Regularized Adjusted Plus Minus (RAPM)

• RAPM is a regression-based framework that quantifies player impact by adjusting for teammates, opponents, and contextual factors.
• It replaces OLS with ridge or Bayesian regularization to stabilize estimates amidst multicollinearity and sparse data.
• RAPM has been applied across sports like basketball, hockey, soccer, CS:GO, and Formula 1, enhancing predictive accuracy.

Regularized Adjusted Plus Minus (RAPM) is a regression-based framework for estimating individual impact while adjusting for teammates, opponents, and deployment context, with regularization used to stabilize estimates under severe multicollinearity and sparse samples. In its canonical form, RAPM keeps the adjusted plus-minus design of APM and replaces OLS with ridge regularization; in Bayesian notation, the same estimator is a Gaussian linear model with a zero-centered Gaussian prior on player effects (Petridis et al., 21 Jan 2026, Matano et al., 2018). Although RAPM is most closely associated with basketball, the framework has been adapted to hockey, soccer, Counter-Strike: Global Offensive, and Formula 1, with observation units ranging from possessions and stints to maps and race entries, and with linear, logistic, multinomial, elastic-net, and hierarchical Bayesian variants (Macdonald, 2012, Bajons et al., 2024, Xu et al., 2024, Rane, 31 Jul 2025).

1. Core statistical formulation

RAPM inherits the central APM idea: model an outcome as the sum of participant effects, conditional on who is simultaneously present. In a standard player-level formulation, the outcome for observation $i$ is written as

$y_i = x_i^\top \beta + \varepsilon_i,$

and the regularized estimator solves

$$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$

This is the formulation explicitly used in soccer APM with FIFA-augmented priors and described as standard ridge APM or RAPM (Matano et al., 2018).

In basketball possession models with separate offense and defense coefficients, the design is often expanded so that each player has an offensive and a defensive term: $$y_i = \gamma_0 + \sum_{j=1}^{m} \gamma_{j,\text{off}} x_{i,j,\text{off}} + \sum_{j=1}^{m} \gamma_{j,\text{def}} x_{i,j,\text{def}} + \varepsilon_i.$$ Here, $\gamma_0$ is an intercept representing league-average points per possession, $\gamma_{p,\text{off}}$ is the offensive RAPM contribution of player $p$, and $\gamma_{p,\text{def}}$ is the defensive RAPM contribution of player $p$ (Petridis et al., 21 Jan 2026).

The Bayesian interpretation is equally standard. Ridge RAPM is equivalent to

$$y \mid \beta \sim \mathcal{N}(X\beta,\sigma^2 I), \qquad \beta \sim \mathcal{N}(0,\tau^2 I),$$

with $y_i = x_i^\top \beta + \varepsilon_i,$0 (Matano et al., 2018). This equivalence is important because it makes clear that RAPM is not merely a computational trick for ill-conditioned regressions; it is a shrinkage model with an explicit prior interpretation.

The “adjusted” component comes from the design matrix. In plus/minus formulations with a single coefficient per player, entries are typically $y_i = x_i^\top \beta + \varepsilon_i,$1 for the focal team, $y_i = x_i^\top \beta + \varepsilon_i,$2 for the opponent, and $y_i = x_i^\top \beta + \varepsilon_i,$3 otherwise. In split offense-defense formulations, offensive and defensive participation are encoded in separate blocks. In both cases, coefficients are identified jointly, so a player’s effect is estimated conditional on the teammates and opponents that appear in the same row (Xu et al., 2024, Macdonald, 2012).

2. Observation units, design choices, and outcomes

RAPM is not tied to a single sampling unit or outcome scale. What remains invariant is the regression-on-participation structure plus shrinkage; what changes across domains is the unit of observation and the response variable.

Domain	Observation unit	Response
NBA basketball	Possession or stint	Points per possession, offensive rating, or net differential (Petridis et al., 21 Jan 2026, Damoulaki et al., 2024, Jacobs, 22 May 2026)
NHL hockey	Shift segment	Goals, shots, Fenwick, or Corsi per 60 (Macdonald, 2012, Macdonald, 2010)
Soccer	No-substitution segment or possession sequence	Goal differential or possession ends in a goal (Matano et al., 2018, Bajons et al., 2024)
CS:GO	One map	Round differential or win/loss (Xu et al., 2024)
Formula 1	Driver-constructor race entry	Weighted finishing rank or top-$y_i = x_i^\top \beta + \varepsilon_i,$4 indicator (Rane, 31 Jul 2025)

This variation is methodologically significant. In hockey, RAPM has been estimated on shift segments with outcomes measured as goals, shots, Fenwick, and Corsi per 60, with zone-start indicators and separate even-strength and special-teams models (Macdonald, 2012). In soccer, one branch uses no-substitution match segments and segment goal differential, while another uses possession sequences defined as consecutive on-ball events and models whether the possession ends in a goal via logistic regression (Matano et al., 2018, Bajons et al., 2024). In CS:GO, the unit is a map, the linear response is round differential, and logistic variants replace round differential with win/loss (Xu et al., 2024). In Formula 1, the row contains exactly one driver and one constructor, and the response is finishing performance rather than a team possession outcome (Rane, 31 Jul 2025).

A further design distinction concerns whether the model estimates a single net effect or decomposes impact into structured components. Basketball and hockey implementations often use separate offensive and defensive coefficients (Petridis et al., 21 Jan 2026, Macdonald, 2012, Jacobs, 22 May 2026). Soccer possession-sequence RAPM splits each player into an involvement coefficient for direct on-ball actions and an on-field coefficient for off-ball presence, including defensive influence (Bajons et al., 2024). This suggests that RAPM is best understood as a family of regularized linear or generalized linear models with sport-specific encodings rather than as a single fixed estimator.

3. Why regularization is central

The rationale for regularization is consistent across the literature. RAPM is needed because OLS APM is unstable when players, lineups, or entities appear together in highly structured ways. Basketball papers describe this as severe multicollinearity and sparse data for many players or lineups (Petridis et al., 21 Jan 2026). Hockey papers identify the same pathology in players who frequently share ice time and in the relative lack of scoring compared to basketball (Macdonald, 2012). Soccer papers add that low scoring and few substitutions weaken identifiability even further (Matano et al., 2018).

Ridge regression is the canonical response because it shrinks coefficients toward zero and guarantees a unique, stable solution. In hockey, this substantially reduces error bounds relative to OLS and improves year-to-year stability, especially when outcomes are shot-based rather than goal-based (Macdonald, 2012). In historical NBA reconstruction, weighted ridge regression is used with the identical mathematical framework applied to modern play-by-play records, and the paper provides explicit posterior covariance formulas and coverage-based rules for $y_i = x_i^\top \beta + \varepsilon_i,$5 (Jacobs, 22 May 2026).

Later work broadens the penalty class. Elastic net is used in CS:GO for both linear and logistic models, with $y_i = x_i^\top \beta + \varepsilon_i,$6 interpolating between ridge and lasso (Xu et al., 2024). Soccer possession-sequence RAPM compares ridge, group lasso, exclusive lasso, and generalized lasso; the groups exploit football-specific structure such as position groups and common strength groups (Bajons et al., 2024). Possession-level NBA work argues that lasso has specific advantages and better performance than ridge regression when compared with selected objective validation criteria, and then uses lasso multinomial logistic regression to construct weighted expected points (wEPTS) (Damoulaki et al., 2024).

The Bayesian reading of RAPM also motivates non-zero prior means. In soccer, FIFA ratings are centered to mean zero and used as the prior mean in Augmented APM, so the shrinkage target is no longer “average player” in the abstract but a FIFA-implied baseline (Matano et al., 2018). In CS:GO, standardized Rating2.0 is used as the prior mean in Bayesian linear RAPM, and a hierarchical extension adds a multiplicative scaling vector to permit player-specific deviations from that prior (Xu et al., 2024).

4. Structured extensions: lineups, hypergraphs, and informed priors

A major development in RAPM research is the move beyond player-only additive models toward lineup-level and interaction-aware estimators. The lineup-level extension in basketball, "Lineup Regularized Adjusted Plus-Minus (L-RAPM): Basketball Lineup Ratings with Informed Priors" (Petridis et al., 21 Jan 2026), keeps the RAPM logic but changes what the coefficients represent. Here the coefficients are lineup offensive and defensive effects, and the regularization shrinks those effects not toward zero but toward player-informed priors built from prior-season player RAPM: $y_i = x_i^\top \beta + \varepsilon_i,$7 The resulting objective is ridge-like but centered at $y_i = x_i^\top \beta + \varepsilon_i,$8, not at $y_i = x_i^\top \beta + \varepsilon_i,$9 (Petridis et al., 21 Jan 2026).

The hypergraph branch generalizes the design itself. "Hypergraph adjusted plus-minus" (Josephs et al., 2024) interprets the usual APM design matrix as the transpose of a lineup hypergraph incidence matrix and extends the construction to observed lower-order combinations. HAPM estimates player effects using weighted ridge regression on an extended incidence structure, while LAPM places a Gaussian random-field prior on generalized lineups through a weighted graph Laplacian. The method is intended to rank individuals, pairs, trios, and full lineups simultaneously (Josephs et al., 2024).

These extensions are closely related to informative-prior RAPM more broadly. In L-RAPM, lineup priors come from sums of player RAPM (Petridis et al., 21 Jan 2026). In Augmented APM, player priors come from FIFA ratings (Matano et al., 2018). In CS:GO Bayesian RAPM, priors can be centered at standardized Rating2.0 (Xu et al., 2024). A plausible implication is that the distinctive contribution of modern RAPM is not only the use of shrinkage, but the choice of what information the model shrinks toward.

5. Empirical behavior and validation

RAPM papers typically validate models by out-of-sample prediction, year-to-year stability, or correlation with domain-relevant targets rather than by in-sample fit alone. This is where regularization, alternative outcomes, and richer priors have usually shown their value.

In hockey, replacing goal-based outcomes with shots, Fenwick, and Corsi produces much smaller standard errors because there are roughly ten times as many shots as goals, and the paper reports higher year-to-year correlations for shot-based ridge APM than for goals-based models (Macdonald, 2012). In soccer possession-sequence RAPM, group lasso and ridge clearly outperform PCV, ELO, and a no-covariate baseline in predicting match results; under ordered logistic regression, Group Lasso Sum attains $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$0 and $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$1, the best values reported in the table (Bajons et al., 2024).

In CS:GO, the comparison against Rating2.0 is explicit. Using data from big events 2018–2023, the correlation between average Rating2.0 and player Plus/Minus is statistically insignificant, with Pearson test $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$2, while RAPM-style models align much more strongly with match-level Plus/Minus. On test data, the Bayesian Plus/Minus rating has $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$3, the logistic plus-minus rating has $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$4, and the elastic logistic plus-minus rating has $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$5 (Xu et al., 2024).

In possession-level NBA work, the final proposal is wEPTS, derived from a lasso multinomial RAPM. The paper states that wEPTS outperforms all other RAPM measures investigated in the study (Damoulaki et al., 2024). In lineup modeling, L-RAPM consistently improves on raw lineup ratings throughout the season, and for unseen lineups the paper reports RMSE improvements around 5% relative to a league-average baseline (Petridis et al., 21 Jan 2026).

RAPM has also been used outside team-ball contexts. In Formula 1, a time-decayed ridge regression with LOESS smoothing is used to separate driver and constructor effects over the Hybrid Engine Era. For the DNF-excluded linear model, the paper reports Partial $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$6, Partial $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$7, and implied constructor influence of $$\hat{\beta} = \arg\min_{\beta} \left\| y - X\beta \right\|_2^2 + \lambda \left\|\beta\right\|_2^2.$$8 (Rane, 31 Jul 2025).

Historical reconstruction work extends RAPM to sparse archival settings. A possession-level database for the pre-play-by-play NBA era reconstructs 2,179 regular-season games, 435,760 total logged possessions, and 1,012 distinct player-seasons, then estimates RAPM via weighted ridge regression on reconstructed stint data (Jacobs, 22 May 2026). That paper’s methodological emphasis is on credible intervals, posterior covariance, and how coverage determines which historical player-seasons are well identified.

6. Misconceptions, limitations, and open directions

A common misconception is that RAPM is simply “plus-minus with a penalty term.” The literature describes a broader object. RAPM can use linear, logistic, or multinomial responses; it can encode offense and defense separately, decompose on-ball and off-ball effects, or shift from players to lineups; and it can shrink toward zero, toward external ratings, or toward player-sum lineup priors (Bajons et al., 2024, Damoulaki et al., 2024, Petridis et al., 21 Jan 2026).

Another misconception is that regularization solves identifiability completely. The papers are explicit that it does not. Stable rosters, low substitution rates, and persistent driver-constructor pairings all produce residual confounding. Soccer Augmented APM uses FIFA-based priors partly to decorrelate highly collinear players (Matano et al., 2018). Formula 1 notes that the model may not be fully separating constructor and driver performance in cases such as Bottas after leaving Mercedes (Rane, 31 Jul 2025). Historical NBA reconstruction emphasizes that heavy collinearity on dynastic teams still constrains what can be learned, even with ridge and posterior intervals (Jacobs, 22 May 2026).

RAPM estimates are also model-dependent because omitted context matters. Hockey models may add zone starts and game state, but not always score effects or coaching effects (Macdonald, 2012, Macdonald, 2010). CS:GO models in the cited work do not explicitly include economy, map-specific effects, side effects, or player roles (Xu et al., 2024). Soccer possession-sequence RAPM is based on whether a possession ends in a goal, a response that is extremely sparse and focused on attacking success (Bajons et al., 2024). Formula 1 excludes weather and circuit-type covariates to preserve interpretability and pre-race predictability (Rane, 31 Jul 2025).

Open directions in the literature are correspondingly structured. The most repeated proposals are richer priors, more granular observation units, and explicit interaction models. Soccer work proposes better priors informed by role, duration, and team structure (Xu et al., 2024). Hypergraph work proposes defense-aware higher-order models and alternative similarity measures (Josephs et al., 2024). Basketball lineup work suggests that separate regularizers for lineup offense and defense could improve predictive performance (Petridis et al., 21 Jan 2026). Historical RAPM suggests that the same Bayesian uncertainty machinery can be carried into partial-coverage settings outside modern play-by-play data (Jacobs, 22 May 2026).

Taken together, these developments indicate that RAPM is less a single metric than a statistical paradigm: regress outcome on participation, regularize aggressively, interpret coefficients conditionally, and validate out of sample. The continuing research agenda is about refining each of those components without losing the central promise of adjusted plus-minus—estimating contribution independent of the strength of teammates, opponents, and other variables that are out of the player’s control (Macdonald, 2012).