A Data Snapshot Approach for Making Real-Time Predictions in Basketball

Area of Science:

• Sports Analytics
• Machine Learning in Sports
• Basketball Performance Metrics

Background:

• Predicting in-game outcomes is crucial for sports analytics and betting.
• Existing methods often lack real-time adaptability during live National Basketball Association (NBA) games.

Purpose of the Study:

• To develop and evaluate a novel approach, "data snapshots," for generating real-time win probabilities for NBA teams during live games.
• To compare the predictive accuracy of different data snapshot models against a baseline.

Main Methods:

• A "data snapshot" method was developed, capturing game state at specific moments.
• Historical NBA game data from 20 seasons was used to train and test models.
• Three models were constructed: Model 1 (point differential), Model 2 (point differential + net team strength), and Model 3 (point differential + score change rate).

Main Results:

• All three proposed models demonstrated superior accuracy compared to the baseline accuracy.
• Model 1, which utilized only the point differential within a snapshot, yielded the best performance.
• The "data snapshots" approach effectively estimates live win probabilities.

Conclusions:

• The "data snapshots" method offers a viable and accurate way to predict NBA game outcomes in real-time.
• Simple game state information, like point differential, is highly effective for in-game win probability estimation.
• This approach has implications for live sports analytics, broadcasting, and potentially sports betting.