Tangotiger Blog

Friday, October 21, 2022

Revenge of the Defense

By Tangotiger 12:49 PM

There is a very simple way to determine how much the offense and how much the defense influences the game. This applies not only to baseball, but to hockey, basketball, football, soccer, and any team sport you can think of. What we care about is the spread in scoring by the offense, and the spread in scoring allowed by the defense, on a team by team basis. If for example defense didn't matter at all, then the spread would be explained purely by random variation. And so the spread in offense would overwhelm the spread in defense. Similarly, if scoring was all about defense (run prevention), then we'd see a huge spread in runs allowed among the teams, and very little spread in run scoring.

Now, what do we actually see in baseball? Well, from 1974 to 2013, the spread in runs scored by 27 outs was one standard deviation of 0.46 runs, while for runs allowed it was 0.48 runs. So, we have slightly more spread in defense, but not that much more. This probably due to the fact that defense is about pitching and fielding, so you have two distributions coming together. I would imagine if I did this in hockey, where we have skaters and goalies, that the spread in goals allowed would be larger than goals scored. (Taking a guess for football, where the QB is so prominent, I'd guess the reverse. I won't even guess for basketball.)

But, since 2014, things have taken a much much larger turn. One standard deviation for run scoring is 0.44, so a bit tighter than we're used to. Runs allowed however has ballooned to one standard deviation of 0.56 runs. Why would that be? Well, in addition to pitchers and fielders being involved as two distributions, we also have a third distribution: fielding alignments. This is making it so that its a differentiator.

The 2022 season is a good example. The two highest scoring teams were the Dodgers and Yankees at 4.98 runs per game and above. We had FIVE teams allowing that many runs or more. The Tigers scored 3.44 runs per game, while we had two teams (Astros, Dodgers) who allowed far less than that. As a result, the spread in offense is tighter than the spread in defense. And so the determination of who will win the game will be more influenced by the defense than the offense.

Here is the year by year chart (click to embiggen) in the spread in scoring by offense and defense, where you can see the clear separation happening in 2014, not coincidentally when Shifts Attacked.

We can also show this as a percentage, taking the spread of defense divided by the spread of defense + spread of offense. From 1974 to 1983, it was 48% for defense. From 1984 to 2013, it was 52% for defense. Since 2014, that's exploded to 56%. Now, is it really about the shift and front office influenced fielding alignments? I don't know yet. It could also be about the shift in starting and relief pitching usage. I don't know yet. The next step is to look at the spread in talent at the player level.

What it could also be is how teams acquire talent. It could be that the spread in talent on the offense and defense side is the same year to year, but the way that talent is collected, there's more discrepancy or deviation on the defense side. This gets a bit complicated because position players contribute on both the offense and defense side.

Anyway, I'll look into it next time. And it'll be very interesting to see the results in 2023, when Infield Shifts are a thing of the past. That will more definitively give us the answer we're looking for. We can wait to see how that plays out too.

#1 TwoScoops 2022/10/21 (Fri) @ 13:56

Given the defensive SD of .56 and offensive of .44, would it be fair to suggest that since 2014 defense is 56% of baseball?

#2 TwoScoops 2022/10/21 (Fri) @ 14:02

2021 NFL
Offensive PPG STDEV: 4.44
Defensive PPG STDEV: 2.89

#3 TwoScoops 2022/10/21 (Fri) @ 14:09

2021-22 NBA
Offensive PPG STD: 3.39
Defensive PPG STD: 3.68

2021-22 NHL
GF/G STD: 0.42
GA/G STD: 0.39

#4 Tangotiger 2022/10/21 (Fri) @ 15:28

Thank you kindly 2Scoops!

I kind of figured NFL would be an easy call, but 61% share is quite the number!

NHL is where I thought it would be directionally, 52%, though a bit lower. I would have figured closer to 55%.

NBA I really had no idea, so seeing it heavier on the defense is very interesting. Also 52%.

So, NHL, NBA, and until 2013, MLB all come in at 52%.

#5 Tangotiger 2022/10/21 (Fri) @ 15:30

As for: does it mean that defense is 56%? Almost. We have to back out the random variation from the run scoring.

But once you do that, you’ll still be pretty close to that 56%. It might nudge to 57%.

#6 jgf704 2022/10/22 (Sat) @ 13:33

Interesting!

I thought it would be fun to get a longer view. Here’s data going back to 1960. My “%DEF” is as you calculated it (i.e. I did not remove the random component). Plus I used Lahman data, so mine is actually “per game” rather than “per 27 outs”. The line is based on a centered 5-year moving average of the StDevs. Those StDev are simply the root mean square of the annual StDev (so square root of 5 year average of variances).

https://i.imgur.com/YjKZ5Cy.png

#7 jgf704 2022/10/22 (Sat) @ 13:35

Well, dang it, the image showed up in the Preview, but not in the comment.

#8 jgf704 2022/10/22 (Sat) @ 13:45

Here’s the data going back to 1920:

#9 Tangotiger 2022/10/25 (Tue) @ 00:07

Excellent, thank you for sharing!

#10 Guy 2022/10/26 (Wed) @ 11:07

Adding a little more data to the discussion: I compared the periods 2005-2013 to 2014-2022, and looked at the SD for BABIP and FIP in each period. Both SDs increased in the later period, but the increase was a bit larger for FIP. I’m sure this analysis can be refined, but this data would imply that about 60% of the increased variation on defense is due to pitching, 40% to fielding.

FIP
2014-2022: .458
2005-2013: .376
CHANGE: .083

BABIP:
2014-2022: .013
2005-2013: .010
CHANGE: .003 (.054 runs/game)

One interesting wrinkle: the correlation between FIP and BABIP at the team level actually declined a bit, from .35 in 2005-13 to .27 in 2014-2022. Not sure if that’s a significant change or not. Surprisingly, that decline in correlation should result in less variation in total defense.

#11 kgfella 2022/12/06 (Tue) @ 08:15

One hypothesis: tanking. When teams like the Orioles and Reds are content to field pitching staffs that legitimately vie for worst of the modern era and do so for 3-5 years at a time, the variation in run prevention will outpace run scoring. There’s a floor to scoring on offense (how bad you can be), there’s no theoretical ceiling to runs allowed on pitching (how bad you can be). As “rebuilding” becomes more acceptably mainstream for more teams at a time, it could partly explain why run scoring has less variation than prevention.

#12 kgfella 2022/12/06 (Tue) @ 08:20

One more hypothesis, perhaps related: position player pitching. Just about all teams are willing to do this with some frequency now, when games are effectively out of reach. Whereas historically it used to be extremely rare. And of course rebuilding teams who can’t even find five guys capable of pitching a bare five innings consistently, find their bullpens taxed like crazy and do even more of it (and are involved in more out-of-reach games in the first place).

#13 Tangotiger 2024/03/19 (Tue) @ 17:48

As a post-script to this:

Here is the year by year chart (click to embiggen) in the spread in scoring by offense and defense, where you can see the clear separation happening in 2014, not coincidentally when Shifts Attacked.
...
And it’ll be very interesting to see the results in 2023, when Infield Shifts are a thing of the past. That will more definitively give us the answer we’re looking for. We can wait to see how that plays out too.

And what happened in 2023? Standard deviation of 0.50 for both offense AND defense!

#14 jgf704 2024/03/20 (Wed) @ 22:52

Heck, I’ll jump in here too… I went back to my spreadsheet, and I actually calculated %DEF two different ways. The first was the way you did. But the second was: in each season, I fit Wpct to a linear model in RS per game and RA per game, i.e.

Wpct = intercept + OFF*RS.g - DEF*RA.g

And calculated %DEF = DEF/(OFF+DEF)

The plot below shows %DEF by the sigma method (blue) and by the regression method (green). The lines are 10-year trailing moving averages. The two methods are in… well decent agreement I suppose.

#15 Tangotiger 2024/03/21 (Thu) @ 11:55

Good stuff, thanks for running that

<< Back to main

Tangotiger Blog

Friday, October 21, 2022

Revenge of the Defense

Latest...