You Cannot Do That Ben Stokes: Dynamically Predicting Shot Type in Cricket Using a Personalized Deep Neural Network
Total Page：16
File Type：pdf, Size：1020Kb
You Cannot Do That Ben Stokes: Dynamically Predicting Shot Type in Cricket Using a Personalized Deep Neural Network Will Gürpınar-Morgan, Daniel Dinsdale, Joe Gallagher, Aditya Cherukumudi & Patrick Lucey Track: Other Sports Paper ID: 1548748 1. Introduction The ability to predict what shot a batsman will attempt given the type of ball and match situation is both one of the most challenging and strategically important tasks in cricket. The goal of each batsman is to score as many runs as possible without being dismissed. Batsmen can be dismissed in several ways, including being caught by fielders or having their wickets knocked over. While simple in principle, the type of shots and style of a batsman is greatly influenced by the format of the game. In short forms of the game such as T20 and One Day Internationals (the focus of this paper), batsmen are typically more aggressive since their team have a limited number of balls from which to score their runs (120 and 300 balls respectively). Getting the right batsman vs bowler match-up is of paramount importance. For example, for the fielding team, the choice of bowler against the opposition star batsman could be the key difference between winning or losing. Therefore, the ability to have a predefined playbook (as in the NFL) which would allow a team to predict how best to set their fielders given the context of the game, the batsman they are bowling to and bowlers at their disposal would give them a significant strategic advantage. In this paper, we present a personalized deep neural network approach which can predict the probabilities of where a specific batsman will hit a specific bowler and bowl type, in a specific game-scenario. As a motivating example let us consider the 2019 Cricket World Cup Final between England and New Zealand, with England needing 9 runs from 3 balls to win. The ball was an attempted “yorker” length delivery, affectionately known as a “toe cruncher” and renowned for being harder to hit long distances. However, the ball missed its mark and failed to bounce - a “full toss” length in cricketing terms - making it much easier to hit (see Figure 10-B for reference to the length nomenclature). In Figure 1 we visualize how the predicted zone likelihood of Ben Stokes’ shot type for this delivery varies using our deep learning model, where we vary the bowling length while holding all other aspects of its original trajectory fiXed. The left plot shows the predicted shot location of the actual delivery; we then gradually decrease the length of the ball to the attempted yorker (center) and finally a short-pitched delivery (right). We can clearly see that the model predicts the outfield zone in the mid wicket region to be Stokes’ preference for this line of delivery but the absolute magnitude changes by almost 10% and 20% when comparing the yorker with the full toss and short lengths respectively. 1 Figure 1: Using our personalized deep neural network model we can predict various shot types based on the specific bowler, game-state and ball trajectory. Above shows the predicted shot charts for Ben Stokes with 3 balls to go against Trent Boult in the 2019 World Cup final for different length deliveries. The importance of this work is that it can be used across many different elements of the sport. First, for team performance applications, teams can create both pre-game strategies and in-game tactics that evolve throughout an innings as the match conteXt changes. The wealth of information in our model will help teams plan bowling tactics and corresponding fielding locations. Secondly, for media where data drives many storylines for broadcasters, having an estimate of where players are most likely to hit shots in a given situation would enable deeper and more powerful storytelling, which would go beyond current score and win predictors which are currently utilized. 2. Background on Cricket Although often viewed as a niche and somewhat impenetrable sport, cricket is more accessible and global than some realize; in fact, the first international cricket match did not involve teams which we typically associate with the sport and was between the USA and Canada on 24th September 1884 in New York. Since then, the sport has grown into one of the most popular and lucrative in the world, with over 100 member nations and huge TV audiences, such as India vs Pakistan in the 2019 World Cup which saw in eXcess of 250 million unique viewers [1]. A key indicator of its extensive commercial clout is reflected in the enormous rights associated with the Indian Premier League in 2019 at US$6.8 billion – despite its recency, this competition is already approaching the financial pull of the English Premier League in soccer [2]. The aim of cricket is simple: score more runs than the other team. Scoring runs is conducted in a similar manner to baseball, with one player bowling (pitching) the ball to a batsman who defends 3 wooden stumps (called wickets) and attempts to hit the ball in order to try and accumulate runs. However, unlike baseball, the legal hitting area in cricket is 360 degrees from the location of the batsman, who plays on a rectangular pitch in the centre of the playing field as demonstrated in Figure 10-A in the appendix. This large hitting area must be covered by 10 fielders, one of which is the wicketkeeper (equivalent to catcher in baseball) who typically stands directly behind the batsman. 2 Scoring runs can done in one of two ways. Firstly, by hitting boundaries, 4 runs if the ball clears the playing area along the ground and 6 runs, if it's hit over the playing area. Secondly - and most frequently - a batsman can score runs (1, 2 or 3) by swapping with their partner who stands at the opposite end of the pitch before the ball is returned to the stumps, similar to players running from one base to another in baseball. Aggression brings risk however, as bigger shots carry a higher likelihood of dismissals such as being caught by fielders. Once 10 of a team’s 11 batsmen have been dismissed their innings is complete, even if there are balls left to bowl. Therefore, a careful balance between aggression and caution is required to maximize the overall team score. The direction and aggression with which a batter attempts to hit a delivery depends on several contextual factors that are a mixture of premeditated decision-making and split-second reactions once the ball is delivered: where the ball bounces, the speed and spin of the ball being delivered, and the current match state. The goal of the fielding team is to dismiss the batsman and/or limit their run-scoring. The three main factors that the fielding team can control to increase the probability of a favorable outcome are: i. The placement of fielders (subject to restrictions on the number of players patrolling specific regions) ii. The choice of bowler at a given stage of a match (subject to a maXimum number of balls per bowler) iii. The speed and trajectory of the ball (subject to the skill and consistency of the bowler) What the fielding team cannot control at the moment of the delivery is the match context, atmospheric/weather conditions and factors intrinsic to the batsman, such as their level of skill or decision-making. However, the fielding team can leverage their understanding of the ability and tendencies of a batsman to restrict their impact. 3 Bowling teams must attempt to process all of this information on the field in real- time to plan their strategies; fundamentally they are trying to predict what type of shot a batsman will play and where they will direct it. As a prediction problem, this is incredibly challenging due to the breadth and depth of cricket – with its 360-degree nature and wide array of shot types that can range from the brute strength required to launch the ball high into the stands, through to deft touches that barely alter the trajectory of the ball as it rolls gently across the grass. All batsmen have their particular strengths, weakness and preferences for shot type, which can drastically alter the expected shot direction for any given Figure 2: Proportion of runs scored in various zones of the delivery. Contrast this with baseball, field by Ben Stokes since the start of 2018. The map is relative where the hitting arc is 90 degrees and to the batsman who is positioned at the top of the rectangle handedness of the batter hugely narrows in the center of the plot. the optimum hitting angle further. Given how many contextual features are at play in cricket, it is paramount that a model has the capacity and ability to capture the “specificity” of the given situation (i.e. the identity of the batsman and bowler, the fielders and the game-state). Unfortunately, however, current methods do not capture these important contextual features and as such, the best analytics currently generated in cricket rely on either broad averages that ignore conteXt or ever-diminishing sample size. An example is shown in Figure 2, where we show Ben Stokes’ hitting chart for England which is an aggregate of his run scoring areas since the start of 2018 in One Day Internationals, but these do not consider the ball trajectory, bowler or match situation. 3. Related Work In terms of cricket, no previous work on personalizing predictions on shot locations has been done before. Previous analyses have concentrated on scorecard level data for performance analysis, such as the rating of batsman performance in Test match [3] and One Day [4] forms of the game.