Estimating Impact, Ability, and Potential in the NBA — Introducing Three New Player Evaluation Metrics
Introduction
This post introduces three player rating metrics that I developed, which will serve as the foundation for future analysis on this Substack. The goal here is to explain how they’re constructed and how to interpret them.
Season Rating (SR): a measure of a player’s impact within a single season
Current Ability (CA): a smoothed estimate of underlying ability, taking into account noise, sample size, and age
Potential Ability (PA): a projection of that ability to the player’s prime age
You can view the follow-up post here, which shows how these ratings are used in roster construction and trade analysis, and explains the player archetypes referenced throughout this post.
Season Rating (SR)
I designed Season Rating to do two things: measure season-level impact, and provide the foundation for Current Ability (CA) and Potential Ability (PA). CA is the backward-looking step. It takes the season-to-season SR signals and estimates a player’s underlying ability after accounting for noise and sample size. PA is the forward-looking step. It projects that estimated ability forward to a player’s prime age, using the player’s current trajectory and the uncertainty implied by the smoothing model.
SR is constructed from two separate models: one for offense and one for defense. Each model is trained on a blend of custom metrics designed to reflect a player’s broader and more complete impact, which capture elements that box stats tend to miss or undervalue (especially on defense). One example is XPP (Expected Points Prevented), a matchup-based metric that estimates how many points a defender prevented relative to expectation against the opponents they actually guarded, benchmarked to league performance against those same opponents. These features are used to predict luck-adjusted O-RAPM and D-RAPM, respectively (you can read about luck-adjusted RAPM here).
Offensive Season Rating (O-SR) combines a base offensive impact estimate with the player’s primary and secondary offensive archetypes. The base model is meant to capture what actually happened on the floor in terms of offensive impact, and the archetype information is used to refine that estimate in a way that is consistent with the role the player plays. Because offensive archetypes already define how a player generates offense, the model does not need separate archetype-specific slopes. Instead, archetypes capture the idea that different offensive roles carry meaningful value beyond the base impact estimate, while the base impact model remains the main driver of O-SR. A spot-up shooter and a shot creator are asked to do very different things, especially in terms of shot difficulty and decision-making burden, and O-SR is built to account for that.
Defensive Season Rating (D-SR) utilizes on-court position to adjust how defensive features translate to impact. What drives defensive value varies significantly by role and responsibilities. A guard, for example, can earn more credit for having a high matchup difficulty rating, as this indicates they are trusted to guard the opponent’s top perimeter threats. A center earns more credit for rim protection events such as blocks, which are more central to defensive value in that position.
Current Ability (CA)
Current Ability takes Season Rating and smooths it across a player’s career so far in order to answer one core question: “What is the best estimate of a player’s true level right now, given what has been observed across multiple seasons and how reliable each season is?”
Statistically, CA is estimated using a state-space smoothing model. Each season’s SR serves as an observed data point (what we actually measured on the court), while CA represents the underlying ability level that the model infers from those observations. Each season contributes evidence, but the model assigns more or less weight based on reliability. Seasons with larger sample sizes (more games and minutes) are treated as more trustworthy, while smaller or partial seasons are viewed as noisier and are regressed more heavily toward the player’s broader track record. This structure allows CA to resist overreacting to seasons shortened by injuries or outlier seasons that do not reflect a true change in ability.
The smoothing model also estimates a trajectory, not just a level. In addition to a player’s current ability level, it tracks whether ability is trending up or down and how uncertain that estimate is. That uncertainty matters because it determines how much CA should move when a new season arrives. When the recent signal is consistent and high-sample, CA updates more decisively; when it is volatile or low-sample, CA updates more cautiously.
CA is age-aware. Rather than assuming ability evolves at the same rate at every age, the smoothing step uses an empirical age curve built from historical year-over-year SR changes. The curve encodes how much real movement is plausible at different ages, with multipliers normalized around prime age, because historically, large upward jumps are more common at young ages and large declines are more common at older ages. In practice, this gives the model more room to recognize genuine growth early in a career and genuine decline later in a career, while treating prime-age players as relatively stable unless the evidence is strong. The figure below shows the growth and decline multipliers by age:
Potential Ability (PA)
Potential Ability is the forward-looking counterpart to Current Ability. CA represents the best estimate of a player’s true level right now, PA takes that estimate and projects it forward to the player’s prime age.
Statistically, PA starts from the same state-space model used to estimate CA. At the end of the season, the model produces a posterior estimate of a player’s underlying ability and trajectory, along with uncertainty around both. PA then projects that estimate forward to prime age without adding any new observations. It is a forecast implied by the smoothing model itself, not a separate model trained on future outcomes.
PA is reported as a single number: a higher-end projection of what a player could look like at prime age. It is meant to reflect a favorable but still realistic outcome based on the player’s current level and recent direction of improvement. As a result, two players with similar CA can still have different PA if one has a stronger upward trajectory.
The age curve matters here too. The same growth and decline multipliers that shape how CA evolves by age also shape how much movement and uncertainty the model allows when forecasting forward. Younger players are granted more upward flexibility because larger true improvements are more common early in careers, while prime-age profiles are treated as more stable unless the evidence is strong. For players already at or past prime age, PA is clamped to CA rather than projecting further growth. The goal is to represent prime-level ability, not a second peak beyond it.
Early in the season, the metrics can be more volatile and will stabilize as more games are played. PA tends to fluctuate the most.
SR, CA, and PA are initially calculated on a per-100-possessions basis, then scaled to a familiar overall rating format where the league average is 75.
You can find the full tables and the latest updates in the NBA Insights Ratings Database.
This is very cool!
I love reading things like this. All of us on the analytics side have our own "secret sauce" when it comes to projections, and how to interpret metrics across time. Good stuff!