Thursday, 5 September 2013

The Win is Dead! Long Live the xWin!

I was reading Joe Posnanski's blog, as usual, when I went through this article and came across a comment by one of his Brilliant Readers, Stephanie. Which got me thinking that it's time for a blog post.

The article is about the building support which has been apparently formalized into killing the win as a statistic because it's dumb. But maybe we should still keep it, because everyone knows the win and its a very easy to understand statistic. 20-5 is a great season. 11-13 is a mediocre season. 3-12 is a bad season. Ok.

I've already been over this, but although wins and losses are certainly correlated with a good pitching season, they can be incredibly misleading. So I am proposing an alternative, a way to turn a very good and robust advanced metric into a win loss record. For now, let's call it xWins and xLosses, or Wx-Lx.

First, a look at old wins and losses.

To get a win, a starting pitcher must
- pitch at least 5 innings
- leave with his team in the lead
- his team never gives away the lead and wins the game

To get a loss, a starting pitcher must
- start the game
- leave the game with his team losing, and his team never comes back to even tie the game

Ok. That's a weird statistic, right? Relief pitcher wins are even worse. For example, you could pitch 1/3 of an inning, give up 5 runs so that your team is tied or losing, but finish the inning. Now have your team mount a comeback in the next half inning, get pulled out of the game, and pick up the WIN.

There is a stat called RE24, and it works for hitters and pitchers. It is similar to WPA, which I have covered already in this blog. Basically, there are 8 base-runner situations that are possible:
1. Nobody On
2. Man on First
3. Man on Second
4. Man on Third
5. Man on 1st and 2nd
6. Man on 2nd and 3rd
7. Man on 1st and 3rd
8. Bases loaded

There are also 3 out situations (0 1 or 2). 8 times 3 is 24 possible base/out states.

Smart people have figured out that there is a certain "run expectancy" for each of the base/out states. With nobody out and the bases loaded, you expect to score way more runs than with 2 outs and the bases empty. When the hitter/pitcher complete a play, they get credit for the positive/negative outcome of that plate appearance. Good players accumulate positive credit.

With every action by a hitter or pitcher, the state of the game changes. Runners on 1st and 2nd with one out? Expect to score 0.97 runs, on average. If the batter strikes out, the run expectancy changes to 0.47. The batter receives -0.50 RE24. The pitcher receives +0.50 RE24. If the batter hits a home run, the next state is one out with the bases empty (0.30 runs) - but three runs have scored (3.00 - 0.97 + 0.30 = 2.33). The hitter gets +2.33 RE24, and the pitcher gets -2.33 RE24.

WPA cares more about how the actions affect the probability of the outcome of the game, so actions matter much much more late in tight games. RE24 only cares how actions relate to the probability of scoring runs, so a run early on in a blowout is worth just as much as the winning run in the bottom of the 12th.

Fangraphs and Baseball-Reference seem to calculate RE24 differently, which is too bad. I will use the bbref version for all of this.

Miguel Cabrera is currently leading MLB with 73.9 RE24, or "base out runs added". Chris Davis and Mike Trout are 2nd and 3rd with 66.3 and 66.1, respectively. These are huge totals, by the way - all of these guys are having monster seasons. Last year was led by Edwin Encarnation with less than 60. Cabrera has the 27th highest total since 1980, so it's a great season by this measure.

Okay. That's the stat. Wasn't I writing about pitcher wins?

So, Stephanie suggested that we normalize some of our more useful stats like WAR into a form that is more digestible by people who understand what a win loss record means. They don't really know or care that a 8.4 WAR season is awesome. They know* that a 20 win season is awesome.


I propose this:

RE24 is in the form of runs - that is kind of abstract. Well, those same smart people can transform RE24 into Base Out Wins Added by normalizing to the run scoring environment of the era (for different scoring eras, different numbers of runs translate into different numbers of wins). So Miguel Cabrera's 73.9 RE24 is worth 6.9 REW (Base Out Wins Added). Clayton Kershaw's MLB leading 41.0 RE24 is worth 4.9 REW.

RE24 is composed of both positive and negative components, which kind of cancel each other out. The websites give WPA+ and WPA-, but they do not give me RE24+ and RE24-, so I will do the best I can. I want to transform RE24 into a win loss record that is about the same magnitude as real win loss records.

Here is the formula I am proposing, at least to start:

Wx = (Innings Pitched / Innings Per Start) / 2 + REW
Lx = (Innings Pitched / Innings Per Start) / 2 - REW

That is pretty simple, I think.

Innings Per Start (IPS) should be the same for everyone - you don't get extra credit for short starts. Let's be tough and make it 9.00 for everyone.

When Innings Per Start is 9.00, Innings Pitched / Innings Per Start basically gives the equivalent number of complete game starts the pitcher made. Dividing it by 2 gives a .500 Wx-Lx record. That is, if REW was zero for a pitcher, that pitcher would have a .500 Wx-Lx record. REW adjusts the .500 record so that a good pitcher has more wins than losses, and a bad pitcher has more losses than wins.

But does it work?
You tell me.

Top 20 in REW for 2013 (at least 10 GS):

Middle 20 in REW for 2013:

Bottom 20 in REW for 2013:

I think it works pretty well.

Let's move away from 2013. What do full seasons look like? Do all-time great historical seasons get the proper treatment? When Felix won the Cy with 13 wins - what was his Wx-Lx record? Did Pedro and Randy Johnson and Greg Maddux and Sandy Koufax and Steve Carlton put up huge Wx-Lx records?

Here are the top 20 REW seasons since they started calculating the statistic:

Two Pedros, three Madduxes, three Koufaxes, two Seavers, Gibson in '68, Carlton in '72... and of course the Doc's ridiculous 1985. 276IP at 20 years old would just not happen anymore.

And here are the worst 20 REW seasons:

Oh right, last two. When Felix Hernandez won the Cy in 2010 with a 13-12 W-L record, his Wx-Lx was 19-9. That makes more sense. On the other hand, when Justin Verlander won the Cy and MVP in 2011 with a sparkly 24-5 record, his Wx-Lx was a slightly less impressive 20-8.

So, what do you think?

No comments:

Post a Comment