Abstract
Basketball shot location data provide valuable summary information regarding players to coaches, sports analysts, and fans. Represented as spatial points, such data are naturally analyzed with spatial point process models. This paper presents a novel nonparametric Bayesian method for learning the underlying intensity surface built upon a combination of Dirichlet process and Markov random field. Our method effectively encourages local spatial homogeneity when estimating a globally heterogeneous intensity surface. Application to the shot location data of 20 representative NBA players in the 2017-2018 regular season offers interesting insights about the shooting patterns of these players.
The "Goldilocks" Problem in Spatial Statistics
In spatial statistics, a "point process" models random events occurring in a defined space. The standard tool for this is the Nonhomogeneous Poisson Process (NHPP), which estimates an "intensity surface"—essentially a heat map of event probability.
However, existing methods face a significant challenge. Parametric models are often too rigid to capture the idiosyncratic shooting patterns of real players. Conversely, current nonparametric models, such as Dirichlet Process Mixtures (DPM), often ignore the "First Law of Geography": the idea that near things are more related than distant things. Without accounting for "spatial contiguity," resulting maps often appear noisy and scattered, making it difficult for analysts to identify a player's true preferred zones.
Methodology: The MRF-DPM-NHPP Model
The researchers introduced the MRF-DPM-NHPP model to address these limitations. The model combines three key components:
- NHPP: The base model for counting points in space.
- DPM (Dirichlet Process Mixture): A nonparametric clustering method that allows the number of "intensity zones" to be learned from the data rather than fixed in advance.
- MRF (Markov Random Field): The "secret sauce" that adds a spatial constraint to the clustering process.
When the model assigns a specific grid square on the court to a cluster, the MRF acts as a "nudge," encouraging it to join the same cluster as its neighbors. This enforces local spatial homogeneity (smoothness). Crucially, the model still allows for global heterogeneity—it can recognize that two physically separated areas (like the left and right 3-point corners) belong to the same intensity cluster because of the court's symmetry.
Data and Application
The model was tested on both synthetic data and real-world NBA shot location data from the 2017-2018 regular season. The study focused on 20 all-star level players across five positions, mapping their shots on a 50x35 ft rectangle of the half-court.
The raw data (shown above) highlights the diverse shooting distributions that the model must interpret. For instance, Stephen Curry's high volume of three-point attempts contrasts with the more mid-range and interior focus of other players.
Key Insights and Results
The MRF-DPM-NHPP model successfully identified player-specific "fingerprints" while maintaining spatial clarity.
- The Power of Contiguity: The inclusion of the MRF term significantly reduces "noise." While previous models might produce a "checkered" pattern of high and low intensity, this model produces clean, continuous "hot zones."
- Symmetry and Style: The model identifies that players often shoot with equal intensity from both wings without forcing the area between them to be high-intensity.
- Quantitative Profiles: The following table summarizes the shot counts and percentages for the four players featured in the shot charts.
In simulation studies, the proposed method was vastly superior at identifying the correct number of clusters. In complex settings, it found the true number of clusters 82% of the time, compared to just 20% for benchmark models. It also achieved a Rand Index (RI)—a measure of clustering accuracy—of over 0.97, indicating nearly perfect recovery of spatial structures.
Conclusion
The MRF-DPM-NHPP model provides a more interpretable and accurate way to map NBA shooting patterns. By balancing local smoothness with global complexity, it offers coaches and analysts a clearer view of a player's "hot hand." Future work may involve incorporating covariates like defender distance or shot clock time to further refine these intensity surfaces.
Summary updated