00:00:00
Welcome, welcome everyone to Wharton Moneyball, the show where sports, statistics, and business all intersect.
00:00:06
Three of my favorite topics. This is Eric Bradlo, professor of marketing and statistics and data science here at the
00:00:11
Wharton School. Some combination of myself, my co-host and friend today, Audi Winer, are here every week.
00:00:18
Sometimes it's us two, Cade Massie, Shane Jensen, but this week it's Eric
00:00:23
and Audi here on the Wharton Podcast Network. I think for those of you that have been on our show or been with our
00:00:30
show for the last 11 plus years know whenever it's me and Audi, I take this opportunity to in some sense interview
00:00:37
Audi. So today there will not be a guest. I will interview Audi for if you'd like an hour or so during our
00:00:43
podcast. The first part of the show will be our standard what caught your eye in
00:00:47
sports segment. Then the second half you're all in for a real treat. As I think many of you know, Audi, besides
00:00:53
being the co-f faculty director of Wasabi, this Wharton Sports Analytics and Business Initiative, which is part
00:00:58
of the bigger umbrella brand that I run, the Wharton AI and Analytics Initiative,
00:01:03
also it runs our faculty research with our undergraduates, and he's going to
00:01:08
talk to us about the research they're doing in sports analytics. So, the first
00:01:11
half will be what caught your eye in sports. The second part will be what Audi is doing with our brilliant Penn
00:01:17
undergraduates on sports analytics and research. Audi, how you doing today? >> I'm doing really well. Excited to have a
00:01:23
good conversation with you, Eric. I know that it's going to be a tremendous temptation for us to spend the entire
00:01:29
time talking about the Yankees, but we won't do that um at least not exclusively. And I look forward to
00:01:34
talking to you about the work that our students are doing, which are not not only includes uh undergrads, but also um
00:01:39
graduate students, our PhD students, as well as some students who are masters in
00:01:43
data science, and we even have a PhD or two from other departments. Well, I know
00:01:48
at one point I think Ryan Bro was AMCS or some students were AMCS. >> Ryan was AMCS. Um, he was in applied
00:01:55
math and now he's with the Utah Jazz. He actually came and visited. He he talked
00:01:59
about one of his um uh he talked about some of the the uh the difficult problems in in general.
00:02:05
>> Well, we'll get to that. We'll get to that in the uh second half of our show.
00:02:08
So, what I wanted to start with today was the NBA. So, something very interesting is
00:02:16
happening in the NBA. So, and it's maybe the most extraordinary start to a season that
00:02:22
I've ever seen. I would just like your take on it from a statistical perspective. A very specific question.
00:02:28
So, the defending champions, the Oklahoma City Thunder, they won the title last year, as you may
00:02:33
remember. >> I do remember. >> They are 17-1 to start the season. Right now, the forecast number of wins
00:02:43
for them is 68. Now, you might say, "Well, that's a huge number, maybe, except they're 17-1." If
00:02:52
you just projected that out, I'm not saying a linear projection. Multiply it
00:02:58
by three, they would be 51 and three in their next 54 games. Now, we all don't
00:03:04
predict that, but for them to beat that prediction of 68, they just need to go better than 51 and 13. So, let's be
00:03:13
clear. They're on a 51 and3 pace right now. 51-13 gets them to 68 wins. So, why
00:03:22
don't we take it piece by piece? Would you go over 68 knowing, by the way, you may remember the highest number
00:03:31
of all time is 73. That was the Golden State Warriors of I think 2017 who lost in the finals, you remember, to LeBron.
00:03:38
>> They were 73 and N. Of course, the Michael Jordan Bulls, I forget which of
00:03:42
the years was 72-10. I think that might be maybe there's one other at 70, but the next gap I know
00:03:50
there's a 69. So, let's just start with that. I've got two other things to say
00:03:54
about the Thunder right now. How would you help our listeners here on Wharton Moneyball on the Wharton podcast network
00:04:01
think about how likely is it above 68 and 14 or possibly even the record of all time?
00:04:07
>> Well, you know, there's two broad ways to approach this. One is the way you did
00:04:12
already, which is just look at what are the best seasons ever and essentially ignore the what we any individual
00:04:20
information we have about OKC. They've won 17 and lost one. Okay, we'll just
00:04:24
put that in the bank and then we'll use that um and we'll ignore that and we'll
00:04:28
just say how likely is it that they're going to be one of these top five teams
00:04:31
ever? And I guess 68 wins would put them what seventh, eighth best team ever. >> Yeah, probably exactly in that range
00:04:37
>> and considering that uh and that and that essentially asks that question in
00:04:42
that way and which case we are really not really thinking too much about the individual team. were just saying they
00:04:47
went 17-1. All the the all the all the data suggests they're they're potentially all-time great team. Where
00:04:53
would you put them? 68 seems reasonable. Um I I don't think that's where it came
00:04:58
from. My guess is the estimate came from some sort of basian updating, which is a
00:05:03
a tool that we've done a lot in this in this program. You essentially shrink or
00:05:08
regress to the mean um to estimate what you might call their true win rate. So there we don't believe they're really a
00:05:15
17 out of 18 team which is like a 93% win rate. We think they're something lower than that. And the question is by
00:05:22
how much do we regress down to the mean. And so we're essentially predicting that
00:05:26
they go what uhif what you say 51 and >> 51 and 13. >> So that's their like 75%
00:05:35
>> little 80%. So we're saying the rest of the way. So they're 93% win up until
00:05:40
now. The best estimate of what their true talent is is 80% and that leads to 68. That's probably what they did,
00:05:47
right? So would you would you be would you reasonably treat this as you know the classic beta binomial situation and
00:05:54
what I mean by that is we have a prior for OKC. We have a prior for every team. Now obviously this this part I can do
00:06:03
this it's not even math it's intuition. If they're 17 and one, which is 93%.
00:06:08
Okay. And the prediction is that they're going to go 80% the rest of the way.
00:06:14
>> The prior must be below 80%. Because if the prior were above 80% and the
00:06:21
likelihood were above 80%, which we know it's 93%, the posterior would have to be
00:06:27
a convex combination of those two, which means that the prior for the Oklahoma City Thunder might have been, let's even
00:06:33
say it was 75%. Which is not unreasonable. Maybe they were projected to be a 6061 win team, which is 75%.
00:06:42
They're 17-1. We're now up to a prediction as you pointed out of about 80% and that's our prediction for the
00:06:48
rest. That sounds about right, doesn't it? >> Yeah, it does. But I don't think so.
00:06:52
There's two ways to do this this basian this shrinkage. Do you shrink to a prior
00:06:56
that's specific to the team or do you shrink to the league overall prior and
00:07:00
the difference is a prior shrunk to the specific team has very has much smaller variance. Right? So if we are if we're
00:07:08
talking about OKC given what we know about everything what they did up until this season, we probably have a 75%
00:07:14
centrid maybe uh or posterior or prior mean and probably a pretty small standard deviation. Um and therefore um
00:07:24
we ignore we shrink pretty heavily despite the fact they got 17 and one. We still shrink pretty heavily back to the
00:07:29
prior. Just put it up a little higher. The other way that you could do it is just ignore, pretend this is the only
00:07:35
thing you know about this team is that it went 17 to1. That's it. You know nothing about the previous years. And
00:07:41
then you shrink to 50%. But your but your prior mean would be huge, much bigger. And so the current data would
00:07:48
end up getting much more weight. Um so you have two choices. You either shrink heavily to a very high mean or you
00:07:54
shrink. >> I'm shrinking towards the OKC one, but I do agree with you. Either one of those
00:08:01
conceptually depending on how much sample size you put in the prior could lead to the 80% number in this case.
00:08:08
>> Either one. It's funny how it's funny. It's interesting how to do this. It
00:08:11
>> is interesting. >> So, um I probably would would have had a a multivariant. I would have lumped
00:08:18
other teams into into that good team prior instead of just prior and try to borrow something else. Um but that's not
00:08:25
I don't want to try to do this myself. We're trying to predict them whether or
00:08:28
not. So essentially, I guess the prior one would get back get to an 80% going out. Um I think that um 68 I think
00:08:37
they're going to go over. That's my that's my point. I >> think so too. So there's something else
00:08:41
about them that's interesting. >> Right now their projection is to win the league by 10 wins. That
00:08:52
means they'll have 10 more wins than any other team. Now, that would be a historic
00:08:57
>> historic mark. Well, can I ask you one question? A particular question. I don't
00:09:00
know. Has their schedule been average, difficult, easy so far? What do we know? >> It's a good question. Um, somebody
00:09:07
knows. >> Yeah. >> This somebody doesn't know, but the answer is I don't know. That's a fair
00:09:13
question. I mean, how I mean, we're starting to get to enough that you and I
00:09:18
would agree. It's hard to believe it's much below 500, if it's at all below 500
00:09:23
of the teams they played, right? Because, you know, eventually it's eventually going to average out to a
00:09:30
certain schedule. >> Well, there's the problem is in the NBA there aren't that many great teams.
00:09:34
>> No, no, that's the problem, right? So, your point is there's no information.
00:09:38
Like, I hate to make this up, but they could be equivalently like five and one. Like, of course they beat those 12 teams
00:09:46
and then there's like six games. All right. You went five and one in in tough
00:09:50
games. All right. Well, all right. That's >> Now, they played How many games going
00:09:54
ahead do they have against really good teams? >> I haven't I haven't looked at that. And
00:09:58
by the way, >> another way to do the prediction, which is we actually build a simulator game by
00:10:05
game. We play that out and then it takes into account who they're playing, their
00:10:10
relative strength. >> That's that's the way to do it. The the difficulty in doing that simulation is
00:10:16
that you have to put in some sort of talent metrics for for everyone. And uh we've been talking about this on our
00:10:22
show with Cade on how to simulate going forward. Do you either put the uncertainty in the parameter estimates
00:10:30
before you start the sim or do you treat the the the parameters as fixed and then
00:10:36
you just reestimate them as you go through the sim. If you are fully basian that those
00:10:41
produce the same and are correct about the priors that produce the same answers. Um if you are not fully basian
00:10:48
if you just and you just want to add uncertainties to some say MLE then uh then you'll get very different results.
00:10:55
>> Agree with that. Let me point out something though that um people might do
00:10:58
though for practical reasons and then we'll move on. This is more of a technical comment here, but the first
00:11:03
one you mentioned where you fix the parameters but then add the simulation error later as you go through because
00:11:09
you up you rerun it kind of that one of course will um take potentially less computation because you've batch
00:11:18
processed the data and now all and you've generated a posterior and now you're just adding one observation at a
00:11:25
time which could there other ways all I'm pointing out is there could be computational iational advantage to do
00:11:31
sequential simulation but it also depends on what your algorithm is. So if you're if you have just some algorithm
00:11:38
that estimates team strengths >> oh then that's fine and that can be done
00:11:41
fast and then you're then you have no then there's no advantage. Just one
00:11:44
other thing about the Thunder before we move off the NBA. >> We talked about this last season in some
00:11:50
metrics they had the greatest season in history last season. Forget number of wins. Their point differential broke the
00:11:55
record. I think it was somewhere around 11.5. So right now, Audi, they're at 16.9.
00:12:02
>> Ridiculous. >> So at some point, we're going to have like, let's say they win the title again this
00:12:10
year. Let's say they win 70 plus games. Let's say they have a point differential
00:12:15
of 13 plus. We're going to have to start talking about this OKC team as being
00:12:22
it's not a dynasty yet. Two doesn't necessarily get you to some definition.
00:12:27
We're going to have to start talking about them as one of the greater teams.
00:12:31
I don't know the top 5% of teams all time. I mean, if this happens, is it? And but the trick is to me, yes,
00:12:38
of course we would. But my question is is a basketball one. What is the anomaly? is the anom because they don't
00:12:45
have is it the anomaly that uh she is really a great player like he's a LeBron
00:12:50
level do you know alltime Michael Jordan level and we've just underestimated him
00:12:54
or is it a team that's extraordinarily well constructed that has no weaknesses
00:12:59
that have that have found that defense is extremely important and while it does it's not flashy it's the way you make
00:13:05
big points >> if you're asking me as a basketball fan what I would tell you is they're an
00:13:11
extraordinarily well constructed ed team that Shay Gildish Alexander is the best
00:13:16
player or top two or three in the NBA right now, but he's not Michael Jordan
00:13:21
level talent or >> prime for or even Seth Curry at his prime or >> right probably not. But he might be in
00:13:26
that next tier down and then they've got a bunch of very very good other players
00:13:32
and so you take a great player and a bunch of very very good other players and you might get there and they seem
00:13:39
extremely well balanced. So either way, I don't want to spend all of our time
00:13:43
talking about the NBA. I was just noticing their records incredible. I glad we talked about some different ways
00:13:48
to do prediction, whether it's through shrinking to their prior, shrinking to
00:13:52
the league prior, a simulationbased prior, but also this point differential is just incredible right now. And at
00:14:00
some point we have to say, I understand it's only 18 games. Okay. So, next week,
00:14:03
Audi, if we're sitting here and it's 23 games and they've got a 16 17 point
00:14:08
differential, eventually we're going to start to have to say we'd be surprised
00:14:12
if they didn't break the overall record. >> Yeah. I mean, not to say I I don't have
00:14:16
the the number exactly off hand, but I believe the RMSSE on predicting wins given your point differential is three.
00:14:24
Meaning that I can predict your total wins just using your point differential to win at about plus or minus three
00:14:29
wins. But how well do you think that'll work in the tails? >> How do you think?
00:14:34
>> So the the data that I fit it to, right? So which is every season up until this
00:14:39
year doesn't seem to have a nonlinearity at the tails. It's it's not drifting.
00:14:45
>> Um >> there has to be some because there is an upper bound. You can only win so many
00:14:49
games, >> right? >> It's predicting. It just seems to I mean it's not it's I'm not predicting. So I
00:14:56
just predict wins. I mean, you could predict winning percentage and then do a logistic uh and then predict a logistic.
00:15:01
That's those are two I now we're getting really in the weeds here for our show. I
00:15:04
don't want to I don't want to get into this. Um but just to say that if they
00:15:08
are if they can hold 16, they are going to break that record easy. >> I think it's easily. Yeah. All right.
00:15:14
So, let's move on. Let's talk a little about the NFL. Now, of course, we could
00:15:18
talk about the Eagles game and all that, but I don't want to talk about that for
00:15:20
just a second. So, I just downloaded the uh ESPN power rankings for the NFL. Okay.
00:15:30
What team do you think is number one? >> Rams. >> That is incorrect, sir.
00:15:35
>> Is it still the Eagles? >> Nope. >> Colts. So, so far you said the Rams, they're
00:15:43
second. You said the Eagles, they are sixth. The Colts are five. Who am I missing here?
00:15:54
The the Patriots. [laughter] >> They're not even They're like 15 or 16.
00:15:59
>> Yeah. I mean, >> it's the Chiefs. >> The Chiefs still got them up there.
00:16:03
>> I I This is the part I This You have to expl So, I know this I know this is
00:16:09
current. Let me just be clear. I know this is current because it has their record at six and five. That's their
00:16:14
actual record. Now, >> it has them at 7.1 points above average. It has the Rams at six. So that means
00:16:23
they're a one-point favorite against the Rams. The Packers are third, but let's
00:16:26
say they played the Eagles. Right now, it has them as a 2.7. Let's say they played the Patriots. It has them as a
00:16:33
6.6 point favorite on a neutral field. Well, um what what are the underlying stats? I heard one stat that I'll share
00:16:43
with you. >> Okay. >> Um that uh I was trying to make some heads or tails of this. I saw this on uh
00:16:48
sports analytics Twitter. Um um someone pointed out that the the KC, the Chiefs have the highest uh um number of yards
00:16:59
per drive. It's a so basically how many yards are they averaging per drive? Um and they're
00:17:07
just about 40. And and they showed then the teams at that level historically go 12 and you know 12 and four or 13 and
00:17:14
three. Just so you know, by the way, I think you remember last year that Chiefs were something like 11-0 in one score
00:17:19
games. Just so you know, they're at least I I well, I guess it just changed up until this last week where they beat
00:17:25
the Colts, they were 0 and5. They're now one and five in one score games. >> One score games. So then so this this
00:17:33
particular stat suggests that this underlying metric which doesn't take into account special teams and and uh
00:17:40
things like that and and position starting positions. So, for example, you'll have more you'll expect more
00:17:45
yards on your drive if you start deep in your in your own territory. You have more more room to go. Um, and uh I don't
00:17:53
know whether that number means anything. It's like a peripheral that's correlates
00:17:56
with with team quality, but they were number one in it. And and usually teams that historically teams with that level
00:18:02
of success in terms of yardage have been have been great teams. So they they're
00:18:07
essentially that one number suggests that the underlying metric suggests that the KC is much much better than they
00:18:14
than they appear based on their record. They've lost a bunch of one-run games.
00:18:17
They've had some bad special team turnovers. They've had some what you might just jump into is and lump into
00:18:23
one big pile called bad luck. And bad luck is not supposed to continue. You should expect neutral luck. So, if their
00:18:30
underlying posh strength parameters are excellent, you might and given their historical performance and and of course
00:18:36
who their quarterback is and the fact that everybody seems to be beating up on everybody else, maybe they do land that
00:18:42
land that high. But I have to tell you, I'm surprised. I am surprised to by the
00:18:46
way, the only other team that's six and five in the top like 12 teams, not surprisingly, and this might be the role
00:18:54
of Prior. Well, who would you guess who's the other six and five team that the Priars are going to bring Ray up to?
00:19:00
Who? >> Bills. >> That's a great question. It's not the Bills. The Bills are seven and four, by
00:19:05
the way. >> Four. Okay. >> It's the Ravens. >> The Ravens, right? There's the Ravens.
00:19:09
Another one. They they they also have prior history. And I wonder what this is ESPN. The ESPN might have a lot of
00:19:15
weight on >> Well, that's why I was asking you because I think most people would find
00:19:19
it shocking that the Chiefs are the top. I think mo a lot of people might find it
00:19:23
surprising that essentially the Eagles and Ravens are equal. The Ravens are better than the 8 and3 Seahawks, the
00:19:30
seven and four ners, the 7-4 Bills. The 9 and2 Broncos are like 12th. >> The Patriots at 10 and two are like
00:19:41
15th. >> Yeah. Yeah. >> So either way, I thought it was just interesting to talk about that.
00:19:46
>> They made a dip twice probably. >> By the way, let me just say for all of
00:19:50
our sports, so as everyone knows, we record Wharton Moneyball on Tuesday. This is Tuesday, two days before
00:19:54
Thanksgiving. There's an amazing schedule on Thanksgiving, Audi. I don't know how
00:20:00
much you're going to get to watch. Um, the first game is Lions and Packers. I
00:20:04
mean, >> that's a great game. >> That's a big game. >> Lions traditionally play on
00:20:08
Thanksgiving. I guess there's more games than one now. They used >> the Lions always do, but there's they're
00:20:13
seven and four and the Packers are 73 and one. The second game all of a sudden now became fascinating. Audi Chiefs at
00:20:20
Cowboys. >> Wow. With the Cowboys beating the Eagles last year. >> So, six and five. I know with six and
00:20:27
five Chiefs against five five and one Cowboys. A week ago was like they're both going to lose. It's going to be two
00:20:32
losing teams. Now all of a sudden these teams are in it. And the night game is an interesting one because Joe Burrow's
00:20:38
coming back. It's Bengals at Ravens. So all of a sudden we actually have three
00:20:43
sort of interesting games on Thanksgiving Day. Hey, one thing I'm going to take advantage of because it's
00:20:48
you and me here. So I I want to talk to you about the Hall of Fame, the baseball
00:20:51
hall of >> course. How could WE NOT, ERIC? IT'S OUR DREAM. >> We're go. We're going to But okay, so I
00:20:58
want to remind everybody there's two separate committees that are going to be
00:21:03
voting this year. Okay. One is what's called the contemporary era committee which used to be called I
00:21:14
don't know if it was the oldtimers committee or anything but there's >> veterans committee. Was it just
00:21:17
>> veterans committ? No. No. But they've split it up. Remember Audi? There's 1980
00:21:21
onwards which is the contemporary era candidates and then there's pre980. So here are the seven candidates in the
00:21:30
contemporary era committee. Okay. Roger Clemens. So we agree he's and by the way you need
00:21:40
75% I think of like 16 voters. So not >> he's not going to get it. Okay. Carlos
00:21:48
Delgado. >> Carlos Delgado. I don't I don't see him getting it. I don't I wouldn't imagine
00:21:54
he deserves it either. How about you? What do you think? >> 473 home runs.
00:21:58
>> Yeah. >> Uh three time silver slugger. I don't know. Right on the border. Jeff Kent.
00:22:06
>> On the border also. You know, this is a Yeah. I mean, Kent, uh Kent, what was
00:22:11
he? His was an middle infield position, wasn't he? Kent. >> Second baseman. I think he has the I
00:22:16
think he has the most home Yeah, he has I it says here in my notes he has the most home runs ever by any second
00:22:23
baseman. >> I think he has a better shot than Dill Do >> Don Maddingley
00:22:29
by the numbers his career was too short. I just can't I can't I can't sanction
00:22:33
it. I mean he he definitely was the best hitter in baseball for about two to three years. um and super competitive
00:22:40
hitting you know 340s winning batting title uh you know 30s one year I think 145 RBI's a statistic we like
00:22:48
>> I remember that year >> like disparaged but is impressive great great fielding first baseman but just
00:22:53
didn't have the longevity I mean he just just you can't have o only if if you're
00:22:58
going to be in only on peak you have to have Griffy level peak Kofax level peak you can't
00:23:06
>> judge level peak >> judge level peak Trout level peak. Look at Trout. I mean,
00:23:10
>> yeah, Trout >> go into the Hall of Fame because of 10 year first 10 years of his career.
00:23:14
>> That's a long peak. All right. How about How about >> Dale Murphy?
00:23:18
>> Also, first baseman. I don't see I I don't see it happen. >> All right. Gary Sheffield.
00:23:25
>> Chef Chef was quite a hitter. I think his career statistics might be a little
00:23:28
higher. I don't have them in front of me, but >> 509 home runs. >> Yeah. And there's a majesty with the 500
00:23:33
homers, right? >> 1,676 RBI's. I think he's got the a good shot. >> Okay. Fernando Venuelo.
00:23:40
>> No. >> Okay. Just No. Okay. Well, those are the seven contemporary. So, you could see in
00:23:47
your mind, you wouldn't be shocked if Jeff Kent got in. >> No, I got you.
00:23:51
>> You would be shocked if Gary Sheffield got in. >> Nope. >> Okay.
00:23:54
>> I'm not I'm not I to listen to our listeners. I'm not staring at their
00:23:58
numbers and uh I just I'm just recalling what I imagine about these players. And
00:24:02
I do think Chef he had he had 500 homers. Um, >> that that's that's a mark that almost
00:24:08
always gets you into the Hall of Fame. >> Kent's leading home runs in a second
00:24:12
base position is really impressive. >> Don't remember if there were uh PhD
00:24:17
performance-enhancing drugs rumors about Sheffield. I think there were, which may
00:24:21
be wellkeeping, >> maybe why he's where he is >> because you would think someone with 500
00:24:25
home runs and almost 1,700 RBI's >> going to shoot in, right? >> Is in. I mean, that's in. It might be
00:24:31
third tier, but you're in. Let's talk about the current ballot. >> Ah, interesting.
00:24:37
>> Now, I don't think there's anyone coming in on the bot on the as first timers.
00:24:41
>> Yeah. So, they're not in my list. I looked at them and I just pruned them.
00:24:45
There's nobody. Nobody. Trust me, nobody. >> Mhm. >> It's a great year for people who hanging
00:24:51
on. So, >> let's talk about a couple hang Let's talk about a couple hangers on and let's
00:24:55
see if you think. So, Carlos Beltran, >> yeah, this is his year. He was he was
00:25:00
close last year with nobody. He's getting in. Yeah, this is here. >> Andrew Jones in his ninth year.
00:25:08
>> Uh, it's either this year or next year he's getting in. So, I would go for this
00:25:11
year because I mean, uh, what's the early reports? Do we have the early reports yet? I know.
00:25:16
>> I haven't even looked. I know you like to look at that. I have not.
00:25:18
>> You like to look at them? I I like to look at I like to look at the Hall of
00:25:21
Fame traffic. >> And that's probably it, right? Because of how far the other people like The
00:25:24
next person on the list is Chase Utley. He's >> Chase Utley, I believe, will make it
00:25:28
eventually, >> but he's in his third year at 39.8%. So he's not jumping to 70.
00:25:33
>> He's not getting anywhere close quite yet. Um so uh there aren't very many
00:25:38
ballots out yet right now. So um uh I I right now there looks like almost nothing out. So
00:25:45
>> okay. And then there's people I mean some other reasonable name you know look
00:25:48
other names that aren't getting in because of possibly performance-enhancing drugs. Alex
00:25:52
Rodriguez, Manny Ramirez, obviously they would be in based on their numbers. Andy
00:25:57
Pettit's in his eighth year won't make it. Here's a guy, you know, I've never
00:26:01
quite understood. If you talk about peak performance over What about King Felix,
00:26:07
>> he was great um for a short period of time. >> Well, I mean, extraordinary
00:26:11
>> question. Is Deg Grom in the Hall of Fame? >> Yes, Deg Gro's I think. I think
00:26:16
>> Well, let me see. I'm going to, you know, >> Deg Gro better than King Felix.
00:26:20
>> You know what? This is This is up my alley because I've done this research.
00:26:23
So, I'm going to give you I I can So, Deg Grom hasn't retired yet. So, um, I'm
00:26:29
going to I'm going to I'm going to pull them up. I This is Yeah, as many of you
00:26:32
know, I I worked with Ryan Brill. We created our Grid War metric, uh, which ranks all which ranks every starting
00:26:39
pitcher based on their their quality. Um, so I'm going to take a look to see
00:26:43
where Deg Grom falls. So the way our metric works is it take the um the geometric average of the rank in your
00:26:53
peak and your based on what? >> So your rank in so the grid war so our it's a long metric so your so I have a
00:27:02
annual grid war metric and you have your career grid war metric um so that that's
00:27:08
how we calculate it. So um and so the so if you want to just look our metric puts
00:27:13
Greg Maddox at number one. Can you just speak by the way can you just just for one second can you say to people the
00:27:18
advantage of taking a geometric mean as opposed to an arithmetic mean here like just take this why don't you just take
00:27:24
the simple average >> yeah you could do that the advantage of doing that is it doesn't if you have um
00:27:30
so the problem with that is that let's say take Sandy Kofax Sandy Kofax is one
00:27:34
in peak um peak uh rank he's number one um and uh he is 48 in career so if you
00:27:43
just took the the average you're not giving enough quality enough enough kick
00:27:48
to that number one. And so in other words, you you essentially it's a decaying term.
00:27:53
>> And the geometric mean puts more kick into peak performance. >> Wait. So the geometric mean for for
00:27:58
right. So so it just it basically says that if you're really really kind of low, you're just low. You're not. And so
00:28:04
if you're 48 or 100, it's not that different because it's it's the square
00:28:09
root metric. Um, and that's and it really values a high value and and it doesn't un doesn't destroy you with an
00:28:15
outlier on one of them. So, it's a little bit uh I think it has slightly better properties. Um, and by the way,
00:28:21
it it's extremely efficient. It it it's everybody's in until you get to a point
00:28:26
where everybody's out. So, it's uh >> you're telling me it has it has
00:28:30
literally a perfect set. >> No, it's not perfect. >> No, it has perfect. It has it has one
00:28:35
person, which we've talked about repeatedly, who's not in it. >> Not Kevin Brown. It's Kevin Brown. He's
00:28:40
the only exception. He's s >> You're telling Oh, wait. You're telling
00:28:43
me I just want to be clear for our listeners here on Wharton uh Wharton Moneyball by rank order person people
00:28:49
pitchers by your metric. There literally is a line I can cut where everybody except for Kevin Brown is in and
00:28:56
everyone below that number is >> not everyone. No, no, no, no, not every. There's a couple exceptions. Um, but it
00:29:02
pretty much it goes from almost everybody above except for Kevin Brown and almost everybody below is not, but
00:29:08
there are a couple exceptions. So, Phil Negro is uh is >> and Catfish Hunter, they're not at
00:29:14
exactly at the border. Perchilling is not I have to exclude those guys, those guys who didn't get in because of
00:29:20
reputation. But you go down and you get every single person all the way down the
00:29:24
list and to skipping over Kevin Brown. You get Max Cerver, he's Sebathia, he's
00:29:29
these guys are are in now. Sherzer's in. And then the one the one borderline
00:29:33
case, this is my favorite one, the actual border one, the guy who sits on the border who's nodding is Dave Steve.
00:29:39
>> Well, I know you've talked about Dave Steve so many times. Look, you've come
00:29:44
to the conclusion that Kevin Brown is more deserving than Dave Steve. I >> He's more deserving. Yes.
00:29:48
>> Okay. >> Yeah. >> So, is there a a planet that we live on where this veterans committee or
00:29:54
whatever they call current era committee like Kevin Brown or Dave Steep gets in?
00:29:59
>> Absolutely. the it really depends on how analytics focused they are, right?
00:30:03
Because the case for Kevin Brown and and for Dave rests on having extraordinarily
00:30:09
good peak performance and nice long careers. There's peak performances are for both of them are majorly undervalued
00:30:16
and they're undervalued because of win loss records. Well, particularly with Kevin Brown,
00:30:21
these years he play he pitched for the Padres's. He was a 500 slightly better
00:30:26
than 500 pitcher, but was the dominant pitcher. Let's make you make a prediction now. So, Eric Bradlo and his
00:30:32
three sons, I'm at Coopertown like I am. It's July whatever 24th or 25th next
00:30:37
year. >> You're telling me I'm going to see Carlos Beltron. >> You're telling me you think I'm going to
00:30:43
see Andrew Jones >> mentally. You're saying and I might see Jeff Kent and I might see Gary
00:30:50
Sheffield. >> That's right. That's about it. >> Okay. >> Well, I found King Felix. By the way,
00:30:56
King Felix is number 42 ranked uh and there's nobody ahead of him. There's a
00:31:02
whole lot of people ahead of him who aren't in. So, >> that were not in that are not in
00:31:06
>> were not in and and I don't think he's going to and if you just look purely
00:31:09
based on the on the quality, I don't think you're you're seeing it. >> I see. I see. But uh but you know he had
00:31:16
a lot of he had a lot people sort of loved him so you never know. >> I got to admit if you tell me I'm going
00:31:20
to Coopertown to see Beltron Jones Kent and Sheffield I'm thinking you know the
00:31:26
s you know I have tears of Hall of Fame. The sum of those four is 12. I don't
00:31:30
care what you tell me. Like there's no question. >> No I'm just saying that doesn't excite
00:31:35
me that much because >> No, it doesn't. >> I mean it's not it's not that exciting.
00:31:38
Well Audi, we've talked about NBA. We've talked about some NFL and rankings.
00:31:42
We've talked about some MLB. I'm glad to hear again about your metric with Ryan
00:31:46
Bril. Uh this has been the first half of Wharton Moneyball here on the Wharton podcast network. Stay with us after the
00:31:52
break and we're going to talk to Audi about uh him and his students and the re
00:31:55
research they're doing. So come and join us after the break. Welcome back to the second half of our
00:32:01
show here, Wharton Moneyball, the Wharton podcast network. This is Eric Bradler, professor of marketing,
00:32:06
statistics, and data science. I'm here with my colleague, co-author, co-author
00:32:09
and friend Audi Winer, professor of statistics and data science. Some combination of the two of us Cade Massie
00:32:15
and Shane Jensen are here every week on Wharton Moneyball. And as I mentioned at
00:32:19
the beginning of the show, one of the advantages, although we love it when everyone's here, one of the advantages
00:32:24
when it's just Audi and me is I basically get to interview him. And so we I just talked to him about some
00:32:28
statistical stuff having to do with the MLB, NFL, NBA in the first half of the show. Now I thought um let's pull back
00:32:36
the curtain on what AI has spent I don't know at least the last 10 plus years
00:32:40
maybe 15 plus years building which I consider the greatest undergraduate mast's MBA PhD level
00:32:49
research opportunity for people that want to apply statistics machine learning now AI data science more
00:32:56
broadly uh to sports research um so AI um I have no guide to you except you know why don't you start by telling us
00:33:04
one of the projects that you're working on now or recently that excited you and
00:33:08
we'll get to as many as we can in the second half of the show. >> So that's great opening just to give you
00:33:13
the listeners a little bit of introductions. We have uh we have many many students who are doing research in
00:33:17
in statistics, machine learning, computer science, mathematics and their their research is in some sort of sports
00:33:23
application which is a really um it's a exciting for them because they they're
00:33:28
they're close to the edge, right? And one of the way one of the way we introduce our in our seminar when
00:33:33
they're research is think about something that's that's that's caused
00:33:36
you to be think think about a problem. What what is what what is what are you curious about and and then try to get
00:33:44
the data. So we have lots and lots of projects going on. Most of these are are uh some of them already been published.
00:33:50
Um some of them have won prizes across different sports. So we have it surprisingly we don't have any
00:33:55
basketball but we have rug rugby a few in in football we have tennis um we have um we obviously we have baseball and we
00:34:04
have um in and question let me ask a question related to that. So, one of the things in the, you know, in our open
00:34:12
source publishing that we have now, let's imagine there's one of our listeners on Morton Moneyball that
00:34:17
wanted to either replicate or uh extend some of the work that you've done with
00:34:22
your students. Like, are these data sets public? And number two, if you publish the paper, do you also publish code and
00:34:29
data with it? And like, you know, which some journals require now, some don't.
00:34:33
like how would one of our listeners if they wanted to say I want to see how my skills are. I want to replicate winer
00:34:38
and brill or somebody else how would they do that? >> So um depends on the sport. So
00:34:42
everything we've gotten from football is either public through event data uh next
00:34:48
generation stats just public event data or from the NFL big data bowl. So the tracking data we've used that if it's
00:34:56
soccer there is public event data but our we have two soccer projects which I'll happily talk about in a moment that
00:35:03
comes from our partnership with PFF so pro football focus FC um you know their their soccer kind of arm they have
00:35:11
collected their own tracking data using video and they've let us have one season
00:35:16
and now we're about to sign on for a second season. I'm not sure we'll be
00:35:20
able to make that data public, but we will make code and everything related to the analysis um public. So, actually
00:35:25
that's a good place for us to start because we just submitted a paper. So, this is probably one of our highest
00:35:30
level teams because it includes um Jonathan Pippi who's a second year PhD student. It includes Tion Shu who's a um
00:35:38
uh who's a master's in data science and he's finishing his second year applying
00:35:42
to PhD programs right now. And it also includes Paul Sabin who's our our senior
00:35:46
fellow. and they've worked on this uh this this soccer tracking data to do something which is uh which they call XG
00:35:54
plus. Um so most people have heard of XG in soccer. XG is the expected goals across a game and so that way you get
00:36:03
credit for goals that aren't goals. So shots every shot has a certain probability.
00:36:08
>> Is XG typically computed at the player level or the team level? And well, it's
00:36:12
computed on a shot level and it's you can associate them with players if you
00:36:16
like, but this is reported at the team level. So, it's usually reported at the
00:36:19
game level >> and it's usually an underlying metric. So, if your if your team if if you you
00:36:25
can lose the game in F in XG but win the game in goals >> and that happens a lot. And so ex
00:36:31
>> would you since I'm trying to play as host here would you relate this somewhat
00:36:36
to what in the first half of the show you talked about about the Kansas City Chiefs like maybe like there's some
00:36:41
underlying metric you're in here it's XG plus before it was you know average
00:36:45
length of drive you know maybe that's a bet there's lots of things as we all
00:36:50
know that are better indicators necessarily of strength than the actual randomness that happens with outcomes
00:36:56
and wins and losses >> abs is exactly a parallel and this is a huge leap up in soccer evaluation
00:37:03
because there are very very few goals in soccer and a game can go one zero 0 with
00:37:08
a shootout and what are you going to do with that but xG can be can accumulate and in fact an individual player can
00:37:14
have substantial xG. Now usually in sports we compare actual goals to XG and we use that to attribute some special
00:37:23
quality um to the player as if they consistently do that. One of the things that we've learned in soccer is that
00:37:30
with maybe well certainly with one exception, maybe a couple others, very very few people seem to outscore their
00:37:37
xG. Their XG is just whatever is you create, by the way, you create your own XG. So that you get responsibility for
00:37:45
that. >> By the way, how give our listeners a sense of how that's computed? Is it I
00:37:51
would imagine you'll just tell me if I'm wrong. I would imagine um where you're
00:37:55
shooting from is part of this, right? >> The distance to the goal. Yep. >> Yep. I would imagine possibly. I don't
00:38:00
know. Is the angle >> angle? Yep. >> Okay. >> Absolutely. People were representative
00:38:06
by how much of the goal is visible, right? So you take your angle and they they can calculate how much of the how
00:38:12
much of of the 360 would be the goal, right? >> The location of the defenders.
00:38:17
>> Absolutely. They have a metric for that. I mean, it's complicated and it's and
00:38:20
these are often proprietary, although we built our own. um how how you actually take the defenders in in to create that.
00:38:26
But basically often what they do is is is essentially how much of the area is open, right? So
00:38:31
>> yeah, you can literally physically say if each player has a certain radius and
00:38:36
you have a certain angle and distance, there's a certain amount of openness and
00:38:39
there you go. You could >> So XG metrics can get very complicated because you could take a look at, you
00:38:45
know, what happened immediately before the shot and integrate that in as well. Like uh are you almost like shooting off
00:38:50
of the dribble? Are you shooting off of immediate pass? Um the Zaxis, the height
00:38:54
of the ball could be in it. Um but usually they are not primary drivers are defenders angle and distance and
00:39:03
distance is >> look all of our listeners are saying Eric when are you going to ask him
00:39:06
what's the secret sauce? So what's XG plus you guys? I mean it's got to be
00:39:10
better than just XG. You call it XG plus otherwise. >> All right. So, just to just to close the
00:39:16
book on XG, uh Messi is the only player who consistently outgo his xG and by a lot. Um he's just you can see that just
00:39:25
immediately. And in fact, this past Nessus or New England sports um um uh conference at Harvard, he just stuck out
00:39:32
in someone's analysis and that that everybody understands. Um okay, so what is XG+? So what our what our what our
00:39:39
team did is they asked the simple question XG is only calculated on shots taken and there are lots of times where a shot
00:39:47
can be taken but for variety of reasons isn't and that isn't account accounted
00:39:54
for in XG. >> Very interesting. So what they did is they they calculated at any instant what
00:40:00
is the probability of the shot being taken and then if that shot would take were taken at that point that would
00:40:06
generate an XG. So you could actually essentially integrate over over continuous time this this this quantity
00:40:15
which is your almost and in most instances the probability of a shot taking is is zero. So it doesn't
00:40:20
accumulate massively, but you'll end up missing all this opportunity if you only
00:40:25
look at XG as opposed to >> could I in XG plus could I penalize a player who should have taken a shot but
00:40:33
chooses not? >> Absolutely. >> That's right. And so PE so players who
00:40:37
have all this opportunity that they don't do, they'll get penalized in XG
00:40:41
Plus. and players who who who take the shots that they should take um you'll
00:40:47
not only see them in XG but they won't get penalized. >> This is now the reason I love this
00:40:53
besides it's interesting now you could have two different qualities of players.
00:40:57
It's almost like false negative false positive. Some players are false negatives. They should be shooting and
00:41:04
they don't. And some players are false positive. They shouldn't be shooting and
00:41:07
they do. So you could actually decompose someone's total XG plus I assume into
00:41:14
almost like false positives and negatives. >> You can do a lot with a metric and it
00:41:18
has a lot of the features that you'd like to see in a metric which is it predicts out of sample well it
00:41:24
correlates with lot it it passes the snip test. The great players show up. Uh this is lots of interesting produces
00:41:30
interesting results and they're just getting started with it and I don't you
00:41:32
know I didn't actually collaborate with the research at all. This is one of the
00:41:35
few projects that I had basically nothing to to do with it. They just did it and presented. But that large measure
00:41:42
is there might as you know a there are a lot of people that listen to us that are
00:41:46
academics like you and me. A lot of people in practice. Let me let me take it from each perspective.
00:41:52
>> First where does Jonathan Paul I forget the other third person's name on the
00:41:57
project. >> Yeah. Where do they try to publish this? >> All right. So that's interesting. Um so
00:42:04
the there are lots there's a whole bunch of sports analytics journals >> and there also statistics journals and
00:42:10
there in some sometimes you can publish in in in sort of operations research journals or econ journals or math
00:42:15
journals. Um so there are um the obvious candidate for a lot of research like this would be the journal of
00:42:22
quantitative analysis and analytics and sports which is probably the the most prestigious sportsoriented statistics
00:42:28
journal. Um there are lots there are others. There's a journal of sports statistics and there's there's a whole
00:42:33
bunch of sports analytics journals. Um it all depends on how substantive of your statistical contribution is. Um so
00:42:40
and how and and and what you're doing with that. So most of our um and so you
00:42:45
could also apply to you can publish in the annals of applied statistics in Jazza. Um these are these are very very
00:42:51
uh um elite journals in statistics. If you're making a methodological contribution to statistics as well as
00:42:58
just doing a really important insight into sports, one of the things that we always do, in fact, if you want to read
00:43:02
about XG+ without having to wait for the full publication of the manuscript, we have what's called uh Wharton Research
00:43:08
Notes on our on our Wasabi web page which has a a popularization of the content and you and they explain what
00:43:15
they've done without getting into the nitty-gritty of the maths of the method.
00:43:18
I also imagine I don't know I assume they may post it like on SSRN so there may actually
00:43:23
>> yeah so the archive is what we've been using that's that's usually the math
00:43:27
place for that you have any what do you think SSRN archive >> post stuff on SSRN but either one I as
00:43:34
long you know I'm always happy to post you know stuff through the review process and post it there and you know
00:43:39
>> yeah so you do SSRN um so we have lots of others um I um so actually >> tell me about another project
00:43:45
>> yeah so why don't we talk about one that I could do really quickly um one of our
00:43:48
one of our undergraduates um is is working on a a a rugby paper. So uh so this is really interesting. So rugby is
00:43:57
a one of the most understudied um sports. So his name is Kenny Watts and he's working with Jonathan Pippen. Um
00:44:04
and so you think rugby is like football, right? There should be a lot of analytics, right? Rugby football
00:44:10
football has analytics, rugby has analytics. Well, guess what? There's nothing. So what do you if I asked you
00:44:16
what do you think is the first and most important analytics paper in statistics goes back around 25 years um in football
00:44:23
what do you think it was >> wait what's the most in in football >> not in football what question in
00:44:30
football was first analyzed with analytics that was really influential >> um the question
00:44:38
>> yeah the paper I know of >> is does the kicker suck or was at the
00:44:44
distance. >> That's one. You know, that's not the one I was thinking of. Um I
00:44:50
>> would you go for it on fourth and one? >> That's one fourth down. Right. So, uh in
00:44:54
fact, it's called Do Firms Maximize and it's by Ror um and it's it's written as
00:44:59
an economics paper, but it does two things. It calculates an expected points model for football given down and
00:45:05
distance and and and yards to go. Um how many points are expected on that drive?
00:45:10
and he uses that to decide whether or not teams are going for it at the right rate on fourth down and he comes to the
00:45:16
overwhelming conclusion that they don't even remotely. So what Kenny and Jonathan did was they said can we do
00:45:22
that for for rugby? >> So explain to me I don't know even know enough about the rules of rugbys and
00:45:28
stuff like if you don't get a certain amount or don't score you have to give
00:45:32
the ball to the other team. I always thought as as long as you guys keep possessing the ball and you're going
00:45:37
forward on the field you keep the ball. >> All right. So the question that they
00:45:40
asked, which I never heard of because I don't know rugby, was that if a if some
00:45:44
sort of penalty happens, and I guess they happen a lot. >> Yeah. >> The team who the offensive team, the
00:45:49
team that has the ball that's subject to the subject to this penalty, >> they get to choose. They can try to kick
00:45:54
what you what amounts to a field goal, >> which I think is worth three points.
00:45:58
>> It is. >> Or they can kick it out of bounds and then they get the ball
00:46:04
essentially about 20 yards downfield wherever they kicked it out of bounds. And it's kind of like going for it
00:46:10
because then they can then score a touchdown which is worth six and then and then they have an extra point or two
00:46:15
um in in that formulation. >> Just be clear, your team has a penalty. I get an opportunity to either kick it
00:46:23
for three or kick it down field and then of course I may end up with zero. But of
00:46:29
course I also have the ball farther down the field. So I'm more like >> get six or seven.
00:46:33
>> That's the same thing. >> It's very similar. Seems >> exactly the same thing. And the question
00:46:38
became why had no one done this before? And the answer is data hasn't been available and the sport's far behind.
00:46:44
And so our our team Kent Kenny is uh knows rugby. Um and uh he he asked let's
00:46:50
do it. And they built an expected points model um very similar to the way the ROR
00:46:55
built his first one with linear regression taking every every play as a having a set of coariantss and and um
00:47:02
the outcome which are highly correlated and have to deal with that and calculated an expense points model and
00:47:08
then he basically calculated what is the optimal decision and turns out that just
00:47:13
as we've seen in almost every sport they are kicking too many field goals. >> That's what I assumed you were going to
00:47:18
say. >> Yep. There it is. So, that was a just it's a wonderful example because it's it
00:47:24
combines um a really important question that people want to know the answer to. >> Um we're talking about undergraduates
00:47:29
here. So, they're trying to flesh out techniques that are pretty standard at
00:47:33
this point and they apply them in a new situation and they tell us something that we almost expected to happen, but
00:47:39
now it actually quantifies it. Can you give us a sense of like how many points or how much win probability a
00:47:48
team is giving up? Like I always like you know since I try to be the effect size guy here.
00:47:53
>> So you know teams aren't optimal. Okay. But are they like you've pointed this
00:47:57
out even with your work with Ryan Bril like people don't go for it on forth not
00:48:02
necessarily always the right time but sometimes it's not irrational depending
00:48:07
like you could come up with a risk aversion story or you come up with an uncertainty story that explains it. Is
00:48:13
it that you know sure they're kicking too many field goals but they're basically not losing much win
00:48:20
probability or maybe they are. I mean how big an effect are we talking about? that's got to be something on their
00:48:25
docket. One of the things that when I deal with students is that I teach them pretty aggressively that it's important
00:48:31
to finish and not do everything. And >> oh, I thought you meant finish by a
00:48:36
different thing. So, let me I'm going to just This is not a story. >> What I mean is that they they they have
00:48:40
to they need to finish a b you you know this. It's never done. Right. >> Right. It's never done.
00:48:46
>> Never done everything. No, but the story I was going to point out is why, you
00:48:50
know, when I while I publish obviously in both statistics and marketing, in some ways marketing's just different,
00:48:56
not harder or easier, just different because let's say you built this statistical model for something. Let's
00:49:03
say kicking field goals in rugby or kicking it. Then someone would you could never publish that. You'd have to go to
00:49:11
the next step. And so what does this mean for firm performance? What does this mean for winning? like they will
00:49:17
never let you like I I thought you meant finishing by and so tell me what this means as opposed to you have to draw the
00:49:24
line somewhere and say you've built a good predictive model you've answered an
00:49:28
interesting substantive question you're not yet you know you could extend it and
00:49:32
say oh what does it affect when probability but at some point the project just ends if you're interested
00:49:36
in that answer it again >> that's a great question because actually the the getting an effect size I think
00:49:41
is something that needs to be done before they submit that that's too important a question. So they built a
00:49:49
model, they have a a conclusion, they're not they're not aggressive enough, and
00:49:52
now you have to kind of turn it into how many points you're giving up or something and which probably then turns
00:49:58
into wins pretty easily. So that piece is is I think is important. But you know, these models are going to be
00:50:04
incomplete. They're models, right? You're not going to be bringing every
00:50:07
factor in. And um you could be look doing looking at win probability models instead of expected points models. And
00:50:13
there's many ways to do these things. How you going to be treating the the the
00:50:17
how you dealing with the independence? Are you how you dealing with the standard errors? There's so many deep
00:50:21
questions that you could ask which eventually should get get answered but can't do them all in the first round.
00:50:26
>> By the way, Audi, I would have lost about a million dollars. And let me tell
00:50:29
you why. If you had told me that you and I were going to talk about related research, no. No. The first two
00:50:38
you would bring up would be soccer and rugby. I'd be like, I'll give you a
00:50:44
million to one odds. So maybe you could tell us about one that deals with the students that I have a little more
00:50:50
knowledge. No, no, I'm not saying there's anything >> but we have one in tennis which you I'll
00:50:57
just >> I want to I love tennis. So get to tennis. But before I do that, I want to
00:51:01
point out to our listeners this what Audi's pointing out is important which is I've always said I joke about it with
00:51:07
my friends. There's only one thing we as academics do really well and that's we
00:51:13
recognize the isomorphism between problems. So if you're not a rugby fan and you're not a soccer fan, who cares?
00:51:20
What Audi's talking about is applicable to so many other sports or so many other
00:51:25
problems. This is why sports is wonderful as a testing ground for statistical methods and learning because
00:51:32
it's, you know, you can change the name from, you know, kicking field goals and
00:51:38
scoring touchdowns in one sport to another or, you know, you have two options A and B and one is more certain
00:51:44
but has lower highend. You know, that's what I love about sports is that you can
00:51:49
learn a lot about business and business decision making through sports type models.
00:51:54
>> All right. So you're going to tell us about tennis. >> Uh so so tennis actually um this has
00:51:58
came up. These are this is a two two students um did this um they uh they put this together um and it's really it's um
00:52:06
and they they presented it and at Carnegie Melon um so let me just tell you who they are. It's I win Amita and
00:52:13
Audrey um Rita we call her and uh so I win Rita and Audrey um they uh they actually won the student competition
00:52:20
award uh for best paper at the Carnegie Melon Sports Analytics Conference. Great. And it was uh uh on tennis and
00:52:26
the feature of tennis was is um they wanted to rank serves and the problem the the problem with most ser service
00:52:32
ranks is that they don't actually measure the service quality as distinct from the player quality, right? Because
00:52:41
you just look how many service points you're winning. Now you can look at aces. Of course that's only one feature
00:52:46
of service. The other feature is getting the the opponent off balance enough to win the the the point fairly quickly.
00:52:52
but other ast as aspects of your play kind of get involved. And so it's essentially a decomposition of the um
00:52:59
the ser it takes you if you're winning a lot of service points, it breaks it down
00:53:03
and to how much of that's due to the quality of your serve and how much is due to the quality of your player. Um
00:53:08
and it gets a >> person serving or the opponent >> uh both of them and adjust for opponent
00:53:13
as well. That was a key key feature. You got you got for opponent. >> So does it have a ranking of player
00:53:20
>> rank serves and it and it and it's players >> uh current I think it has some of that
00:53:24
current players I I'll see in there I can look through their players and see
00:53:28
who they rank as excellent but I'm sure you would enjoy that right so that's
00:53:31
>> I would I would enjoy knowing but it's it is good to know because it's one of
00:53:35
those things where like simple metrics I have a prediction of who the best server
00:53:41
is but it's not going to be someone if if there is a table is there a table should I guess somebody
00:53:46
>> uh I don't actually have it so you can you can guess it for next time
00:53:50
>> I'm I'm gonna guess it's a lessernown player who is 6'10,
00:53:56
>> right? >> His name's Riley Opelka. >> I'm going to guess that he and you
00:54:02
remember from our childhood Rosco Tanner. >> Well, it was Rosco Tanner. Yeah.
00:54:06
>> No, no. I'm just saying Rosco >> 50 mph. Rosco Tanner. >> Rosco Tanner could serve. Rosco Tanner
00:54:12
could really serve. That was the only great. But either way, that's a fascinating
00:54:19
project. Um, so yeah, and that and those had great success and and it's funny
00:54:23
because I mean it's actually really really nice to see uh these students all
00:54:27
work with me or they were all Moneyball Academy TAs in part of our lab this summer. Um, and that's where they did
00:54:33
their work. Um, I can finish up with a few others. We have we have two actually, you know, pretty good ones in
00:54:38
football. >> Two more. Two more would be great. >> Yeah. Well, so one one of them in
00:54:42
football. We have we have a soccer, we have football, we have a baseball one that's just getting started that that uh
00:54:47
we just submitted to Saber >> question in the baseball one >> and the baseball question actually it um
00:54:53
so it has to do with my uh historical admiration of the RBI and and um so the RBI statistic would know everyone likes
00:55:01
to make fun of because it's contextual, right? It's it's so dependent on on the
00:55:06
the settings the opportunities you get. So that that that just essentially is just asking you to adjust for context.
00:55:16
So the question we asked was what are your RBI's above expected? >> That's very reasonable.
00:55:23
>> I I'd love to know like Lou Garri's 180 or >> Lou Garrick the thing this is what
00:55:29
you're pointing out. The greatest season I've ever thought about was when Lou
00:55:34
Garri hitund this 1927 the year Ruth hit 60. So, we know he came up at least 60 times with nobody on base.
00:55:41
>> He had 175 RBs. I think Ruth had 160 something. So, how many guys were goddamn left on base by the time Garrick
00:55:50
got up there? And he still hit 175, I think, is the number. >> So, to me, that that's an obvious, you
00:55:56
know, you always ask for, isn't that the one of the most obvious unanswered questions in baseball? If you report
00:56:01
RBI, you should also be reporting RBI's above expected immediately. And then you
00:56:06
can also um and this came from our conversations with some of the analysts at the Phillies. They want to know what
00:56:11
your RBI's above expected is after controlling for the quality of player that you are. So we know that right so a
00:56:19
lot of that RBI's above expected could be due to the fact that you are a great
00:56:22
player. So that's that's that. So they want to decompose the RBI above expected
00:56:28
into two pieces. one which is what you'd expect to get given the quality that you
00:56:33
are and then the additional part which was essentially be the luck or the um you can call it luck or skill to drive
00:56:40
in those extra RBI or pressure performance expected part I can imagine a bunch of I
00:56:47
can imagine a statistical model but I can also imagine just binning players into quality buckets or tersiles let's
00:56:54
say and then say of the players in this tersile based on some metric or something or some per, you know, skill
00:57:01
estimate. Let's look at how many RBIs above expected they have for their decile. That that would be one way to do
00:57:08
it. Would you favor that over some sort of let's call it parametric typo or some
00:57:13
sort of model? >> Well, I I generally my my general tendency is to use a parametric model
00:57:19
and then and essentially predict your RBI's given your context and your say WOBA. That's a simple way to do that. I
00:57:26
like your bucket. In fact, Eric, I'm I'm gonna ask you. You want to be a you want
00:57:29
to be the faculty adviser to our to our little team? They'd certainly love to
00:57:33
have you. >> I'd love it. Let's do it. >> So, we have we have we have two
00:57:36
undergraduates working on it right now. Talia and uh Lev are are sophomore and freshman. In fact, Leev um took
00:57:43
Moneyball Academy and he's now here as a freshman and is already jumping in to do
00:57:47
research. Um I'll finish off with a football um in Harvard um and did a great job. They
00:57:54
did something that everyone is interested in. You know, we we look at wind pass uh you know that block block,
00:57:59
you know, brush win rate, right? That's a standard statistic that is get calculated. But guess what it's not
00:58:04
doing? It's not adjusting for um quality, the quality of your opponent. And that is something they want to do
00:58:14
and they so this team has tried to adjust for the quality of the opponent and the context and the actual what's at
00:58:20
stake, right? So failing to block in a situation that doesn't really matter is
00:58:24
much less important or damaging than failing to block or succeeding to block and that leads to a pressure or sack. So
00:58:31
essentially they want to do is a much more sophisticated evaluation of both the blockers and the and the and the um
00:58:37
the rushers and they come out with great great results and it's much better than
00:58:40
just you know success rate block you know black uh block pass success rush win rates. Well, the reason I love all
00:58:48
these stories is first of all, um, it's great that you're giving back to our
00:58:52
students and giving them an opportunity to do research. That's obviously number
00:58:55
one. That's great. It's great. And a lot of these people, as you said, may go on
00:58:58
for research jobs in industry or PhDs, etc. >> Well, they're applying for statistics,
00:59:03
PhD departments, >> I mean, that's that's incredible. Um the second thing is it shows people that um
00:59:11
you know you don't need a PhD to do interesting research too. And so that part is is interesting. That part is is
00:59:21
really interesting as well. Well, this has been one hour here of Wharton Moneyball on the Wharton podcast network
00:59:27
on behalf of myself, Eric Bradlo, my colleague, co-author and friend Audi Winer uh in Absentia, Cade Massie and
00:59:33
Shane Jensen who'll be back next week. Thanks to our associate producer and sound engineer, Dion Simpkins. Thanks to
00:59:39
our producer and if you like our big bosses, Marissa Rena and D Patel. Thank you for joining us here on Morton
00:59:46
Moneyball between now and next week. Enjoy your sports. Enjoy your statistics. We'll see you back here on
00:59:51
the Wharton Podcast Network.