Scholarlink

What makes many decisions in sports difficult is that they involve a trade-off between risk and reward. Actions such as taking a three-point shot, carrying a puck, or dribbling with a ball carry a higher risk of failure and require exceptional skill to pull off, but also bring a higher potential reward. This paper describes computational tools for risk analytics to model the risk inherent in the choices faced by teams and athletes. We leverage distributional reinforcement learning (RL) as a source of concepts and techniques for computational risk analytics. Distributional RL techniques allow us to model a dynamic distribution of outcomes for 1000+ games in the National Hockey League. We find strong evidence that strong teams take many risks (0.90 correlation between team season standing and team season standard/Gini deviation). For players, we also find strong evidence that stronger players take more risks (e.g., 0.86 correlation between a player’s season goals and their value-at-risk metric).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Risk, Reward, and Reinforcement Learning in Ice Hockey Analytics

  • 章节

  • 发表日期：

    2025年

查看全文
  • Sheng Xu,
  • Oliver Schulte,
  • Yudong Luo,
  • Pascal Poupart,
  • Guiliang Liu

摘要

What makes many decisions in sports difficult is that they involve a trade-off between risk and reward. Actions such as taking a three-point shot, carrying a puck, or dribbling with a ball carry a higher risk of failure and require exceptional skill to pull off, but also bring a higher potential reward. This paper describes computational tools for risk analytics to model the risk inherent in the choices faced by teams and athletes. We leverage distributional reinforcement learning (RL) as a source of concepts and techniques for computational risk analytics. Distributional RL techniques allow us to model a dynamic distribution of outcomes for 1000+ games in the National Hockey League. We find strong evidence that strong teams take many risks (0.90 correlation between team season standing and team season standard/Gini deviation). For players, we also find strong evidence that stronger players take more risks (e.g., 0.86 correlation between a player’s season goals and their value-at-risk metric).