Filter posts by...
Run_Win_Expectancy
Thursday, July 15, 2010
Sean added Clutch to Play Index. And here you go. The greatest single-season clutch: David Ortiz 2005 at +3.4 wins followed by +3.3 wins of Eddie Murray 1985 and Mickey Stanley in 1968. The guys who saved all their best performances for when it didn’t count: Bill Mueller, 2003 at -3.5 wins, and Gary Gaetti 1983 at -3.3 wins. ARod’s 2002 was fourth worst at -3.2 wins. I just love when a stat verifies the obvious (Ortiz and AFraud), and gives you the expected (Eddie Murray), and so adds legitimacy when you get some surprises. Bobby Thomson’s 1953 comes in at -2.9 wins in Clutch.
And, the career leaders:
+13.4 wins Nellie Fox
+10.5 Tony Taylor
+10.0 Tony Gwynn
+9.3 Pete Rose
...
+6.1 Mark Grace (23th place)
+6.1 George Brett
...
+5.7 Ichiro (34th place)
...
+5.4 Tim Raines (39th)
This list goes toward what the fans were telling my in my Color of Clutch project: they love the guys who put the g-dd-man bat on the g-dd-mn ball.
And the guys whose stats were compiled when the game mattered the least:
-16.8 wins Sammy Sosa
-15.9 Frank Robinson <--- !!!
-13.0 Jim Thome
-12.7 Lance Parrish
-10.6 Mike Schmidt
-10.2 Richard Hidalgo
-10.0 Jermaine Dye
...
-8.9 ARod (12th place)
-8.9 Jeff Kent
-8.9 IRod
-8.6 Jim Rice (the man most feared ever)
Tell me you don't love this.
If we take Sammy Sosa, we we see that in the 20% of his PA that were in high leverage situations, he had a .264/.341/.479 line, compared to his career .273/.344/.534. Was this “luck”? Well, do we even care it was luck? It happened. His gaudy numbers did not come as often when the game mattered.
***
If you do care if it was luck, his performance in high lev situations was 2 SD from the mean. Given that this is the most extreme of all the players, it is almost certainly bad luck on Sammy, since *someone* is going to be 2SD from the mean just by luck alone. Still, that the best clutchers are populated by non-HR hitters and the worst-clutchers are filled with big swingers, it’s certainly not all luck. There is *some* clutch skill. In any event, we are not celebrating talent level, but performances. And Sammy Sosa’s numbers were very inflated relative to when his team needed him, and Nellie Fox came through like no one else.
***
Thanks Sean.
Wednesday, July 07, 2010
From Jack Moore, in looking for relievers with a worse than -1.00 WPA in a single-game. That was since 2002.
Sean Forman will give it to us for the Retrosheet era. There were 26 of them… TWICE by Rollie Fingers.... in the same year.
http://www.baseball-reference.com/boxes/OAK/OAK197409060.shtml
http://www.baseball-reference.com/boxes/CLE/CLE197407200.shtml
That’s what happens when you give up the go-ahead run in one inning, the tie-ing run in the 9th inning, and the eventual winning run in extra innings. He didn’t necessarily pitch terribly. It simply tells the story of a pitcher who ended up giving up a run at the wrong time, every time, late in the game.
It’s a story stat. And WPA makes it very easy, by quantifying the story, to find these games.
To those who think WPA is useless: just because you haven’t found a use for it, doesn’t mean that it has no use.
By , 09:30 AM
Regarding last night’s STL/COL game, in which the Rox were trailing by 6 runs in the bottom of the 9th:
“Fangraphs gave the Rockies just a 0.9% of winning this game coming into the bottom of the 9th.”
That seems awfully HIGH…
Wednesday, June 23, 2010
What do you do when the result of a positive act results in a negative value in win expectancy? e.g. Bottom of 5th, runners at corners, 1 out, home team up by 5: win exp = 98.6% (137/139). Next batter singles and runner on 1st goes to 3rd, making situation the same, but home team up by 6: win exp = 96.7% (87/90) (numbers from winexp.walkoffbalk.com - if there is another source, I imagine that somewhere in the list of possibilities that the above situation has occurred)
A win PROBABILITY is a probability, or an expectancy, or a future outcome.
What the reader is quoting here are TWO POOLS of past historical win rates. In one case, teams found themselves in one situation 139 times and won 137 times. In another case, a better situation to be in, COMPLETELY DIFFERENT teams found themselves in that situation 90 times and won 87 times.
You cannot simply compare the two, and come up with a delta win expectancy. This is why I am completely opposed to looking at win expectancy charts based solely on the empirical / actual / historical. You get hundreds, if not thousands, of such inconsistencies. Don’t do it. This is why Markov chains are far more preferable.
It’s important to make the distinction between past win rates of a bunch of unrelated game states, and future win probabilities of game states that are completely dependent on each other.
I hope I explained that well enough.
Somebody asked me about creating a pythag for a 7-inning game. Here’s how you can do it.
First, let’s look at this data, which shows that the home team in the 1980s win 54.2% of their games. I don’t have their runs scored and runs allowed (per 27 outs), but we can work backwards to figure that out.
If we give the home team 4.6 runs per 27 outs and the road team 4.2 runs per 27 outs, and we put it through PythagenPat (4.6 plus 4.2, raised to .283), we get exactly .5420. (Note: the .283 could be anywhere from .28 to .29… I just used .283 because it made the numbers work out cleanly.)
Now, how about for a 7-inning game? That’s easy enough: start the game tied in the third inning! (*) The home win% is .528. To match that, we need to set the pythagenPat exponent from .283 to .096, if we keep the runs as per 27 outs, or to .109 if we make the runs as per 21 outs. If you want a general rule, just make the exponent .100; and it won’t matter much if you put the runs as per 21 or 27 outs.
(*) Yes, this changes the lineup order, since if we start the game in the third inning, we are not guaranteeing that Rickey Henderson will lead off every game. I agree. But, you look for easy reaonable solutions where you can get them, and that’s what I’m doing here. If someone wants to try it another way, feel free.
Interestingly enough, if you start the game in the 4th inning, or the 8th inning, tied, you end up with the same home win%. This obviously makes no sense. The reason is because of the strategies employed, and personnel available. Meaning, bullpen usage. That, and sample size.
We can look at far more years here (that chart shows the win% for the batting team, not home team). Numbers make a bit more sense.
Friday, June 11, 2010
This is going to require alot of effort from you, to read and to follow. It’s a step-by-step kind of thread. And anyone out there that has ever disparaged, questioned, or just been plain flummoxed by wOBA or WPA/LI, well, this thread may be for you. I hope all your questions will have been answered after you read this. (Though it is possible that new questions you hadn’t considered will pop up!)
It took me about an hour to write this. I know I’m asking alot for your time and patience in return. The reward should be there. Let’s go…
Read More
Wednesday, June 02, 2010
From Ken Pomeroy.
Compiled a few years ago.
Wednesday, May 19, 2010
Two interesting threads.
H-H
H-A
A-H
A-A
“H” is home (or heads if you like). “A” is away (or ... well, a crude tail). If H=A=.50, then H-H plus A-A would be .50. If H=.6, and A=.40, then H-H plus A-A equals .52.
In baseball, if you win the first game, you probably had the better starting pitcher. That means in the second game, you probably have a worse starting pitcher. But you still have the home advantage. So, what might it look like?
If H1 wins, then chances are their presumed odds was say .560. If A1 wins, then chances are, their presumed odds was say .480. That is, given that you know that H1 wins, then there’s a decent chance they had an above average pitcher (and they might have been the better hitting team to begin with). On top of which, we also know that pitcher will not be available the next day.
So, H1-H2 = .56 x .53
A1-A2 = .48 x .45
Add the two together, and you get .513.
What were the empirical results?
The true answer: 51.3%, a little better than a coin toss.
I swear that I got lucky. I started with .540 for the home team, and .460 for the visiting team, and added +.005 for likely having the better hitting team and +.015 likely the better pitching team. That’s for game 1. For game 2, I used +.005 and -.015.
Tuesday, May 18, 2010
It’s the same old story, but for those who can’t get enough (like me for some reason):
Read More
Monday, May 17, 2010
If we think of wpaWins and wpaLosses, as I introduced them last week, as a step above a pitcher’s W/L record, but still worse than something else, does it make wpaWins and wpaLosses better? That is, once you have wpaW and wpaL, is there any reason, at all, to look at a pitcher’s W/L record? Here’s some food for thought.
Here it is from Boyd’s World. Could we have come up with something like that? Let’s see, if we try my Markov calculator. You really should put in the actual stats, but I’m going to take a shortcut and wing it. We know there are .63 runs per inning, so I’m going to change the AB value until I get that. Change AB to 34.7, and you get the RE chart for 5.68 runs per 9 inning MLB game. Ideally, you would set the HR and BB and SO numbers to match to actual college, but, I just want to do something quick. (Hint: someone ELSE do the work until then.)
Let’s compare what Boyd has to what Markov says. At bases empty, it’s pretty similar. It’s higher in mine, almost certainly because MLB has more HR than college. (Just guessing.) With one man on, the two charts are pretty similar. Indeed, the two charts are all pretty similar except:
Runners on 2b and 0 outs is much higher in college than Markov would say. Same thing with runners on 3B and 0 outs. And 2b, 3b, 0 outs. That is where the interesting thing looks. I’d like to know how college plays in those states.
If Boyd is out there, I’d like to see the batting lines by the 24 base out states.
***
Good article with a great database on bunts.
Sunday, May 16, 2010
This is a great way to use WPA and tell the story.
Thursday, May 13, 2010
Thanks to Colin for inspiring this post.
Have you ever asked yourself this question:
Read More
What Albert and Katie did on ESPN a few weeks back is exactly what I did with WPA: they used a pitcher’s earned runs allowed and inning pitched, and picked a threshhold to count as a good start. They did it simply enough: looked at how often a team won when a pitcher had a certain combination of runs allowed and innings pitched. They set their threshhold at anywhere where a team won at least 75% of the time for that combination (I would have selected closer to 60%, as that would more closely match the 1700 starter wins). They did not do the flip side (bad starts).
But, all this is is taking something we know that goes on a continuous scale (WPA, ERA, whatever), and turn it into something binary (W or L). That’s all it is. If you don’t buy ERA, you won’t buy what ESPN did. If you do buy ERA, then you’ll like what they did. Same deal with WPA for starters.
Friday, May 07, 2010
My answers below the fold:
Hi Tom,
I think the RE24 stat pretty much says it all in terms of measuring offensive performance. Very impressive.
I’ve tried to learn more about it and have a few questions I can’t find answers to:
1 - How are errors dealt with? Does a player get credit for advancing himself/runners on an error the same as a hit? Does the erred player or pitcher take the loss in RE?
2 - When a player steals a base, does he get creditted? Does the pitcher or catcher get docked? If the hitter drives in the base stealer, does the base stealer’s share any added RE or does all of the added RE go towards the batter after the steal? Same with Caught Stealing…
3 - For non groundout force double plays, is the base runner tasked with any loss of RE?
4 - And lastly, Couldn’t RE24 be adjusted at end of year to show the exact expectancies that occured in that single season by itself, instead of using broader averages? A yearly adjustment would keep all RE24 numbers in direct correalation with the actual season they performed in.
Thanks again for all the work.
Jason
Read More
Wednesday, May 05, 2010
It’s the bottom of the 9th, 2 outs, bases empty, down by 1 run. What are the chances the home team will win the game?
Well, we can look at the historical record and see that they win .035 times per game. We can look at a Markov model and see a .042 win%. You can even get your hands dirty and try to figure it out yourself. The chance of scoring 2 runs or more in that situation has historically been 2.2%. And the chance of scoring exactly 1 run (and sending the game into extra innings, with 50/50 shot of winning) was 4.5%. So, you get 2.2% plus half of 4.5% for a total of 4.45%, or .045 win%.
For purposes of this illustration, I’m going to use .044 chance of winning with bases empty, two outs, down by 1, bottom of the 9th.
Here are the chances of winning after each event, along with the frequency of each event:
freq win% event
0.665 0.000 out
0.100 0.090 bb
0.160 0.090 1b
0.045 0.140 2b
0.005 0.170 3b
0.025 0.530 hr
So, 66.5% of the time, you make an out, and the game is over. You hit a HR 2.5% of the time, and your chances of winning is 53%. And so on. The overall average, in this illustration, is .044 wins. So, if you hit a double, your chance of winning goes from .044 to .140, or a gain of +.096 wins. So, if Mike Sweeney hits a double, we assign +.096 wins for that particular double. In these cases, we are uninterested what happens after Sweeney gets the double. The next batter can make an out or a HR, and Sweeney is still assigned +.096 wins.
But, what if we DO care about the next batter? So, now we have a new chart:
freq win% event
0.665 0.000 out
0.260 0.044 Reach base, doesn’t score
0.050 0.383 Reach base, ends up scoring
0.025 0.530 hr
The weighted sum is still .044 wins. So, if Sweeney reaches base, but doesn’t score, he basically had no value. The chances of winning before he came to bat was .044, and so, his WPA (after the fact) was 0, which is why the win% assigned (after the fact) remained at .044. If he reaches base, and scores, we give him .383 - .044 = +.339 wins. What we did here was account for the fact that him scoring is more valuable than not, and it depends on the batter after him.
So, for those who think that we need to change the credit for getting on base, based on whether the runner eventally scores or not, this is how you do it.
But, do you really want to do this? Do you really want to decide whether to give Sweeney +.000 or +.339 based on whether he is eventually driven in or not by a future batter? Or do we simply want to give him credit for reaching on base, regardless of what the next batters do?
This is what you have to decide as a reader and analyst. You have to answer that question first. And once you answer that question, then I have given you the path that you need to follow to get the answer.
Tuesday, May 04, 2010
I don’t have a good word for it yet. “Fires extinguished” would be what I’m looking for. As Jeff notes, the idea of saves can be improved upon. So, what can we track?
How about any game in which a reliever enters (regardless of score, ahead/behind) and he comes out with a WPA of at least +.05, we call that a “douse” (or whatever better word you can think of). And if he leaves the game -.05 or worse, that’s an arson (or whatever better word you can think of). Last year, Mariano Rivera had 38 douses and 5 arsons. Just a matter of playing around with the threshholds to see what kind of results we’d like to get.
There’s a certain comfort in counting things in binary form like this.
Two points:
1. ignore if it’s sellable
2. we need a good name
Thursday, April 29, 2010
If you didn’t watch the game, this is what you missed. And now you know you missed something really good. All bow down to the Win Expectancy gods.
Just what the doctor ordered.
I’ve been meaning to construct a Markov program for football since, well, forever. I used to play this great sports board game called CFL Sports Action Football. Hands-down, the best sports board game I ever played. The layout was pretty good, where each QB and each runner had different ratings from a “1” play to a “5” play. The offense had to call the play as run or pass and the defense had to call the play as 1,2,3,4,5. If the defense called “1” (run defense) and you called a run, well, your yardage gain would be from -1 to +5 or something. If the defense called “5” (pass defense) and you called a run, well, your yardage gain would jump dramatically. And the passing works similarly, with an extra provision on the offense of calling a short, medium, or long throw. So, there was a strategy call on every play. Just beautifully well-done.
My thought was always: well, how the hell am I supposed to know what to do, other than by instinct? I was playing against my brother, and, he was in the same boat, so, it was instinct v instinct. That did not stop me later on to try to figure it out, by making some decent guesses as to how much yardage you should get based on the strategy of the opponent. The CFL put out this great little guide book (can fit in your pocket) that had alot of good data.
Anyway, I got a flashback to all that, and now Brian’s giving us the play-by-play data. I’m hoping to get to this at some point before the season starts. When you mix the three things I love to do (numbers,programming, and sports), I am in heaven. Just what the doctor ordered.
from Amazon
Recent comments
Older comments
Page 3 of 342 pages « First < 1 2 3 4 5 > Last »Complete Archive – By Category
Complete Archive – By Date