Basketball Analysis Workspace

Lead Dominance Index (LDI) is treated here as a continuous-game measure of scoreboard control rather than as a box-score efficiency statistic. For each interval between observed score changes, the score differential is held constant and multiplied by the interval length; the team-season value is the signed sum of these point-time areas. Formally, for interval j, LDI = sum(D_j * delta_t_j), where D_j is the team's score differential and delta_t_j is elapsed game time. The displayed predictor, Avg LDI/Game/48, divides each team's season total by games played and normalizes the result to a 48-minute game, which makes lockout, overtime, and pace-adjacent timing differences less dominant. This construction follows the same premise that motivates score-process models in basketball: the evolving score differential contains information about team strength and win probability, not merely the final score. Stern's Brownian-motion model of sports scores is the statistical precedent for treating the game margin as a time-indexed process, and Kubatko, Oliver, Pelton, and Rosenbaum provide the broader basketball-analytics basis for evaluating team performance through normalized, possession-aware and efficiency-oriented summaries.

The predictive test is intentionally out of sample. All available regular seasons are ordered chronologically. For a held-out season s, the model fits a one-variable linear regression on prior seasons only: Predicted Win% = alpha + beta * Avg LDI/Game/48. The fitted equation is then applied to each team in season s, predicted win percentage is bounded to the interval [0, 1], and predicted wins are obtained by multiplying by games played. The table reports season-level accuracy, mean absolute error, root mean squared error, R2, correlation, and the largest team miss. Accuracy is defined as 1 - total absolute win error / total games, so it expresses the share of scheduled games not lost to absolute prediction error. For the earliest seasons, where no prior-season training set exists or the prior set is too small for a stable regression, the model trains on all other seasons and flags that fallback in the coverage statement. This design follows the out-of-sample validation logic emphasized by Sill in adjusted plus-minus work: a metric intended to explain winning should be judged by its performance on games or seasons not used to estimate the model, because in-sample fit can reward noise.

LDI should not be read as a proven replacement for plus/minus. If plus/minus means team point differential or net rating, LDI is usually expected to be less directly predictive of season wins than plus/minus, because plus/minus is the realized scoring margin that most closely determines wins and losses. NBA Stats defines net rating as team point differential per 100 possessions, and Basketball-Reference's Pythagorean wins framework estimates expected wins directly from points scored and allowed; both conventions reflect the empirical strength of aggregate scoring margin as a win predictor. LDI adds a different signal by preserving when the margin occurred: a team that leads by double digits for 40 minutes and wins by one will rate very differently from a team that is even all night and wins by one, even though both games have the same final margin. That temporal information can describe control, pressure, and game script, but it also creates reasons for weaker win prediction: late-game fouling, garbage time, deliberate clock management, and comeback volatility can make sustained control diverge from final margin. In short, LDI is best interpreted as a complementary measure of dominance shape. It is more descriptive of how a team controlled games, while team plus/minus or net rating remains the cleaner benchmark for predicting the number of games won.

If plus/minus instead refers to raw player plus/minus, the comparison changes. Raw player plus/minus measures the score change while a player is on the court, but Sill notes that unadjusted plus/minus is confounded by teammates, opponents, lineup collinearity, and overfitting; adjusted plus-minus and regularized adjusted plus-minus were developed precisely to correct those problems. LDI in this project is a team-season statistic, so it avoids individual lineup attribution but also does not identify player value. The appropriate conclusion is therefore limited: team LDI may be more stable and interpretable than raw single-player plus/minus for describing team control, but it should be expected to trail team net rating or point differential as a pure season-win predictor unless a direct head-to-head model shows otherwise.

Sources: Hal S. Stern, A Brownian Motion Model for the Progress of Sports Scores; Justin Kubatko, Dean Oliver, Kevin Pelton, and Dan T. Rosenbaum, A Starting Point for Analyzing Basketball Statistics; NBA Stats, Stat Glossary; Basketball-Reference, Glossary and Pythagorean Wins definition; Joseph Sill, Improved NBA Adjusted Plus-Minus Using Regularization and Out-of-Sample Testing.