Golf Strategy Optimization and the "Drive for show, putt for dough" adage

TL;DR

This paper develops a scalable, data-driven Markov Decision Process model to optimize golf strategies, demonstrating its application using PGA data and challenging traditional beliefs about driving and putting.

Contribution

It introduces an exact, scalable MDP approach for golf strategy optimization, overcoming previous computational limitations and enabling real-world application.

Findings

MDP model can be solved efficiently on large-scale golf data

Analysis shows the relative impact of driving and putting skills

Results challenge the 'Drive for show, putt for dough' adage

Abstract

This study explores strategic decision-making in professional golf's Stroke Play format through a computational lens. We develop a Markov Decision Process (MDP) model-specifically, a stochastic shortest path formulation-to optimize a golfer's strategy on any given course, incorporating both course layout and player skill data. While MDPs have been widely used in sports analytics, applying them to golf presents significant scalability challenges due to the curse of dimensionality. Our primary objective is not to predict player performance with high precision, but rather to demonstrate that an exact, data-driven MDP approach is computationally tractable on full scale, real-world instances. We show that, with careful problem structuring, low-level coding, and efficient memory management, it is possible to solve such large-scale models without resorting to heuristics or Q-learning approximations, as used in existing approaches. To illustrate the model's potential, we show how one can use PGA Tour data and aerial course imagery to simulate strategic outcomes and analyze how different skill profiles influence performance. In particular, we assess the relative impact of driving and putting, challenging the popular adage "Drive for show, putt for dough." These results support the value of our methodology as a robust proof of concept and a foundation for future enhancements. All code and analyses (in R and C++) are made available as open-source resources to support reproducibility and further research.