Random Walk Picture of Basketball Scoring

We present evidence, based on play-by-play data from all 6087 games from the 2006/07–2009/10 seasons of the National Basketball Association (NBA), that basketball scoring is well described by a weakly-biased continuous-time random walk. The time between successive scoring events follows an exponential distribution, with little memory between different scoring intervals. Using this random-walk picture that is augmented by features idiosyncratic to basketball, we account for a wide variety of statistical properties of scoring, such as the distribution of the score difference between opponents and the fraction of game time that one team is in the lead. By further including the heterogeneity of team strengths, we build a computational model that accounts for essentially all statistical features of game scoring data and season win/loss records of each team.

💡 Research Summary

The authors investigate whether the evolution of the score in NBA basketball games can be described by a simple stochastic process. Using play‑by‑play logs from all 6,087 regular‑season games played during the 2006‑07 through 2009‑10 seasons, they first quantify the basic scoring statistics. A “scoring play” is defined as any sequence of baskets occurring without any elapsed game‑clock time; on average there are 94.8 such plays per game, corresponding to a mean scoring rate of λ≈0.0329 plays per second, which is essentially constant throughout the 48‑minute regulation period (aside from brief anomalies at the start and end of each quarter that are ignored in the subsequent analysis).

The distribution of time intervals between successive scoring events is examined in two ways: (i) t_e, the interval between any two scores regardless of which team scores, and (ii) t_s, the interval between two scores by the same team. Both distributions display an exponential tail, P(t)∝e^{‑λ_tail t}, with λ_tail≈0.048 plays s⁻¹ for t_e and half that value for t_s, indicating that scoring follows a continuous‑time Poisson process with negligible memory. Autocorrelation analysis confirms this, as the lag‑n correlation C(n) stays below 0.03 for all n≥1.

Next, the authors study which team scores after a given play. Because possession changes after a score, the same team scores again with probability q≈0.348, while the opponent scores with probability 1‑q≈0.652. This anti‑persistence (a step in one direction is likely to be followed by a step in the opposite direction) is a key ingredient of the model. They define a “streak” as a run of consecutive points scored by one team before the opponent scores. Assuming an average of s̄≈2.0894 points per play, the probability of a streak of s points is Q(s)=A q^{s/s̄}, which reproduces the observed exponential decay of streak lengths.

To improve realism, the model incorporates the empirical distribution of play values: 1‑, 2‑, 3‑, and 4‑point plays occur with probabilities w_1,…,w_4 (Table 1). By treating a streak as a sequence of plays whose point values sum to s, the exact streak probability becomes a convolution over all admissible sequences. This leads to the recursive relation Q(s)=q∑_{α=1}^{4} w_α Q(s‑α), which can be evaluated numerically for any s. The resulting Q(s) matches the data far better than the simple exponential, confirming that streaks arise purely from random fluctuations rather than “hot‑hand” effects.

The authors also detect a systematic dependence of scoring probability on the current lead. The winning team’s scoring rate declines linearly with its lead size, while the losing team’s rate rises, with a slope of about 0.0022 per point. This linear restoring force is analogous to the Ornstein‑Uhlenbeck drift term and reflects a modest “coasting” behavior by the leader and increased effort by the trailer.

Finally, to capture season‑long variations in team quality, each team i is assigned a baseline scoring rate λ_i=λ exp(σ η_i), where η_i is a standard normal variable and σ controls the heterogeneity of team strengths. Monte‑Carlo simulations of entire seasons (including the anti‑persistence rule, the linear restoring force, and the heterogeneous λ_i) reproduce a wide array of empirical statistics: the distribution of final score differences, the fraction of game time a team spends in the lead, the number of lead changes per game, and the win‑loss records of all teams over 20 seasons. The agreement is quantitative, with no adjustable parameters beyond those directly measured from the data.

In summary, the paper demonstrates that NBA scoring dynamics are well captured by a weakly biased, continuous‑time random walk with a small linear restoring force and an intrinsic anti‑persistence due to possession changes. By adding realistic play‑value probabilities and team‑strength heterogeneity, the model accounts for virtually all observed statistical features of basketball games, suggesting that complex tactical narratives or “hot‑hand” phenomena are not required to explain the bulk of scoring behavior.