If you have been watching the Big Bash T20 competition you may have noticed that they have been talking about rating fielders as + or – runs. This came about as one game, whilst commentating, Andrew Symonds would mark down when he believed a fielder saved or conceded runs. This inspired me to write something about cricket fielding metrics – a topic which gained a little coverage towards around this time last year.
The Current State of Fielding Metrics:
Traditionally cricket has recorded three fielding statistics: catches, stumping, and run outs. Now these occur infrequently throughout matches and are a very small percentage of the actual instances of fielding which occur during games. Cricket Australia announced at the end of 2016 they now had a fielding average statistic. This, in effect, combined these traditional statistics with missed chances and also run out assists to produce the percentage of chances a player successfully took. Although a small step forward, as Jarrod Kimber notes in an article on cricinfo, this only penalizes players for dropping the simplest chances and doesn’t assign extra credit for taking difficult ones. This is the same as baseball’s fielding average which has been replaced by more accurate stats such as Defensive Runs Saved and Ultimate Zone Rating (UZR).
The fielding average also fails to deal with the other problem of traditional fielding statistics, that it doesn’t cover the majority of fielding actions. This is where Cricviz has fielding runs saved. Fielding runs saved significantly improves upon standard fielding metrics as it apportions credit for common fielding acts such as saving boundaries. It does this by taking the percentage time that a piece of fielding is made and then charges or credits the runs saved. Say a ball hit to the boundary is fielded 80% of the time for 1 run and missed 20% of the time for 4 runs. If the fielder misses the ball they are charged (0.2-1)*(4-1) = -2.4 runs or conversely if they save it they receive (1-0.8)*(4-1) = +0.6 runs.
Fielding runs saved is not without its problems though. In the same article Kimber notes that the percentage assigned is subjective – and I can find no evidence on cricviz to the contrary – which leads, in his estimation, to fielders who circle a ball rather than running direct routes to gain too much credit. Cricviz also appears to have it’s runs saved/conceded dominated by catches due to their implementation which appears to be (0/1-p(catch taken) * average runs scored by batsmen in that position. Now that is the implementation for test cricket and I assume the same is in place for t20 – in which case it would be better replaced by the change in expected runs scored due to the loss of the wicket (with the bowlers credit subtracted of course).
A Starting Point For Advanced Fielding Metrics:
My starting point for advanced fielding metrics would be to build a metric which rates players as +/- runs. It would do this by assigning a probability of each delivery going for 0, 1, 2, 3, 4, and 6 – overthrows would be dealt with later. It would assign probabilities to balls being hit in the air being caught. It would also assign probabilities of run outs occurring. The proportioning of credit between bowler and fielder for catches would start at 50/50 although this would be altered in future iterations. Run outs would need to have credit split between fielder and batsmen in some way according to mix ups or great pieces of work.
The main part of the model though would be working out the probabilities of each batted ball going for a certain number of runs. In his article Kimber touches on some of the components of how my preferred fielding metric could be built saying:
“With a SportVU camera, spatio-temporal pattern-recognition software, and cricket-specific algorithms, we could work out important and previously unanswerable cricket questions. How long it took the ball to get to the fielder, how far the ball was from the fielder, whether the fielder went straight at the ball, if the fielder took off slowly, how often a second run was successfully made in that situation, and the accuracy and speed of the throw. From that, once enough data is brought in, we could start to work out who played the biggest part in those two runs, and it could be used for everything from wicketkeeping dives to run-out chances. We could tell which fielders make plays, and which ones only execute grade-one chances that any player could make.
But this is quite advanced, and while something of the kind might end up coming to be reality, we are a long way from a system of this sort.”
I don’t believe a simplistic version of this is that far away. Currently we know that television broadcasters have access to the speed of the ball off the bat as well as its trajectory – how else can they make wagon wheels. They also will have a wide angle shot which can show where players are at the point of release. An algorithm could be written which takes these images and gives every fielder xy co-ordinates. Then, based off batted ball data, the ground dimensions, and these xy co-ordinates a multinomial logistic regression could produce a model that provides the probability for each run scoring outcome. With this we can easily compute expected runs and thus fielding runs saved.
This simple model is not perfect, as already mentioned we need to figure out how to provide credit for dismissals as well as dealing with overthrows. We also need to split credit for cases where one fielder saves the ball, and the other throws it in. Still, I believe this to be a useful starting point for a good fielding metric that can truly answer whether someone is ‘worth twenty runs in the field’.
Area Undy the Curve 