How to Predict NBA Player Points Props
The method our model uses, step by step — and the graded record that says points is the hardest prop in the sport to beat.
The short version
A points prop is a question about a distribution: given everything we know about a player and tonight's game, what is the chance he scores more than the number the book posted? The method below produces that probability from public box scores. It is the same method that runs on our board every game day, and the record at the bottom of the page is what it produced last season: 462-451 (50.6%) at the play tier (one change to the method, described in step 3, arrived a month before the end of that record). That is not a number to bet blind, and this page does not pretend otherwise. It is a number to understand.
The method, in four steps
1. Measure recent form, not the season average
The book line is anchored on season averages. The box scores move faster. For every player the model keeps a rolling window of his last eight games and pulls two rates from it: minutes per game and points per minute. Splitting scoring into minutes and rate matters — a player at 34 minutes and 0.65 points a minute is a different prop from one at 28 minutes and 0.8, even though both average about 22. It also keeps the spread of his points across those eight games, which becomes the width of the distribution later.
Games under 20 minutes are dropped from the window, and a player averaging under 15 minutes is not projected at all; those lines are not worth modelling.
2. Project a mean with a regression
An XGBoost regressor, trained on the current season's player box scores, takes seven inputs — the two rate features, rest-day difference, whether it is the second night of a back-to-back, home or away, the opponent's defensive rating and the opponent's pace — and outputs a projected points total for tonight. Seven features is deliberately few. The point of the regression is to adjust the rate for the matchup without inventing signal from noise; a deep model on player-game data memorises last month.
3. Put a distribution around it
A projected 24.1 is not a pick. The model wraps the projection in a normal distribution whose width is the player's recent standard deviation — with two corrections that the graded record forced on us. The width is floored at 60% of his full-season deviation (an eight-game cold streak compresses the window and made the model far too sure of itself), and it is then multiplied by 1.5, a calibration added on May 14, 2026 after a re-grade showed the 90%+ predictions were hitting 45.6%. Wider distributions mean fewer plays, and the plays that remain are ones the model is actually entitled to. Note the date: most of the record below was produced before that multiplier existed.
4. Score the line with a z-score
The probability of the over is the area of that distribution above the line. Half a point is added to the line first so “over 24.5” means 25 or more, and the same calculation is repeated at every alternate line from 5.5 to 40.5 in one-point steps — that is where the alt-line probabilities on the board come from. The “safe line” is the highest line the model puts at 85% or better. A play is any side at 65% or more at the primary line, which is also the grader's threshold, so the record below counts exactly what was badged.
What it hit, graded
Every play above the threshold from March 12, 2026 to June 13, 2026 (79 game days, playoffs included) was settled against the official box score the next morning. Injured players and anyone who did not play are excluded, not counted as losses. The break-even at -110 is 52.4%.
| Model probability | Record | Hit rate |
|---|---|---|
| 65–70% | 147-150 | 49.5% |
| 70–80% | 184-178 | 50.8% |
| 80–90% | 108-90 | 54.5% |
| 90%+ | 26-35 | 42.6% |
| All plays (65%+) | 462-451 | 50.6% |
Tier rows are a replay of the grader over the archived pick files (465-453 in the replay; the tracker's per-day count is 462-451); the total is the tracker on the NBA performance page. Green clears -110; amber does not.
Read the table before the story. The hit rate is flat. A 65% prediction and an 85% prediction hit at about the same rate; the 80-90% row is the only one over the -110 line and it is 198 plays, not a trend; the 90%+ bucket did worse than any of them. Two caveats on the tiers. First, 779 of these plays were made before the 1.5× calibration of May 14; the model with the multiplier has 69-70 (49.6%) since, which is the same coin flip on a smaller sample, so the multiplier fixed the overconfidence without producing an edge. Second, by side, unders went 167-153 (52.2%) and overs 298-300 (49.8%): a small gap in the direction the market's blowout risk points, and within noise.
Compare the other over/under prop we publish with the same code: assists went 57.2% over the same window. Same regression, same distribution, same threshold, different market. The difference is the market, not the model. (Our rebounds board is a different, simpler thing — a projection against a self-set line, not a graded market — and is not a comparison.)
Why points is the hardest prop to beat
Three reasons, and they compound.
The line already knows what the model knows
Points is the most heavily bet player prop in the sport. The books put their sharpest pricing on it, and by tip-off the line carries minutes, pace, the matchup and the injury report. Everything in step 1 and step 2 is public. A model built on box scores can find the line's mistakes in a market nobody prices carefully; in this one it mostly re-derives the line.
Scoring is the noisiest thing a player does
Rebounds follow minutes and position. Assists follow role. Points follow whether shots go in, whether the player picks up two early fouls, and whether the game is close in the fourth quarter. A 24-point scorer's eight-game range is routinely 14 to 34. When the distribution is that wide, most lines sit near its middle and most probabilities sit near 50% no matter how they are dressed up.
Overs die together
The over side of a points prop is priced on a full workload. A twelve-point lead at the end of the third quarter takes the starters off the floor, and every over on that team dies in the same minute. The model sees the opponent's defensive rating; it does not see that the game will be over by the fourth. That is why unders graded a shade better and why the projected competitiveness of the game matters more than any player feature.
How to actually use a points projection
Use the distribution, not the direction
The over/under call at the primary line is the least informative thing on the card. The alt lines are the model's real output: if a player is 62% at 22.5 but 84% at 18.5, the second number tells you where his floor sits, which is what a same-game parlay leg or a ladder bet needs. The last-10 bars under Details show whether he cleared each line by playing well or by playing 40 minutes.
Check minutes and the injury report yourself
The model's minutes feature is an eight-game average. It cannot know that the second scorer was ruled out 90 minutes before tip-off and the primary option's usage just jumped. That is the single most exploitable situation in the market and it lives in the injury report, not in a model.
Go live
Once the game starts the pre-game line is stale and the box score is not. Our live tracker re-projects each player's final points from his actual minutes and rate at each checkpoint — that is where a points model has information the closing line did not, and it is the tool we point points bettors at.
Tonight's points board
Every player on the slate with his line, the model's over/under split, alt lines and last-10 bars. The top five are free; the full board is for subscribers. First board of the season: opening night, October 20. Every play is graded the next morning and added to the record on this page.
Frequently asked questions
How do you predict NBA player points props?
Four steps: a player's recent form (rolling 8-game minutes and points per minute, and the spread of his points), an XGBoost regression that turns that form plus rest, back-to-back, home/away and the opponent's defensive rating and pace into a projected points mean, a normal distribution around that mean using his recent volatility, and a z-score against the book line that gives the probability of the over. A play is any side at 65% or more.
What hit rate does your NBA points model have?
462-451 (50.6%) over 913 plays at the 65% tier, graded March 12 to June 13, 2026. That does not beat the 52.4% a -110 price needs, and it is flat at every confidence level, so we say on every points page that the projection is context, not a bet.
Why is points the hardest NBA prop to beat?
It is the most heavily bet prop in the sport, so the line already carries minutes, pace and matchup, and scoring is the noisiest stat a player produces — shot volume, foul trouble and blowouts move it by ten points either way. A model built on public box scores adds little the book does not already know.
Where does a points model add value, then?
In the distribution rather than the direction: the alt-line probabilities show where a line sits inside a player's real range, unders have graded slightly better than overs, and the live tracker re-projects from the box score during the game, which is where the pre-game line stops being informative.
What is a z-score in a player prop?
The distance between the line and the projected mean, measured in standard deviations of the player's recent scoring. A z-score of zero is a coin flip; a line one standard deviation below the projection gives roughly an 84% over. We add half a point to the line first so that 'over 24.5' means 25 or more.
The week ahead, once a week
Last week's graded record — wins and losses — the plays for the week ahead, and a subscriber discount code. One email, no daily noise, unsubscribe in a click.