This is part 3 of the Player Evaluator and Calculated Expectancy (PEACE) model, which is an alternative to Wins Above Replacement. This article will introduce evidence that z-scores can be converted into runs (or points in other sports) with accuracy and reliability, as well as analyze the results that zDefense has produced.
Recall that zDefense is broken down into 4 components: zFielding, zRange, zOuts, and zDoublePlays. The fielding and range components depend on the accuracy of Calculated Runs Expectancy, which I introduced in Part 1. Outs and double plays, though, use a different technique: they take z-scores for the relevant rate statistics, then multiply by factors of playing time. Here were the equations:
- zOuts = [(Player O/BIZ – Positional O/BIZ) / Positional O/BIZ Standard Deviation] * (Player Innings / Team Innings) * (√ Player BIZ / 2)
- zDoublePlays = [(Player DP/BIZ – Positional DP/BIZ) / Positional DP/BIZ Standard Deviation] * (Player Innings / Team Innings) * (√ Player BIZ / 2) * Positional DP/BIZ
We can set up models in other sports that estimate point differentials using very similar techniques. I’ve developed one for college football and another for the NBA.
For the first model, I’ve used the data for every Division I FBS football team from 2000-2014 (1,802 teams), and I defined the relevant statistics and their “weights” as such:
- zPassing = [[Completion Percentage z-score * Completions per Game] + [Passing Yards per Attempt z-score * Passing Attempts per Game]] / 10
- zRushing = [Rushing Yards per Attempt z-score * Rushing Attempts per Game] / 10
- zTurnovers = [Turnovers per Game z-score]
- zPlays = [Number of Offensive Plays per Game z-score]
These 4 components summed make up zOffense, while taking each team’s opponents’ calculations results in zDefense.
What I found after summing the different components was that the resulting number, when divided by the number of games played, was a very accurate estimator for a team’s average point differential.
Among the nearly 2,000 college football teams, the average difference between zPoints (calculated margin of victory) and actual MOV was just 3.21 points, with a median of 2.77, and a max difference of 13.97 points. About 20% of teams’ MOV were calculated to within 1 point or less, 53% were accurate to 3 points or less, 79% to 5 points or less, and 99% to 10 points or less. The regression model for this dataset can be seen below:
The NBA model has similar results using 6 parts:
- z3P (3-point shots) = [[3P FG% z-score * 3-point attempts * 3] / 10
- z2P (2-point shots) = [2P FG% z-score * 2-point attempts * 2] / 10
- zFreeThrows = [FT% z-score * free throw attempts] / 10
- zTurnovers = [Turnovers per Minute z-score * League Average Points per Possession] * 2
- zORB (offensive rebounds) = [Offensive Rebounds per Minute z-score * League Average Points per Possession]
- zDRB (defensive rebounds) = [Defensive Rebounds per Minute z-score * League Average Points per Possession]
Similar to the football model, these 6 components make up zOffense, while each team’s opponents’ calculations make zDefense. I particularly like z3P, z2P, and zFT because they multiply the z-score by the “weight”: 1, 2, or 3 points. Recall that zRange is multiplied by the IF/OF Constant, which is just the difference, on average, in runs between balls hit to the outfield vs. balls that remain in the infield.
I’ve only done the calculations for the 2013-2014 season, where teams averaged 1.033 points per possession. To convert to zPoints in this model, add zOffense and zDefense, then divide by 5.
In most seasons, elite teams will have an average point differential of +10, while terrible ones will hover around -10. On average, the NBA model had an average difference between the calculated and actual differential of just 1.331 points, with a median of 0.800. 17 out of 30 teams were calculated within 1 point, 25 within 2, and 29 out of 30 were accurate to within 5 points per game.
The fact that these models can be created using the same general principle (rate statistic z-scores multiplied by a factor of playing time equates relative points) provides some evidence that similar results are calculable in baseball. This is the basis for zDefense in PEACE. Let’s look at the results.
Most sabermetricians would turn to the Fielding Bible Awards for a list of the best fielders by position in any given year, so we’ll use those results to compare. If we assume that the Fielding Bible is accurate, then we would expect zDefense to produce similar conclusions. Comparing the 2014 winners to the players ranked as the best at their position by zDefense, we can see some overlap. The number in parentheses is the positional ranking of the Fielding Bible Award winner by zDefense.
- Position: Fielding Bible Winner (#)…zDefense Winner
- C: Jonathan Lucroy (12)…Yadier Molina
- 1B: Adrian Gonzalez (1)…Adrian Gonzalez
- 2B: Dustin Pedroia (2)…Ian Kinsler
- 3B: Josh Donaldson (2)…Kyle Seager
- SS: Andrelton Simmons (8)…Zack Cozart
- LF:Alex Gordon (1)…Alex Gordon
- CF: Juan Lagares (3)…Jacoby Ellsbury
- RF: Jason Heyward (1)…Jason Heyward
- P: Dallas Keuchel (5)…Hisashi Iwakuma
The multi-position winner, Lorenzo Cain, was also rated very favorably by zDefense. While most positions don’t have a perfect match, every single Fielding Bible winner was near the very top of their position for zDefense. This is the case for almost every instance, which isn’t surprising: if there were drastic disagreements about who is truly elite, then we would suspect one of the metrics to be egregiously inaccurate. Instead, we see many similarities at the top, which provides some solid evidence that zDefense is a valid measure.
As always, feel free to comment with any questions, thoughts, or concerns.
Walter is an undergraduate researcher at The Ohio State University studying Economics. He hopes to have a career in baseball operations after graduation with a focus on sabermetric analysis.