This article proposes a novel approach, called data snapshots, to generate real-time probabilities of winning for National Basketball Association (NBA) teams while games are being played. The approach takes a snapshot from a live game, identifies historical games that have the same snapshot, and uses the outcomes of these games to calculate the winning probabilities of the teams in this game as the game is underway. Using data obtained from 20 seasons worth of NBA games, we build three models and compare their accuracies to a baseline accuracy. In Model 1, each snapshot includes the point difference between the home and away teams at a given second of the game. In Model 2, each snapshot includes the net team strength in addition to the point difference at a given second. In Model 3, each snapshot includes the rate of score change in addition to the point difference at a given second. The results show that all models perform better than the baseline accuracy, with Model 1 being the best model.

中文翻译：

---

一种用于在篮球中进行实时预测的数据快照方法。

本文提出了一种称为数据快照的新颖方法，可以在比赛进行时为美国国家篮球协会（NBA）的球队生成实时获胜的概率。该方法从实况游戏中获取快照，识别具有相同快照的历史游戏，并使用这些游戏的结果来计算游戏进行中该团队在该游戏中的获胜概率。利用从20个赛季的NBA比赛中获得的数据，我们构建了三个模型，并将其准确性与基准精度进行了比较。在模型1中，每个快照都包含了比赛给定时刻主队与客队之间的得分差。在模型2中，每个快照都包括团队的净实力，以及给定秒数上的点差。在模型3中，除了给定秒数的点差外，每个快照还包括得分变化率。结果表明，所有模型的性能均优于基准精度，其中模型1是最佳模型。