NBA Player Projection Models: Intro & Test Data

Our NBA player props were introduced in late 2023 and quickly became one of our most popular models. We utilize a 3rd party's regressed Kalman-optimized outputs for each player's base box score projections, and then adjust them based on a consensus minutes played estimate for each player and on the game pace values generated by our team spreads model. In early 2024, we released a Version 2.0 that further enhanced both the raw projection accuracy and the probability conversion process, with the changes as follows:

• In Version 1, player-level consistency was calculated using a coefficient of variance estimate applied to the daily projection. This had the effect of underestimating upside/downside values and thus making the model overly confident in its probabilities and did not allow for varying levels of consistency in different box score items. In Version 2, all metrics are calculated independently for each market (see table below left)
• Before, all projections assumed a simple Gaussian distribution. While this was fine for high-volume markets like points, it was less than ideal for lower-total ones like blocks (see skew values in center table). Now, the distributions are customized for each market, and include both discrete and continuous methods.
• In Version 1, the pace adjustment was applied equally to each box score item. However, certain stats are more heavily affected by pace changes than others. For example, in the correlation matrices below right, you can see that PRA is much more heavily impacted by pace changes than steals+blocks are. Now, pace adjustments are optimized for each individual stat.

Points

Model Test Performance Summary

All Test Datapoints

Rebounds

Model Test Performance Summary

All Test Datapoints

Assists

Model Test Performance Summary

All Test Datapoints

3-Pointers Made

Model Test Performance Summary

All Test Datapoints

Steals

Model Test Performance Summary

All Test Datapoints

Turnovers

Model Test Performance Summary

All Test Datapoints

Steals + Blocks

Model Test Performance Summary

All Test Datapoints

First Basket Scorer

Model Test Performance Summary

All Test Datapoints

*No separate test data set for this market; set was built over course of season and will be added to in months ahead