of a player and the minutes total is not. A guard who averages eight assists per game is really a player who distributes at a certain rate per minute while playing a certain number of minutes, and both halves of that sentence can change independently.
When a teammate returns from injury, the minutes redistribution is the whole event. The returning player gains, everyone who absorbed their role loses, and the resulting shift is often larger than any matchup effect the market has priced. A model that treats minutes as a slowly moving average will be behind the market on exactly the days that matter most.
There is a second-order reason to lead with minutes. Because the minutes estimate is itself uncertain, the honest output is a distribution with a width that reflects that uncertainty. A points projection built on a rotation guess and presented as a precise number is asserting more than the underlying estimate can support.
Where the minutes number comes from
- Demonstrated rotation: which player starts, which closes, and what the coach has actually done in comparable games.
- Availability of teammates at the same position, because minutes are a shared and finite resource.
- Expected competitiveness of the game, since the spread governs how likely a fourth-quarter truncation is.
- Rest situation, because a back-to-back changes both the rotation and the minutes distribution around it.
Blowouts delete the high side of a distribution
A counting distribution fitted to a season of games blends two different situations. In a close game, a starter plays into the final minutes and the distribution has a long right tail. In a game decided early, the same starter sits the fourth quarter and the tail is cut off.
If you fit one smooth distribution across both, the fitted curve will include a tail that only exists in the close games. On a main line near the middle of the distribution this produces a small error. On a high alternate rung, far out in the tail, it produces a large one, and it errs in the direction of making long-shot overs look more likely than they are.
The adjustment is not exotic, because the input is public. The pre-game spread gives an estimate of how competitive the game is expected to be, and the truncation can therefore be modelled before tip-off rather than noticed afterwards. It is one of the few places in prop research where a structural correction is available for free.
How truncation behaves
- A large lead reduces the minutes of the leading team's starters, cutting the high side of every counting distribution for those players.
- The effect is concentrated in the right tail, which is precisely where large alternate lines and high-side prices sit.
- A symmetric distribution fitted to a blowout-exposed player will overstate the probability of a very large game and understate the probability of a very small one.
- Blowout probability is derivable in advance from the spread, so the truncation adjustment is available before the game rather than inferred from it afterwards.
Each market has its own dominant driver
Treating points, rebounds, assists and combined markets as one modelling problem is how a prop tool ends up accurate on the main lines and wrong on everything else.
Points
Minutes multiplied by per-minute scoring, adjusted for the expected pace of the game and the defensive matchup. Scoring rate is the noisiest of the three core rates, so it warrants the most shrinkage, and the minutes term is where the projection actually lives or dies.
Rebounds
A function of minutes, positional matchup and shot-profile. Long misses produce long rebounds, so a game between two teams that take and miss a lot of three-pointers generates more rebound opportunities than the raw pace figure alone suggests. Opponent shot profile is a more informative input than opponent defensive rating.
Assists
Sensitive to role, not only to minutes. A guard who initiates every possession and one who plays the same minutes off the ball have different distributions. Assists also depend on teammates making shots, which introduces correlated noise on top of the player's own passing rate.
Combined markets
Points plus rebounds plus assists is the sum of three correlated counts. Adding means is straightforward; adding variances is not, because the components share a game environment and a minutes term. Treating the combined line as three independent props in a trench coat overstates the spread and misstates the tails.
Variance is a primary input, not a footnote
Rebounds and assists are overdispersed counts, and blowout truncation makes their distributions asymmetric. A method that reports a mean and stops has discarded the part of the answer the line depends on.
Rebounds and assists are non-negative integers, and their game-to-game variance is typically larger than their mean. A Poisson model forces those two quantities to be equal, which produces distributions that are too narrow and tails that are too thin. A negative binomial adds a dispersion parameter and matches the observed spread, and for lines away from the middle of the distribution the choice between the two changes the answer materially.
The minutes term introduces a second source of spread that a pure count model misses. Minutes are not fixed, they are estimated, and their own distribution propagates into the player's counting distribution. The cleanest way to handle this is a mixture: draw minutes, then draw the count conditional on those minutes, and read probabilities from the resulting compound distribution. The compound distribution is wider than the conditional one, and the width is real rather than a conservative fudge.
A third consideration is dependence across players. Two teammates' rebound totals are negatively correlated because they compete for the same available rebounds, while two teammates' assist totals may be positively correlated because they share a fast pace. Any combination of legs that ignores this is pricing a joint event as a product of marginals, which is the same error as treating a same-game parlay as independent and is just as damaging here.
- Shape over centre. On a main line, a slightly wrong spread does little damage. On an alternate rung three or four steps out, the shape of the tail is the entire calculation, which is why alternate markets are where a crude distribution is most clearly exposed.
- Compound, not single-stage. A count conditional on a known minutes total understates the true spread. The distribution over minutes has to be carried through into the final probability rather than collapsed to a point estimate early.
- Teammates compete and cooperate. Rebounds are a shared pool, so two teammates are substitutes for each other. Assists and scoring can move together when a team plays fast. Neither direction of dependence is captured by treating legs as independent.
What moves an NBA prop, and how fast
Basketball prop lines move later and faster than football ones, because the information that moves them arrives later.
| Driver | Why it matters here | Timing |
|---|---|---|
| Inactive list | Removes a player and redistributes their minutes and usage across the rotation. The second effect is often larger than the first. | Shortly before tip-off, sometimes within the hour |
| Starting lineup confirmation | Confirms who is on the floor and, for the replacement, confirms a role that was previously hypothetical. | Pre-game, occasionally later for a late replacement |
| Spread and total movement | A total move reprices every player on both teams at once, because pace is a common factor across all of them. | Throughout the day, accelerating near tip-off |
| Rest decision | A late scratch for rest removes a high-usage player and changes the distribution of minutes for several others simultaneously. | Often announced late, sometimes without prior indication |
| Sharp-side pressure | A prop number moving against public sentiment on a thin market is a signal that someone with a view has acted. | Continuous, most visible in the final hours |
A daily basketball prop routine
The last step is the one that separates a working process from a broken one, because basketball's information arrives late.
Project the rotation
Establish who is available, who starts, who closes, and how the minutes are likely to be distributed. Do this before opening any price.
Apply shrunk rates
Use per-minute rates pulled toward a positional baseline, not raw season averages. Rate stability is the justification for projecting forward at all.
Adjust for blowout exposure
Use the spread to estimate how competitive the game is likely to be, and truncate the high side of the distribution accordingly.
Carry the uncertainty through
Produce a compound distribution that includes the spread in the minutes estimate, then read probabilities for every rung off it.
Price every rung
Compare each alternate line separately. A ladder is a set of bets with separate margins, and one rung is frequently out of step with the rest.
Re-price at the news
When the inactive list and lineups arrive, re-derive rather than patching. A model that only runs in the morning has not seen the information the market has.
What this approach cannot do
It cannot tell you what the coach will do. Rotation decisions are made by people with incentives a model does not observe, and late-season games in particular can be shaped by considerations that have nothing to do with the basketball. Modelling the rotation more precisely than it is knowable is a way of manufacturing false confidence.
It cannot make a thin market liquid. Prop markets on less prominent players are easier to move and harder to bet in size, and an edge you can only get a small stake into is a small edge regardless of how it looks on a screen.
And it cannot rescue a stale input. The single most common failure in basketball prop research is not a modelling error at all, it is a projection that was correct when it was generated and was never regenerated when the availability picture changed. That is a process problem, and no amount of distributional sophistication compensates for it.
quant+ publishes the research surface, not the wager. For the team-level side of the same slate see the NBA betting model, and for the general method see the expected value guide.
Open the prop research tool · Scan for prop edges
Start from the rotation
Project the minutes, apply a rate you have shrunk toward a baseline, carry the uncertainty into a distribution, and price every rung against fair value.
Questions people ask
Why do minutes matter more than recent scoring for NBA props?
Because a player cannot score from the bench. Points, rebounds and assists are all a per-minute rate multiplied by time on the floor, and minutes are far less stable week to week than the rate is. Recent scoring averages implicitly assume the rotation has not changed, and in the NBA the rotation changes constantly through injuries, rest and coaching decisions.
How do blowouts affect player prop probabilities?
They cut off the high side of a starter's distribution. In a game decided early, starters sit the fourth quarter, so the large-outcome tail that exists in close games does not exist here. A distribution fitted across both kinds of game will overstate the probability of a big performance, and the error is largest on high alternate lines, which is where the biggest prices sit.
Should combined points, rebounds and assists markets be modelled separately?
Not separately in the sense of being unrelated, but not as three independent props either. The three components are positively correlated through shared minutes and a shared game environment, so adding the means is fine while adding the variances as if they were independent is not. The combined market needs the joint distribution, not three marginal ones multiplied together.
How late do NBA prop lines move?
Later than in most other sports, because the information that moves them arrives late. The inactive list and confirmed lineups land close to tip-off, and a late rest decision can change the minutes distribution for several players at once. Any workflow that generates projections once in the morning and never revisits them will be working from a picture the market has already moved past.
What does quant+ show for an NBA prop?
The cover line and its alternates, best price across books, the market consensus line, a fair price with the margin removed, an independent model projection, recent line movement with timestamps, and the number of books posting the market. quant+ does not place bets and has no account connection to any sportsbook.
Sources
- Research player propositions with Outlier's Prop Finder, help.outlier.bet
- Building an NBA betting model | Rithmm, rithmm.com
