← All projects

October 26, 2026

Building an NBA assists and threes model

A projection model that prices every player inside his team's full context, tested against several seasons of real lines before a single live bet.

NBA / Player props
Live trackerNBA Player PropsFull tracker →

Record

394-357-26

Units

+28.06u

ROI

+3.7%

Most prop models project a player in isolation. In basketball that breaks the moment a starter sits: his assists don't vanish, they move to teammates. This model projects every player in context, so a whole roster's numbers fit together the way a real game does.

Architecture

  1. InputRosters, availability, historyAs of tonight
  2. ProjectEvery player, in team contextConsistent across the roster
  3. PriceA full distribution for every lineNot just a point estimate
  4. SelectOnly the clearest edgesAt the best available price
01

Availability

Who's playing tonight and in what role drives everything downstream. When the injury report changes, every teammate's projection updates with it, not just the missing player's.

02

Team context

Players are never projected one at a time. A team's production is shared, so one player's bigger night has to come from somewhere, and the model keeps that accounting consistent.

03

Distributions

Props live in the tails, so a projected average isn't enough. Every player gets a full probability distribution for each stat, and every line, including pushes, is priced from it.

04

Selection

Most of the board is priced close to fair. The model only plays lines where its edge over the best available price clears a fixed bar, and it skips players whose current role the data can't describe yet.

The backtest

The rules were written down before any test season was scored. Earlier seasons were used to build and tune the model, and the 2025-26 season was held out and scored once.

Season Bets ROI 95% range
2025-26 (held out) 2,908 +5.4% +1.8% to +9.2%
2024-25 (out of sample) 4,106 +2.4% −0.4% to +5.5%

Ranges come from resampling whole betting nights, since bets on the same night share the same news.

The audit

A result like that deserves suspicion, so the backtest was audited before going live.

  • No leakage. Every input was checked for information from after tip.
  • Placebo. Randomized versions of the model never once beat the real result.
  • Price sensitivity. The edge is thin and depends on getting a good number. Line shopping is part of the strategy, not an afterthought.
  • Concentration. A meaningful share of profit came from a small group of players. That’s normal for props, and it’s a reason to stake small.

Live

Every play is logged before tip and graded from the box score, 1 unit each, on the model page. A bet posted on more than one card counts once. A few hundred live bets is still too few to separate skill from luck, so the backtest is the stronger evidence until the season gets deeper.