Methods and indicators for baseball modelling

Intro

Baseball modelling will always be associated with Michael Lewis book – Moneyball. It is the story of Oakland A’s and their groundbreaking analytical approach to player evaluation and signing amateur talent. This book revolutionized baseball, sports modeling and analytics. The term “Moneyball” is now used to describe the analytics-based approach to scouting and development of players and teams. Baseball has a discreet and specific gameplay structure that allows the gathering of a huge amount of statistical data. Having a prediction on future players’ and teams’ performance based on stats, brings enormous advantages for managers and punters.

Data

The explosion of data in baseball began in 2007 – earlier than soccer. New game tracking software, performance models and human resources added to the new sabermetric idea – the science of learning about baseball through stats. Data is now being generated in a continuous stream and teams are finding ways to exploit that amount of information in the best possible way.

Some of the data that is being tracked today is the velocity of the pitch, the three-dimensional movement of the ball, the pitcher’s arm angle, movement of pitchers and baserunners.

Statcast is an automated, high-speed and high-accuracy tool that has been developed to track players movements and events in Major League Baseball. In 2015, that tool was installed on all thirty MLB stadiums. Compared to OPTA in-house developed software for soccer tracking (more about OPTA can be read in our previous articles about: Football modelling and expected goals), it provides the same detailed raw data for baseball analysis. Statcast use a combination of cameras and radars to track the ball in every position it reaches as well as every player on the field at all times. It is important to mention this tool as it provides an incredible wealth of information such as where the pitcher released the ball, how hard he threw, what sort of spin he had, how fast it appeared to the batter, exit velocity, launch angel, batted ball direction, sprint speed etc.

Baseball methods and models

In baseball, pitching is very difficult skill to learn and an essential component of obtaining victories. In baseball you have to outscore your opponent and pitchers’ main job is preventing runs. Pitchers aren’t responsible for the opponents runs entirely, because the defense plays a major role as well. Pitchers force great weight over walks, home runs and strikeouts and much less control over hits as these are being handled by the defense. For this reason, evaluating pitchers’ form pre-game and in-play is crucial in finding value and winning the match. (Read here about MLB’s Most Profitable Pitchers)

Also, substituting the pitcher with a reliever is important decision-making problem, because the reliever needs 10-15 minutes warm-up. Moreover, this decision involves the “accurate” judgement of the pitcher’s form – is he fatigued, faltering or underperforming? Knowing all this helps managers, but at the same time it provides an enormous value to the bettor.

One example of estimating pitcher’s performance is by modelling and analyzing pitch-by-pitch data as time series data. It is known that historical statistics and past data could show crucial patterns in many baseball plays. Past pitches records and outcomes is needed for a time series analysis.

Time series data can be defined as a data that collectively represents how a system/process/behavior changes over time. Here the time isn’t just a metric, but the main axis. The main idea here is to create a predictive model for determining when a starting pitcher is about to falter using time series classification methods. By assigning different scores to each possible pitcher’s thrown outcome, pitch events are modelled as time series data. A time series model will, in general, reflect the fact that observations close together in time will be more closely related than observations further apart.

After that we use the k-Nearest Neighbours classification algorithm for predicting the future performance. k-Nearest Neighbours is classification by finding the most identical data points in the sample data, and making an intelligent guess based on their classifications.

A player’s fielding is also important part of the baseball defense and can be modelled and evaluated. Fielding can be broken down in the following skills: positioning, range and arm. For a variety of reasons, positioning is often overlooked when evaluating a player. It can make a tremendous difference if a fielder changes his position with a few steps. A good position will allow the fielder to not work as hard to get there.

Range reflects how much ground a fielder is able to cover in order to make a play. We do have to mention that range can be summed up with three important skills: reaction time, reading the ball off the bat and speed. A good reaction time is always better, but for certain position, third base, reaction time is much more important.

Deciding where the ball will end up based on angle and bounces is very difficult and it only takes a split second to decide. Last but not least, speed is crucial when fielding because it can make up for the poor reaction time and difficulty reading the ball off the bat.

A fielder with a strong arm is nothing without accuracy and the right combination of these two attributes provide a better play.

For the above-mentioned reasons, a player’s fielding ability is more difficult to evaluate. If a batted ball moves through the left side of the infield, it will be very difficult to know whether a faster or better positioned fielder could have made the play.

Another problem in evaluating the fielding in baseball is the large playing surface. There mentioned approaches that attempt to counter this problem. The Ultimate Zone Rating is based on a division of the playing field into 64 large zones and evaluating fielders’ successful plays within each zone. The Probabilistic Model of Range does a similar division into 18 pie slices on either side of the second base and fielder evaluated within each slice. The problem with these models Is that every ball that moves into a certain zone is treated equally.

A hierarchical Bayesian model can be used to evaluate the success of each individual fielder. One logic here is to estimate the number of runs that a given fielder saves or costs his team during a season compared to the average fielder at his position. This can be done by modelling the binary variable of whether a player successfully fields a given ball in play as a function of its location. Then, we integrate over the estimated distribution of ball in play locations and multiply by the estimated consequence of a successful or unsuccessful play, and we will have a final estimate of the number of runs saved or cost by a given fielder in a season.

There are other statistical approaches that can be utilized and model pitchers/batters and teams’ performances. Bookmakers have their own in-house models that compile odds based on starting pitcher, home field advantage and recent performance. Current MLB lines and odds from top European and Asian bookmakers can be found here: Oddshark

Out of experience in the field of building sports betting models, the best approach is to start small and simple, add results and make notes of steps. After that you can add levels of complexity based on angles and situations that you have tested. Also,

Indicators

Measuring player performance involves the use of all the available statistics, leading to important methods, indicators and derivatives. All this is carefully calculated and used to evaluate players’ past contributions to their teams, their current and future value. With the flooding of advanced metrics, it is important to understand the significance and insight that some of these indicators bring to us.

We will start with the most common one called ERA – earned runs average. This is the number of earned runs a pitcher allowed and divided by the number of innings he pitched. All this is then multiplied by nine and the calculated number will give us an expectation of pitcher’s earned runs over the course of nine innings. This metric is very basic and easy to understand and shows us exactly what we want a pitcher to do – prevent runs. These stats can also be deceptive depending on his performance in the bullpen area.

The Fielding Independent Pitching (FIP) is a popular metric that attempts to eliminate the role that defense can play in affecting the number of earned or unearned runs a pitcher allows. It only measures events that cannot be affected by the defense – home runs, strikeouts, hit by pitchers and walks. Here we need to point out that by stripping out balls put in play, you are eliminating the influence of the defense, but we are also assuming that pitchers have no control over the quality of contact that is made at the plate.

The Expected Independent Pitching (xFIP) is a regressed version of FIP and it is calculated in the same way with the exception that it uses a pitcher’s fly ball rate to estimate how many home runs the pitcher should have given up, assuming the league average home run to fly ball rate. We can say the xFIP is a better indicator of skill and it is a better metric to predict how a pitcher will perform in the future.

Strikeout and Walks per nine innings measure the averages of these attributes for a pitcher per nine innings. These are very important when evaluating pitchers because pitchers, in general, play a prominent role in strikeouts, walks and home runs while defense is important in balls in-play. Because pitchers have control over their strikeouts and walk rates, these attributes are a good measure on the pitcher’s performance and skills. Pitchers who gather strikeouts routinely prevent runs and those who allow walks typically allow more runs. From this point of view, these stats are quite important in guessing a pitcher’s performance.

Another important indicator is the Skill Interactive Earned Runs Average or shortly SIERA. It uses a pitcher’s strikeout-rate, walk-rate and groundball-rate to estimate a pitcher’s performance. The mentioned three variables are used because they do not vary as much as others from year to year. The main idea behind SIERA is that pitchers who produce a high number of strikeouts also generate a weaker contact. This indicator gives more credit for strikeout than xFIP does and less weight to fly balls.

If we have to quickly sum up the above-mentioned indicators, the biggest disadvantage to ERA is the quality of defense behind the pitcher because it can have a larger effect on the earned runs he gives up. FIP attempts to solve this by eliminating balls in play and it assumes that pitchers have no control over the quality of contact that is made. SIERA, on the other hand, assumes that pitchers have control over the quality of contact made and they want to generate a weak contact.

There are statistics that consider the hitter’s performance as well. One of these is the Weighted On-Base Average (wOBA) which attempts to credit a batter for the value of each outcome rather than treating all hits or times on base equally. It is one of the most important and popular catch-all offensive stats. The concept behind the wOBA is very simple and it assumes that not all hits are created equal. Batting average and on-base percentage assumes the opposite with the difference that on-base percentage includes walking or being hit by a pitch. Slugging percentage (the total number of bases player records per at-bat) weights hits, but not accurately and again ignores other ways of reaching base.

Weighted On-Base Average sums up all aspects of hitting into one metric and weights each of them in proportion to their actual run value. While batting average, on-base percentage and slugging percentage are less accurate, WOBA calculates offensive value more accurately.

Measuring how often a ball in play goes for a hit is called Batting Average on Balls In Play (BABIP). Here we consider a ball in play when the plate appearance ends in something other than a strikeout, walk, hit batter, catcher’s interference, sacrifice bunt or home run. This metric can be useful by telling you something about the quality of the player. Furthermore, big deviations from a player’s career average BABIP can give you hints about their luck or the performance of the defense. BABIP is influenced by three main factors: defense, luck and talent.

Defense is a factor in BABIP because a player have no control over the opponents’ defense and they can direct their hits to a limited extent. Good contact when batting doesn’t mean that the ball will be going in a good direction.

Luck is always a factor and there has been numerous events to prove that not only in baseball but in other sports as well. For example, a well hit ball may go right to where a fielder is standing even thou the batter struck the ball at a very high velocity.

Talent plays role in this metric because a better hitter will usually have a higher BABIP than a worse hitter and a worse pitcher will have a slightly higher BABIP than a better pitcher.

To sum up, this metric is good for batters and pitchers in different ways. BABIP can be a good indication about the batter’s overall quality of contact. For pitchers, BABIP is even more important because pitchers have no control over what happens to a ball once it is in play.

All these indicators may have predictive power, but it is unclear which is most reliable or consistent. There are written logics and models that address this question with the help of statistical tools. It is important to run such tests and decide which indicators would be most suitable for your model. One way of conducting such test is via a Bayesian random effects model where all the parameters are considered random within the Bayesian framework.

The idea here is to determine which metrics have signal and which are dominated by noise. If a metric was pure noise, pitchers would be inconsistent from year to year and in this case the best prediction one could give would be the league average.

Conclusion

Baseball is a huge area of research in the realm of predictive analytics. Important metrics and attributes are analyzed, and their predictive power evaluated. This process is crucial when building predictive models and the more time you spend on researching, analyzing and testing metrics, the more reliable your model will be. Evaluating the performance of pitchers and batters using pre-match attributes will allow you to spot value in each game and make educated bets (Have a look at this must-read guide on: How to place an educated bet and MLB Playoff Betting). Predicting the winner of a Major League Baseball games is an interesting and challenging task. There are no definitive formulas or models that will provide accurate information on factors that will lead a team to a victory. The analysis of historical data is the only way to discover new trends in this field. Sabermetrics will be leading baseball in terms of teams and player analysis along with forecasting games.