A calibration method with dynamic updates for within-match forecasting of wins in tennis

In-match predictions of player win probabilities for professional tennis matches have a wide range of potential applications, including betting, fan engagement, and performance evaluation. The ideal properties of an in-play prediction method include the ability to incorporate both useful pre-match information and relevant in-match information as the match progresses, in order to update the pre-match expectations. This paper presents an in-play forecasting method that achieves both of these goals by combining a pre-match calibration method with a dynamic empirical Bayes updating rule. We present an optimisation rule for guiding the specifications of the dynamic updates using a large sample of professional tennis matches. We apply the results to data from the 2017 season and show that the dynamic model provides a 28% reduction in the error of in-match serve predictions and improves the win prediction accuracy by four percentage points relative to a constant ability model. The method is applied to two Australian Open men’s matches, and we derive several corollary statistics to highlight key dynamics in the win probabilities during a match.