Today's episode takes us into the dynamic

intersection of Bayesian statistics and

2

sports analytics with

3

Patrick Ward, the Director of Research and

Analysis for the Seattle Seahawks.

4

With a rich background that spans from the

Nike Sports Research Lab to teaching

5

statistics, Patrick brings a wealth of

knowledge to the table.

6

In our discussion, Patrick delves into how

these methods are revolutionizing the way

7

we understand player performance and

manage injury risks in professional

8

sports.

9

He sheds light on the particular

challenges of translating

10

complex Beijing concepts for coaches and

team managers who may not be versed in

11

statistical methods but need to leverage

these insights for strategic decisions.

12

Patrick also walks us through the

practical aspects of applying Beijing

13

stats in the high -stakes world of the

NFL.

14

From selecting the right players to

optimizing training loads, he illustrates

15

the profound impact that thoughtful

statistical analysis can have on a team's

16

success and players'

17

For those of you who appreciate the blend

of science and strategy, this conversation

18

offers a behind -the -scenes look at the

sophisticated analytics powering team

19

decisions.

20

And when he's not dissecting data or

strategizing for the Seahawks, Patrick

21

enjoys the simple pleasures of reading,

savoring coffee, and playing jazz guitar.

22

This is Learning Bayesian Statistics,

23

Welcome to Learning Bayesian Statistics, a

podcast about Bayesian inference.

24

the methods, the projects, and the people

who make it possible.

25

I'm your host, Alex Andorra.

26

You can follow me on Twitter at alex

.andorra, like the country.

27

For any info about the show, learnbasedats

.com is Laplace to be.

28

Show notes, becoming a corporate sponsor,

unlocking Bayesian Merge, supporting the

29

show on Patreon, everything is in there.

30

That's learnbasedats .com.

31

If you're interested in one -on -one

mentorship, online courses, or statistical

32

consulting,

33

Feel free to reach out and book a call at

topmate .io slash alex underscore and

34

dora.

35

See you around folks and best patient

wishes to you all.

36

And if today's discussion sparked ideas

for your business, well, our team at Pimc

37

Labs can help bring them to life.

38

Check us out at pimc -labs

39

Hello, my dear patients, I have some

exciting personal news to share with you.

40

I am thrilled to announce that I have

recently taken on a new role as a senior

41

applied scientist with the Miami Marlins.

42

In this position, I'll be diving even

deeper into the world of sports analytics,

43

leveraging Bayesian modeling, of course,

to enhance team performance and player

44

development.

45

And honestly, this move is so exciting to

me and solidifies my commitment to

46

advancing the application of Beijing stats

and sports.

47

if you find yourself in Miami or if you're

curious about the intersection of Beijing

48

methods and baseball or team sports in

general, don't hesitate to reach out.

49

OK, back to the show now.

50

Patrick Ward, welcome to Learning Bayesian

Statistics.

51

Thanks for having me.

52

I listen to every episode.

53

I think every year at the end of the year,

Spotify tells me that it's one of my

54

highly listened to podcasts.

55

So it's pleasure to be here.

56

Hopefully.

57

I don't know if I can live up to your

prior.

58

You've had some pretty big timers, but

yeah.

59

No, yeah.

60

So first, thanks a lot for being such

61

faithful listener.

62

I definitely appreciate that.

63

And I'm always amazed at the diversity of

people who listen to the show.

64

That's really awesome.

65

And also I want to thank Scott Morrison,

put us in contact.

66

Scott is working at the Miami Marlins.

67

He's a fellow colleague now.

68

That's change for me, that's great change.

69

I'm extremely excited about that new step

in my life.

70

But today we're not going to talk a lot

about baseball.

71

We're going to talk a lot about US

football.

72

So today, European listeners, when you

hear football, we're going to talk about

73

American football, the one with a ball

that looks like a rugby ball.

74

And so Patrick, we're going to talk about

that.

75

But first, as usual, I want to talk a bit

more about you.

76

Can you tell the listeners what you're

doing nowadays?

77

So I gave your title, your bio in the

intro, but maybe like tell us a bit more

78

in the flesh what you're doing and also

how you ended up doing what you're doing.

79

Yeah.

80

Well, currently I'm at the Seattle

Seahawks, which is one of the American

81

football teams in the NFL.

82

And I'm the director of research and

analysis there.

83

So we kind of work across all of football

operations.

84

So everything from player acquisition,

front office type of stuff to team based

85

analysis and opponent analysis.

86

And just kind of coordinating a research

strategy around how we attack questions

87

for the key decision makers or the key

stakeholders across

88

coaching, acquisition, even into player

health and performance and development and

89

things like that.

90

And I got here, this is my 10th year, I

got here from Nike.

91

So I was at Nike in the sports research

lab actually working for nearly two years

92

as a researcher.

93

And the way that I got

94

was I was doing some projects for Nike

around applied sports research and they

95

had just at the time, I think they had

just become like the biggest sponsor of

96

the newly minted National Women's Soccer

League.

97

And they said, we want to do something

around this.

98

And so, you we were kind of kicking around

ideas.

99

And one of the ideas we had was what if we

went out and we tested all of the women in

100

the league, like tested them

101

sprinting and jumping and power output and

things like that.

102

And then we could basically build like

archetypes and that would be useful for,

103

you know, like apps in your watch and on

your phone and girls could in the field

104

could compare themselves to their favorite

athletes and stuff.

105

So they let us do it.

106

And they sent me on the road for an entire

off season, the entire off season training

107

of the national women's soccer league.

108

went around the country to every single

team, myself and four colleagues, and we

109

tested.

110

every woman in the league.

111

And so we had the largest data set on

women's soccer players that anyone could

112

have.

113

So we did some conference presentations

and things like that with that data.

114

And lo and behold, Nike was there and the

Seattle Seahawks called down to Nike and

115

said, hey, we hear there's this test

battery and we'd love to see what our

116

players do on it.

117

And so I went up and I did a project for

them around that.

118

And

119

then they kind of just said like, what if

you just did this kind of stuff all the

120

time?

121

And so that's how I started out 10 years

ago.

122

And I basically started out just in

applied physiology, which was my

123

background.

124

And I was doing like wearables, wearable

tech for the team, like GPS and

125

accelerometry and things like that.

126

And then that kind of progressed into

draft analysis and player evaluation and

127

things like that.

128

And it just kind of growing until

129

Yeah, 10 years later, here we are.

130

Yeah, that's a great, yeah, it's it's a

great, uh, background.

131

love it because I mean, definitely it

seems like you've been into sports since

132

you are, uh, at least a college graduate,

but also there is a, uh, a bit of

133

randomness in this.

134

So sorry, love that.

135

Uh, of course, as a, as a fellow Bayesian,

always, always interested in.

136

in the random parts of anybody's journey.

137

Actually, how much of the Bayesian stats

do you have in that journey and also in

138

your current work?

139

How Bayesian in a way is your work right

now, but also how were you introduced to

140

Bayesian stats?

141

Well, mean, anyone who has watched

American football knows it's a game of

142

very, very small sample sizes.

143

So we only play up until two years ago.

144

We play 17 games now.

145

We used to only play 16 games.

146

So unlike most of the other sports,

baseball has got 162, several hundred at

147

bats, basketball, hockey, 82 games.

148

many attempts.

149

Also the players in a lot of these sports

are all doing the same things in baseball.

150

Aside from the pitchers, everybody's going

to go to the plate and hit in basketball.

151

Everybody has a chance on the court to

dribble the ball, shoot, score, pass, get

152

assists, get blocks, et cetera.

153

Football is really unique because it's a

very tactical game.

154

There's discrete events in terms of plays,

stop and start.

155

But because of the tactical nature of it

and one ball, there's only certain

156

positions that touch the ball.

157

There's only certain opportunities that

players are going to have.

158

So that was always an issue.

159

And when I did my PhD, a big part of my

PhD was using mixed models to look at

160

physiological differences between players

on the field with GPS and accelerometry.

161

And I always thought of mixed models, even

though I didn't know it at the time,

162

because I hadn't really...

163

learned anything Bayesian yet.

164

I always thought of mixed models.

165

I think of them as like this bridge to

Bayesian analysis because you have these

166

fixed effects which behave like our

population averages, our population base

167

rates, I guess you could say.

168

And the random effects are sort of like,

hey, we know something about you or your

169

group and therefore we know how you

deviate from the population.

170

And then with those two bits of

information, we're also like, hey, here's

171

someone new in the population, or maybe

someone that we've only seen or observed

172

do the thing one time.

173

best guess therefore is the fixed effects

portion of this until proven otherwise.

174

So I always had that in the back of my

head going through this, but you know, my

175

first two or three years in the NFL, we

always just used to kind of throw our

176

hands up when we see small samples, we'd

be like, yeah, it's this, it's

177

50%, but it's such a small sample, we

can't really know.

178

And we didn't really have a good way of

like sorting out what to do with that

179

information.

180

Because as you know, know, something like

one out of 10 and 10 out of a hundred and

181

you know, a hundred out of a thousand,

those are the same proportions, but

182

different levels of information are

contained within those proportions.

183

And I stumbled upon a paper, it was like a

19, I think it was like 19.

184

77 or something by Efron and Morris and it

was called the Stein's paradox and I

185

probably stumbled on it because I was like

You know, there's so much in Saber Metrics

186

someone in baseball has probably figured

this out before and so I I was probably

187

googling something like Small samples

baseball statistics Saber Metrics blah

188

blah blah and I stumbled upon this paper

about Stein's paradox and The crux of the

189

paper was if we observe these I think it

was

190

12 or 18 baseball players through the

first half of the season up to the All

191

-Star break.

192

And we see the number of times they went

to the plate and what they're, you know,

193

and the number of times they hit, we have

a batting average.

194

If we take the observed batting average

through the first half of the season, how

195

well does that predict the batting average

at the end of the season?

196

Meaning now they've gone through the

second half.

197

And

198

You look at that and you're like, okay,

let's, you know, what's this all about?

199

And so the first thing they do is they set

up this argument that like, well, that

200

doesn't do a very good job because some of

these players batted, you know, five times

201

or three times, certainly a player who

went three for three has a hundred percent

202

batting average.

203

We don't think this is the greatest

baseball player of all time yet, because

204

we've only seen them do this thing three

times.

205

So, the basic naive prediction of using

the half

206

first half of the season to predict the

second half wasn't very good.

207

And so in that paper, they introduced this

kind of simple Bayesian model of saying,

208

well, we know something about average

baseball players.

209

What if we weighted everybody to that?

210

And lo and behold, that did a bit better

of a job constraining the small sampled

211

players who had these, you know, a guy

that goes 0 for 10, which is totally

212

possible in baseball when you have

hundreds of it bats.

213

We don't think that's the worst hitter in

baseball.

214

so, you know, constraining those players

told them something about what they

215

expected to then see at the end of the

season.

216

And so through that paper, then I found

this blog by David Robinson, who's an R

217

programmer.

218

And it was all about like using empirical

Bayesian analysis for baseball.

219

And then he made it into a nice little

book that you could buy on Amazon for

220

like, I don't know, $20 or something.

221

You know, and I read those two things and

I was like, this is incredible.

222

This is exactly what I've always wanted to

know.

223

And so like I went in the next day to our

other analysts at the time, there was only

224

two of us.

225

And I said, I think I figured out a way we

could solve small sample problems.

226

And, and that was it.

227

Like then after that, you really couldn't

convince me otherwise that this wasn't a

228

great way of thinking.

229

That doesn't mean that everything we do

has to be Bayesian.

230

Certainly like there's other things that

we do that are used.

231

you know, different tools like machine

learning models and neural networks and

232

things like that.

233

But certainly when we start thinking about

like decision -making, how do I

234

incorporate priors, domain expertise?

235

How do I fit the right prior?

236

You know, like if you went 0 for 5 and

you're first at bats, let's say in

237

baseball, but you were a college standout

and you were an amazing

238

player in the AAA, I probably have a

stronger prior that you're maybe a

239

slightly better than average baseball

player than if you went 0 for 5 and you

240

were a horrific college player and you

weren't very good in AAA and you were

241

really the last person on the bench that

we needed to call.

242

And maybe that prior is much lower.

243

so utilizing that information in order to

help us make decisions going forward,

244

that's really

245

That was kind of the money for me.

246

And so how much do we use it?

247

mean, if we have a new analyst start, one,

you know, one of our new analyst starts,

248

started two years ago.

249

I think the first thing was like, how much

do you know about BASE?

250

And it was like, well, I never really

learned that in school and blah, blah.

251

And it was like, okay, here's two books.

252

Here's a 12 week curriculum.

253

We're going to meet every week and you're

going to do projects and homework and

254

reading.

255

And that was it.

256

Like, it was like, you have to learn this

because this is how we're going to think.

257

And this is how we're going to,

258

process information and communicate

information.

259

Well, what about that?

260

I told the listeners that we were not

going to talk a lot about baseball, but in

261

the end we are.

262

It all comes back to baseball, think.

263

Yeah, in sports analytics, all comes back

to baseball, Certainly, yeah.

264

Yeah, okay.

265

If I understand correctly, was motivated a

lot by low sample sizes and being able to

266

handle all of that in your models.

267

That makes a ton of sense.

268

As a lot of people, I've seen a lot of

clients definitely motivated by a very

269

practical problem that you were having.

270

I mean, most of people enter the Beijing

field through that.

271

Something that I'm actually very curious

about, because like I could keep talking

272

about that for hours, but I really want to

dive into what you're doing at the

273

Seahawks and also, you know, like how

Beijing stats is helpful.

274

to what you guys are doing.

275

I think it's the most interesting for the

listeners who understand basically how

276

themselves could they apply patient stats

to their own problems, which are not

277

necessarily in sports, but I think sports

is a really good field to think about that

278

because you have a lot of diversity and

you have also a lot of somewhat controlled

279

experiments.

280

You have a lot of constraints and that's

always extremely interesting to talk about

281

that.

282

Maybe you can start by basically

explaining how patient stats are applied

283

in your current role for analyzing player

performance and injury risk.

284

Because now that I work directly in

sports, something I'm starting to

285

understand is that really player

286

projecting player performance and also

being able to handle injury risk are two

287

extremely important topics.

288

So maybe let's start with that.

289

What can you tell us about that, Okay.

290

Let's see.

291

Which one should I start with?

292

I guess I'll start with injury risk, I

suppose.

293

Injury is like...

294

I mean, this is like a super difficult

problem to solve.

295

You know, I've written a number of papers

on those.

296

think you can link to my research gate.

297

And there's a number of methodology papers

that we've written that have looked at

298

things like this.

299

And I think it's complicated because one,

there's like a ton of inter -individual

300

differences as far as why people get hurt.

301

There's a ton of things that we probably,

you know, don't know they're important yet

302

because we can't measure them or we at

least can't measure them in the real world

303

applied.

304

setting, maybe in a lab you can.

305

And then there's other things that we just

don't know because we're like, it's a

306

epistemic problem.

307

Like we're just stupid about it.

308

We're naive that there's other things out

there that maybe we're just unaware of

309

yet.

310

And so it's a really hard problem to try

and solve.

311

So when I see papers that basically come

out and say like an injury prediction

312

model and they're estimating

313

prediction as like a one or a zero, like a

yes or a no, like a binary response, and

314

they give a nice little two by two table

and they talk about how well their model

315

did.

316

I'm always like, I don't, how is that

useful to the people who actually have to

317

do the work?

318

Because in reality, what we're dealing

with is it's probably not unlike a hedge

319

fund manager managing the risk of their

portfolio.

320

And if you think of each player,

321

or each athlete that you deal with as a

portfolio, they each have some level of

322

base risk.

323

So if we know nothing about you, you

really have to have a pretty good handle

324

in your sport of what's the base rates of

risk of injury for position groups and

325

players of different age and things like

that.

326

So that might be an initial model, right?

327

And then from there...

328

The players go out and they do things and

they play and they perform and they

329

compete and they get dinged up and they

take hits and they get, you know, hit by

330

hit by pitches or they get tackled really

hard or things like that.

331

And we collect that information and we're

basically just shifting the probabilities

332

up and down based on what we observe over

time.

333

And when that probability reaches a

certain threshold.

334

And of course you could use a posterior

distribution.

335

So you have an integral of like how much

of the probability distribution is above

336

or below a certain threshold.

337

Then you have the opportunity to have a

discussion about when to act or what to

338

do.

339

And how you act and when to act is going

to be dependent on your tolerance for risk

340

or your coach's tolerance for risk.

341

If it's your best player, if it's the MVP

of your team and it's week two of the

342

season and the risk probability, or let's

say we're using this as a model.

343

Some of the stuff that you mentioned Scott

earlier that we've worked on is like

344

return to play type of models where it's

like, okay, the athlete has, you know, saw

345

an ankle sprain and we're there rehabbing

346

And we have a, you know, we have a test or

several tests, a test battery that tells

347

us where that athlete is on their return

to play timeline.

348

Um, let's say it's week two of the season

and we say, well, there's a, you know, the

349

probability distribution, the posterior

distribution looks like this.

350

Here's the threshold that we'd feel

comfortable releasing this athlete back to

351

full on competition.

352

And there's a 30 % chance

353

they're in good shape and there's a 70 %

chance that they're below that threshold.

354

In week two of the season, we probably

want to say, you know what?

355

Let's not take that risk this week.

356

Let's be a little bit more risk averse

here because it is the best player.

357

And let's wait till we have more

distribution on the right side of the

358

threshold.

359

Alternatively, if it's

360

final game of the season, it's the Super

Bowl or the World Series or the Champions

361

League final or something like that,

you're going to probably take that risk

362

because you need the best player out

there.

363

And so when I think about injury risk

modeling, what I really think about is how

364

do we evaluate this individual's current

status?

365

on our sort of risk score or our risk

distribution.

366

And when do we feel like we need to

intervene and do something?

367

And when are we going to feel like, this

is fine and continue training as is.

368

And I think that's the tricky part.

369

I don't think it's not easy.

370

I don't think I've solved anything.

371

I don't think anyone has, but...

372

Certainly from the perspective of our

staff, we can all sit down with a

373

performance staff of strength coaches and

dieticians and strength coaches and

374

medical people and sit down and have these

conversations.

375

And what makes it nice about using a

Bayesian approach is that we can also take

376

into account domain expertise that we

might not have in the data.

377

So if we sit down on a Monday meeting and

then we say, you know, this player,

378

This is where they're currently at and

this is their risk status, which I don't

379

know, I don't really feel comfortable with

that.

380

How do you feel about it?

381

And then one of the medical people say,

you know, he's been complaining that his

382

hamstring feels really tight and he's been

getting treatment every morning.

383

Well, that's not data that we would be

collecting, but that's valuable domain

384

information that this individual who's

working with the player now adds to this.

385

And it's just like anything in

probability.

386

It's like if we

387

two or three or four independent sources,

all kind of converging on the same

388

outcome, on the same end point, we

probably need to feel really good about

389

making that decision and saying like, hey,

let's do something about this, let's act

390

now, right?

391

So that's kind of how we, you know, that's

how I think about it in that, you know,

392

from that side of things.

393

From the performance side of things, the

development side of things,

394

It's probably going to be, I mean, it'd be

way different for you guys in baseball

395

because you draft a player.

396

You don't expect them to maybe get to the

major leagues and contribute till 23, 24,

397

25 years old.

398

You know, for us, you draft a player and

those are going to be the, you know, next

399

year they're playing, they're ready, you

know, they're in, they're in the mix.

400

So in that regard, you'd be thinking of

models that would probably be, in my head,

401

I would be thinking of it as like models

that are mapping the growth potential of

402

an individual.

403

How are they progressing through the minor

leagues, which attributes matter?

404

And then maybe from there answering

questions like what's the probability that

405

this player makes 20 starts in the major

leagues or starts for three seasons

406

whatever end point makes sense to the

decision makers, obviously.

407

You know, for us, it's more about like

player identification.

408

And again, football is a, is a sport of

small samples.

409

And so in their college years, some of

these kids might really only be a starter

410

or a full -time player in their junior and

senior year, or maybe just their senior

411

year of college.

412

Additionally, you know, unlike, unlike the

NFL where

413

you know, at that highest level, the

talent is much more homogenous.

414

You get to the college football ranks and

you have just this diversity of talent

415

where you might have a big time team

playing a really lower level opponent.

416

And so, you you have to adjust things,

being able to hand off the ball to your

417

running back who's playing against a very

low level opponent.

418

And he goes for 500 yards or something

absurd, 200, 300 yards in a game.

419

that has to be adjusted and weighted in

some way because it's not the same as

420

going two or 300 yards against a big time

opponent.

421

And the big time opponents are more

similar to the NFL players that they're

422

going to play against.

423

And so, you know, all of these types of

things fit into models and hierarchical

424

models and Bayesian models, which help us

utilize prior information.

425

And the other way that the Bayesian models

are useful here

426

You know, sometimes we're dealing with

information that's incomplete because we

427

can't observe all of the cases.

428

You know, for example, in college sport,

division one is the top division.

429

You know, and then you have FBS and then

they division two and division three.

430

So if you pull all the division two kids

that have ever made it as a pro athlete,

431

the list is very small.

432

but they're kids that made it.

433

And so if you were to just build a normal

model on this, it would say like, well,

434

the best players clearly come from these

lower level schools because all of the

435

ones that we have seen have made it, have

been successful.

436

And in theory, there's hundreds of

thousands of kids from that level that

437

have never made it.

438

So we have to adjust that model in some

way.

439

We have to weight that prior back down.

440

Yeah, this guy is really, really good at

that level.

441

but our prior belief on him making it is

very, very low.

442

And you mean he'd have to be so

exceptional in order to, and this is where

443

like, oftentimes people rail on like, use

weekly informative priors, let the data

444

speak a little bit.

445

But there are times where in these

situations where I feel like you could

446

probably put a slightly stronger prior on

this and be like, man, this guy's really

447

gonna have to do something outstanding to

get outside.

448

of the distribution that we believe is on

this just given what we know.

449

Okay, yeah, that's very interesting.

450

That's a very good point.

451

Since I, yeah, related to survivor bias in

a way.

452

How concretely, how do you, how do you

handle these kind of cases?

453

Is it a matter of using a different prior

for these type of players or something

454

Try to do this in a few different ways.

455

One is you try and make basically like

equivalency metrics, like saying if you

456

did X at this low level, it in some way

relates to Y at this other level.

457

So you try and normalize players based on

players that you've seen that have moved,

458

say, between levels

459

of the game.

460

so like, again, if you think about it from

a baseball perspective, you know, hitting

461

40 home runs in AA baseball might be

related to, you know, might be in some way

462

convert to like 33 home runs in AAA and 24

home runs in the MLB or 12 home runs in

463

the MLB or whatever it might be.

464

Right.

465

So trying to,

466

identify equivalencies between those that

we can then like constrain everybody.

467

Other ways is just like, like you said,

like putting a prior on it.

468

knowing the level that the person is

playing at, you would have like a lower

469

level of prior.

470

For example, it's just like playtime.

471

If I think about playtime and performance

as sort of this,

472

this kind of like rising curve that goes

to an asymptote of some upper level of

473

performance.

474

The players way at the left who have very

small number of observations, it would be

475

silly to say that my prior for those

players is the league average.

476

There's a reason why they're not playing

very much.

477

It's probably because people don't think

they're very good, right?

478

So somewhere in that curve,

479

for each of those numbers of observations

across whatever performance metric we're

480

looking at, there's going to be a specific

prior on that continuous distribution.

481

And that's where I would, you know, that's

where we would kind of draw a stake in the

482

ground and say like, we probably think

based on what we know that this player is

483

closer to these players than he is to

those players.

484

Okay, yeah, yeah, I see.

485

Yeah, definitely makes sense.

486

And yeah, yeah, like that point of play

time already tells you something.

487

Because if the player plays less, then

very probably already you know you have

488

information about his level.

489

And that means he's at least not as good

as the A level players that play much more

490

The only time you get in trouble with that

is like an endowment effect where if you,

491

you know, like in major league baseball,

there's been some research on players who

492

are drafted very high in the first round,

second round get progressed up and through

493

the minor leagues faster than players who

were drafted lower, even if they don't

494

outperform those players just because

they're high as a consequence of being a

495

high draft pick.

496

That one's a tricky one, but there has to

be, at some point it's like actually, and

497

this is where like, know, posterior

distributions, you can really, I mean,

498

it's almost like doing an AB test.

499

Like we've got two players and what's the

probability that this guy is actually

500

outperforming the other guy, even though

the other guy might've been, you know, a

501

higher draft pick or something like that.

502

And so you try and at least display, you

know, we try and at least display that

503

visually and have those conversations.

504

It's,

505

kind of in my head, at least maybe I'm

wrong, but a nice way of like helping

506

people understand the uncertainty, you

know, which is really important.

507

always, maybe it's try, you know, I used

to work with a guy who whenever I would

508

present some of the stuff at work and he'd

be like, stop doing that.

509

Like every, every time you present, you

talk about like what the uncertainty and

510

the assumptions and the limitations are,

like just give them the answers.

511

And I'm like, well, it's important that

they know what the limitations

512

and what assumptions are behind this

because we can't, we don't want to talk

513

past the sale and sell them on something

that, you know, isn't really there.

514

Like there's been times where I've had to

stop someone and just be like, hold on.

515

This analysis definitely can't tell us

that.

516

Like what you're saying right now, it

can't tell us that.

517

like, let's not, let's not try and make

this more than it is.

518

And also just, you know, conveying your

uncertainty.

519

mean, that's just super important because

520

It's really, really hard.

521

I mean, we're all going to fail at trying

to identify talent.

522

It's really hard to identify why one

player is going to succeed over another.

523

so, you know, in some way it's not binary.

524

It's not a like, do you like this guy or

not?

525

Is he good or bad?

526

Is this guy better or worse than the other

guy?

527

there's a lot of factors that go into why

someone has success.

528

And so I think conveying that uncertainty

is really important.

529

And obviously, the more observations that

we have of you doing the thing, the more

530

certain we are that this is your true

level of performance.

531

But it takes a while to get there.

532

So we have to just be honest about that.

533

Yeah, yeah.

534

I think that's actually related to

something I wanted to ask you about also a

535

bit more generally, you know, but

536

the most significant challenges that you

face when applying Bayesian stance in, in

537

sports science and, and how you address

them, because I'm guessing that you, you

538

already started talking a bit about that.

539

So, let's go there.

540

And then, then I have other technical

questions for you, but the kind of, of

541

models and, and, and usefulness that

Bayesian stance has in your field.

542

But I think this is a good moment to, to

address these.

543

questions.

544

think the biggest or there's a few

challenges.

545

One challenge is not everybody is excited

about a posterior distribution like you

546

might be.

547

Most of the time, they just want an

answer.

548

Tell me what to

549

Give me the yes or no, make it binary.

550

And so that's always tough.

551

And you're trying to oftentimes convey

this to non -technical audiences or people

552

who are good at doing other things.

553

They're not math people or they're not

stats people.

554

And that's okay.

555

So that always makes it challenging is why

are you showing me this distribute?

556

I don't understand what I'm supposed to

take from this.

557

Just tell me.

558

What to do?

559

Tell me which guy's better.

560

Tell me which guy's worse.

561

So that's always hard.

562

And that takes a lot of patience and

communication.

563

For a while, we used to do just weekly sit

downs with our scouts where we would teach

564

them about like one stat a week.

565

And we'd go slow.

566

And we'd also try and...

567

as best as possible, relate things back to

the currency that they speak in.

568

And scouts and coaches, the currency they

speak in is video, not charts and graphs.

569

So the more that we can connect our

analysis to video cut -ups, because then

570

they can see it.

571

And then they understand why a model says

what it says or makes a decision or why it

572

has assumptions.

573

And this is also super valuable too,

because they give

574

And they say, it's, saying that, you know,

the model is saying that, this is, is the

575

outcome, but I can see why it's because

these four other things happen.

576

It's like, wow.

577

Well, we could probably account for that.

578

And we never, I just didn't know it,

right?

579

That's why they're domain expert and, and,

and I'm not.

580

so.

581

You know, the patience around

communicating stats and numbers is always

582

difficult and also knowing what people

like.

583

When I first started, everybody would tell

you, need to have, you know, got to have

584

an amazing dashboard, got to have like

charts and graphs, you know, and all that

585

stuff.

586

And what I found was there was a lot of

people who were like, I don't, what do

587

you, I don't even know what I'm looking

at.

588

Like, I hate these things.

589

Just give me the table of numbers.

590

It's like, okay.

591

Well, maybe a table of numbers with just

some conditionally formatted information.

592

And also, you know,

593

I have an academic side, I do supervise

PhD students and master students, and I do

594

teach a master's class in statistics at

college.

595

So I guess what I'm about to say would,

know, people on the academic side would

596

hate it, but you have to like recognize

the environment you're in.

597

And sometimes just like changing the

verbiage helps, like instead of calling

598

things the...

599

low credible interval and the high

credible interval, like we just call it

600

the floor and the ceiling.

601

And people are like, yeah, this guy's

floor, it's a bit higher than the other

602

guy's floor.

603

And that guy's ceiling, this guy's got a

better ceiling.

604

And like, know, academically you'd get

shot for that, it's like, those kinds of

605

things go a long way because it brings the

information to the end user.

606

And if you want them to start to...

607

take this information into their decision

calculus, you have to get them

608

comfortable.

609

And sometimes it's just meeting them with

terminology that helps.

610

And so I think that's a really, you know,

that's a big one.

611

Those are big challenges in communicating

this stuff.

612

Yeah, definitely.

613

And I resonate with that.

614

I've had the same issues.

615

I'll be able to tell.

616

talk more precisely about sports in a few

months.

617

But when it comes to a lot of other

fields, whether it's marketing or biostats

618

or electrical forecasting, yeah, the

issues are related to these.

619

They're also extremely diverse.

620

So that's interesting.

621

You definitely don't have a one size fits

all.

622

Definitely what's extremely important

basically is to know the model extremely

623

well from my experience.

624

And yeah, if you have coded the model

yourself, you usually know it really well

625

because you spent hours on it to try and

get it to work and understand what it's

626

doing.

627

And when it's not able to do as you were

saying, I think it's extremely important

628

to be able to tell people what the model

cannot tell you.

629

And yeah, I think these are extremely good

points to try and balance what people are

630

usually wondering about.

631

And that's also where I think having the

Bayesian model is extremely interesting,

632

right?

633

Because the Bayesian model by definition

is extremely open box and you have to run

634

it down your assumptions.

635

And so you know much better what the model

is doing than a black box model.

636

Yeah, I mean, that's another good point

is.

637

If you go into a meeting and you have

model outputs and your only reason when

638

asked, why does it prefer this over that?

639

Your only reason is because the model said

so.

640

If people aren't going to be super excited

about that.

641

knowing why things are happening, know,

this also, you know, I mean, this really

642

plays into like how you validate and check

your models.

643

And so buildings, you know, we kind

644

within that Bayesian sort of world,

building simulations is a big part of it.

645

And building simulations to see how the

model behaves under different constraints

646

and different pieces of information,

that's really important because it gives

647

you useful context to talk about and it

gives you useful information in order to

648

head things off at the pass when you know

there's gonna be some gotchas and some

649

trouble if, you

650

people have certain types of questions.

651

You can head things off of the past

because you're already aware of them.

652

Another thing that I do think is really

useful in this and maybe in some of your

653

prior work in consulting, I'm sure you've

like stumbled on like, or used frameworks

654

like crisp DM and things like that.

655

Like in statistics, there's a PPDAC

problem plan, data analysis and

656

conclusion.

657

Those types of frameworks help just

because again,

658

A lot of times we're dealing with non

-technical audiences and they're trying to

659

give you a question and say like, Hey, can

we look at this?

660

And oftentimes these things are very vague

and very sort of like, you know, not, not,

661

not clearly defined.

662

like, you know, my younger self would take

that and run away and, know, do something

663

for a week or two and then come back and

be like, Hey, here's this thing, you know,

664

and you ask about

665

you know, they're usually like, the reply

is, that's kind of cool, but I was

666

thinking of it like this and I would do

this with it.

667

it's like, man, if you, you know, if you

told me that two weeks ago, I would have

668

done something else.

669

So using those kinds of frameworks, one,

does a few things.

670

One, it gives us the opportunity.

671

Like I always tell our analysts like

question the question, like, you know,

672

question the question.

673

Right?

674

So when they have a question, I'm always

sitting there and I'm like, okay, well,

675

you know, what would you want to do with

this?

676

How, do you foresee yourself using it to

make a decision?

677

What's the cadence that you would need to

access this information?

678

If I were to get it to you tomorrow, you

know, what would you, what kind of

679

decision would you want to make?

680

Like really kind of Socratic questions,

you know, question the question.

681

And, that does a few things.

682

One, we get, we get to two, you know,

683

usually two different results.

684

Both of them are good.

685

The first is I get them to then walk

through that five minutes with me and

686

clearly define what it is they're looking

for.

687

That's great.

688

The other result is the opposite, but it's

also a good result, which is we get about

689

three minutes in and they go, you know

what?

690

I haven't thought about this well enough.

691

Let me think through it a bit more and

come back to you.

692

In which case I didn't waste the time

building things and scraping and cleaning

693

data and doing all that stuff.

694

The other thing that those frameworks do

is, and I try and get analysts to think

695

like this, is utilize each step within

those frameworks as touch points back to

696

the person who asked you the question.

697

Hey, this is where we're at.

698

We've collected this kind of data.

699

These are the things we're thinking.

700

These are the features that we're thinking

about using.

701

What do you think about that?

702

Anything else you can think of.

703

By doing that, along each step of the way,

they get to see the model developed.

704

They get to provide input.

705

And what that does is it gives them a bit

of ownership over it.

706

So when you get to the end result, they're

like, geez, this was built exactly in my

707

vision, and now I'm excited to use it.

708

And that's a really cool thing too.

709

Yeah.

710

Yeah.

711

Thanks for that detailed answer, Patrick.

712

I can definitely hear the 10 years of

experience working on that.

713

That makes me think about a lot of other

things.

714

Yeah, definitely the same for me, would

say, where my personal evolution has been

715

trying to really understand the question

the consumer of the model is trying to get

716

to, right?

717

Like what actually is your question?

718

Because you have something in mind, but

maybe the way we're talking about it right

719

now and the way I have it in mind is not

what you want.

720

And so, yeah, as you were saying, a good

model is really that's custom made, that's

721

fine and hard work and that takes time.

722

so before investing all that time in doing

the model, let's actually make sure we

723

align and agree on what we're actually

looking at in studying.

724

That's, think it's extremely important.

725

Yeah, no doubt.

726

I think that's often the hardest part,

because it's just getting people to really

727

define.

728

that's probably, I mean, that and making

sure that you have good data.

729

Those are the two biggest things.

730

The model building part and things like

that sort of happen a little bit easier

731

once you do the first two things.

732

That's always the tough part.

733

Yeah, yeah, yeah.

734

Actually, continuing on that topic, how do

you communicate these statistical

735

concepts?

736

And honestly, a lot of them are really

complex.

737

So how do you communicate that to non

-stats people in your line of work?

738

I'm guessing that would be scouts, as you

talked about, coaches, players.

739

How do you make sure they understand?

740

what you're doing and in the end are able

to use it because we talked about that in

741

episode 108 with Paul Sabin.

742

If your model is awesome but not used,

it's not very interesting.

743

So yeah, how do you do that?

744

First, trying to really understand what

kind of cadence this is going to be on.

745

So some questions.

746

especially in sport, get asked.

747

And they're more asked from the knowledge

generation standpoint, meaning that I have

748

a question.

749

I think it'll help us with, you know,

updating our priors, our prior beliefs

750

about the game.

751

Maybe things have changed.

752

Maybe rule changes have altered things or

something like that.

753

Can we study this?

754

A question like

755

for knowledge generation requires a

different output than something that's

756

like weekly or daily consumption.

757

So if it's for knowledge generation,

that's usually communicated in the form of

758

like a short written report.

759

The question at the top, the bottom line

up front, here's the four bullet points,

760

and then the nitty gritty.

761

Like this is how we went about studying

it.

762

charts and graphs and usually it's like a

page or two and a PDF or maybe like an

763

interactive HTML file that they can see

things and have a table of contents and go

764

to different sections.

765

If the question is directed at stuff

that's required to be evaluated weekly or

766

daily, like I need to see this every week

because we're going to be evaluating a

767

certain player or an opponent or I need to

see this daily because it's

768

player health related, something like

that.

769

We're always thinking in terms of like web

applications.

770

So how do I get, you now I have to think

through the full stack pipeline of like,

771

where do we get the data?

772

Where does it live in the database?

773

What's the analysis layer?

774

Kick it out to an output.

775

Where's that output stored?

776

And then how does the website ingest that

output and make it consumable?

777

And for that,

778

It's usually some form of charts and

graphs and a table.

779

And usually it's interactive stuff.

780

So they can sort and filter and hover over

points and access the information.

781

And again, as best as possible, I'm always

thinking to try and develop that in the

782

way that they're going to use it.

783

So like I was sitting down, for example,

today with our director of player health,

784

and he was like, you

785

I'd love to have this information daily so

that I can relay it to the new coaching

786

staff.

787

And I want to say it, you know, say these

things.

788

Okay, great.

789

I have all that information.

790

I have all of that, those models, but come

over to the whiteboard and draw for me the

791

path that you want to take to going from

sitting at your desk.

792

and reading the information from a webpage

to how you want to communicate it.

793

And as soon as he started drawing it out,

it's like, okay, I know exactly what to do

794

now.

795

That's perfect.

796

Otherwise I would have built something

that in my head I thought would be useful,

797

but maybe not useful to him.

798

And then he uses like part of it or maybe

because he's super motivated, he's going

799

to use it.

800

And he's also going to,

801

use like 10 other things to get the other

stuff he wants, but he's a nice guy and he

802

doesn't want to tell me that it doesn't

have all the things that he needs.

803

And so then like four weeks later, I walk

in his office, I'm like, what are you

804

doing?

805

It's like, oh, I go here and then I get

this information from this webpage, but

806

then I go to this other three webpages

again.

807

So, whoa, whoa, whoa, why didn't you just

tell me that?

808

Like I'll just, I could make this all into

one thing.

809

Like you don't have to, and so.

810

That's a really important piece is knowing

how the data is going to be utilized,

811

making sure that it's exactly in the order

that the decision maker requires it.

812

Yeah.

813

Awesome points.

814

Yeah.

815

Thanks for that, Patrick.

816

And I think it's also very valuable to a

lot of listeners because we're talking

817

about a professional sports team here, but

it is definitely transferable to

818

basically, I think, any company where

you're working

819

different people who are using the models

but are not themselves producing the

820

models.

821

It's like almost every company out there.

822

yeah, I think and also from my experience

doing consulting in a lot of different

823

fields, I can definitely vouch for the

things you've touched on here.

824

yeah, thanks.

825

That's definitely, I think, very valuable.

826

turn back a bit more to the technical

stuff because I see time is running and I

827

definitely want to touch a bit more on the

spot side of things and how patient stance

828

is applied in the film.

829

Obviously a very important part of your

work is, I'm guessing,

830

drafting players, player selection

processes.

831

So yeah, how might Bayesian methods be

applied here to improve the drafted

832

strategies in the player selection

processes?

833

Yeah, well, again, like I think I said

earlier, everybody's going to miss.

834

It's impossible to be, you know...

835

to have a good hit rate and always be

picking, you know, picking players who are

836

going to reach high level success.

837

And a lot of that is just because, you

know, performance and talent are extremely

838

right -tailed.

839

You know, you have a whole bunch of

players that never make it.

840

You have a small group that make it and

are good enough to make it.

841

You have an even smaller group that are

good enough to make it and like really

842

good to play all the time.

843

And then you have

844

a few Hall of Famers sprinkled in, right?

845

So it's really right -tailed.

846

it is very hard to do this stuff.

847

So, you know, understanding or modeling

your uncertainty, that's really important.

848

And

849

information from the domain experts, know,

scouts see things on film that we can't

850

see in numbers and vice versa.

851

One of the values that we have is we can

process way more players than any one

852

human can actually watch.

853

So we have the ability to build models

that can identify players and hopefully

854

get them,

855

over to the domain experts who have to

then watch the film and write the reports

856

and say like, hey, did you know this guy

was really good in these things?

857

This is his potential ceiling.

858

And we think that we have, you know, we

think that this would be valuable for our

859

team, right?

860

Building models like that, that help us.

861

Identify talent, give us a range of

plausible outcomes.

862

One, it helps us get information to the

people who have to watch the film and make

863

the decisions.

864

Two, it helps us have discussions about

where the appropriate time to acquire

865

people

866

If you're sitting there, obviously, you

know, in the major league draft, major

867

league baseball draft, it would be the

same thing.

868

Everybody knows who the first round picks

are and the second round picks.

869

It's after that, that things become pretty

sparse.

870

And if you can identify players that have

unique abilities later in the draft, that

871

opens up a lot of opportunities to,

872

select players that might be able to

contribute successfully to your team.

873

And so that's really where those models

help us.

874

The other area that they help us in is, I

always talk about with our analysts, like,

875

what is the benchmark that you're trying

to beat?

876

So every model, like you can't just build

a model.

877

I mean, I remember one of our analysts,

she had a model and she said, I built a

878

model and I think it's really good.

879

And I said, cool.

880

How well does it do against the benchmark?

881

She's like, well, what do you mean?

882

And I was like, well, like how well does

it do against if we just use, let's say

883

scout grades or if we just use public

perception, how well does it do

884

historically against that?

885

She's like, no, no, no, I don't care about

that.

886

Like this model is just with their stats

and you You know, it's like, no, no, but

887

you have to care about that because if

it's not better than those things, then

888

why would we use it?

889

Right?

890

You have to be able to beat that

benchmark.

891

One of the areas where we can really beat

a benchmark is when we combine the domain

892

experts information with the actual

observed data information.

893

And a Bayesian model allows us to do that,

right?

894

It allows us to take down the domain

expert who's maybe scoring the player a

895

certain way, writing information about the

player.

896

It allows us to take that information.

897

mix it with the numbers and get a model

that is, I guess, man and machine, right?

898

And those models beat our benchmark much

better than any one of these alone, right?

899

If we just use numbers, never watched any

film, never knew anything about the

900

player, or if we just use domain expert

information.

901

When we combine those things, we tend to

do a much better job.

902

And so that's where Bayesian analysis

really helps us.

903

And also,

904

That's where you start to get interesting

discussions about the floor and the

905

ceiling of a player.

906

Because now once you run their posterior

distribution and the domain experts

907

information is in there and you're saying,

yeah, this guy, he's awesome at tackling

908

and he'll be a great tackler and blah,

blah, blah.

909

And these are his numbers.

910

the numeric model says like, yeah, I think

this guy's a pretty good tackler.

911

Domain experts saying like, no, no, no, I

watched him and he doesn't play against

912

great competition, but his technique is

really bad.

913

It's not going to translate against these

bigger players.

914

It's like, well, that's not information

that maybe our stats would have.

915

But when we combine those two bits of

information, all of a sudden, our maybe

916

overly bullish belief in this player gets

brought down a bit.

917

And utilizing the information like that

918

is interesting and it also makes it unique

to the people that are in that room, the

919

domain experts that you have in that room

and things like that.

920

How you weight those things is really

important.

921

For our own analytics staff, we'll do

things like we'll build our own separate

922

models and have our own meetings and we'll

build our own analysis.

923

So we'll have independent models all

against each other and maybe we'll have

924

them weighted or we'll use

925

you know, like triangle prior and build

them together and, you know, mix them

926

together and get posterior simulations.

927

And we try and do those things in a way

that allows us to understand all the

928

plausible outcomes that might be relevant

for this individual.

929

It's fascinating.

930

Yeah.

931

And I really love both that feel the fact

that you have to blend a lot of different

932

information.

933

Like the domain knowledge from the scouts,

the benchmark from the markets, the models

934

that you have in house, also scientific

knowledge of all the scientists that the

935

team has inside of it.

936

that makes all that much more complicated,

right?

937

I'm guessing sometimes as the modeler, you

would probably be like, my God, that'd be

938

so much easier if we could just run some

very big neural network and that'd be

939

done.

940

at the same time, I think it's what makes

the thrill of that field, at least for me,

941

is that, no, that stuff is really hard.

942

There is a lot of randomness.

943

There is a lot of things we don't really

understand either.

944

And you have to blend all of these

elements together to try and make the best

945

decisions you can, even though you know

you're not making the optimal decisions,

946

as you are saying.

947

And I think it's a fascinating field to

study important decision -making under

948

uncertainty.

949

Yeah, for sure.

950

I think that's the thing that's most

951

interesting about it to me.

952

Like, yeah, I think that's the most that

stuff is fascinating just knowing

953

Yeah, Decision making under uncertainty is

really challenging and I think that's the

954

thing that makes this the most, you know,

the coolest stuff to work on.

955

Yeah, yeah, no, definitely.

956

Actually, maybe a last question on the

technical side.

957

Now if we look, so we've talked about the

beginning of the career of a player,

958

right?

959

Like the draft.

960

We've talked about...

961

kind of the whole lifetime of the player,

which is projection, performance

962

projection over the whole career.

963

Now I'm wondering about the day -to -day

stuff.

964

What can Bayesian models tell us here or

how can they help us in predicting the

965

impact of training loads on the athletes'

wellbeing and performance?

966

I know, I think it's kind of a frontier

967

almost all the sports, but I'm curious

what the state of the art here is,

968

especially in US football.

969

Yeah, it really is, I think, the sort of

one of the final frontiers, I guess, in

970

sport.

971

Team sport is just challenging because you

perform well or you win or you lose due to

972

a whole bunch of issues that sometimes

973

have nothing to do with you.

974

For example, I can train you, you know, we

could train you and you could be very fit

975

and strong.

976

And if in the last play of the game, the

quarterback throws the ball to a patch of

977

grass and you lose, it had nothing to do

with you being fit and strong.

978

know, counter that to like individual

sport athletes.

979

If you're a 400 meter runner, a cyclist, a

swimmer, a runner, a marathoner, you know,

980

physiologically.

981

If we build you up, we have a much more

direct line between how you develop and

982

how it directly relates to your

performance.

983

There's not a lot of other information

there.

984

No one's trying to tackle you on the bike

or in the pool or something like that.

985

So that makes, that makes a sport much

more difficult.

986

Baseball is probably the closest because

even though it is a team

987

It really is this sort of zero sum duel

between a pitcher and a batter.

988

And one guy wins and one guy loses.

989

And the events are very discreet.

990

The states of the game have been played

out, know, runner on first and second with

991

two outs, bottom of the third, blah, blah,

blah.

992

So it's maybe a little bit more clear in

baseball.

993

I think in the other team sports, in the

kind of invasion sports,

994

what makes this challenging is

identifying.

995

I always try and take it back to

identifying the discrete events that we're

996

trying to, trying to maybe measure

against.

997

like, for example, I can give you example,

a pretty clear example from basketball.

998

was talking with a friend in a, in an NBA

team and, he was like, yeah, you know,

999

our, our, our coach and our scouts and

the, you know,

coaches, feel like our players don't close

out three pointers fast enough.

And I was like, well, is that a tactical

problem or is it a physical problem?

And he's like, well, how would we look at

that?

And I was like, you have the player

tracking data.

And if you know every time your team's on

defense, which is easy to know, and you

know every three pointer that's been shot

against your defense, if you were to take

that frame,

out of the player tracking data and maybe

like the frame a second to a second and a

half before that.

So all of that information for every one

of those three pointers.

You have an idea of the relationship

between your player and the player who's

taking the three point shot.

You have an idea of the relationship

between your player and the other players

on his team.

So you know from a technical, a tactical

standpoint.

you know what type of like formation or

defense you're trying to run.

So first things first, are the players in

the right position to close out that three

pointer?

Maybe, you know what?

Our guys consistently mess up the

defensive shape and when they get in

there, they give too much ground to the

guy shooting a three pointer.

The other is the physical standpoint of,

well, no, they're in good position, but

when they go to close it out over that

second and a half,

They're not fast enough to get there.

Okay, great.

Now roll it back to what you can measure

in the gym.

Is there some measure, let's say on a

force plate of the amount of impulse or

force under the force time curve that the

player outputs that can tell us something

about their ability to move rapidly, apply

force into the ground, move rapidly to

close out that three pointer?

And maybe if you look at several years

worth of data, you'd find

The top players on your team all do this

thing really well, and some of the worst

players at closing out the three do this

thing poorly.

And so now you have something to say about

like, hey, what if we develop this quality

in the off season and our players, would

we be able to close out the three pointers

more effectively, more efficiently?

And so I think from that standpoint,

linking the development piece to sport,

team sport, invasion sport.

You have to really think about the

discrete events of the game and how you

can kind of tease those out of, let's say

the player tracking data.

And it's like super hard in something

like, you know, in football, because

players all do really different things.

You know, the linebacker does something

totally different than the offensive

lineman.

And so you have to really get down to the,

the domain of each of those positions and

say like, gosh, what are the discrete

events?

that define what this position does, then

how do we measure success in those?

And then if we can measure success, how do

we identify the archetype of players who

are good at those things?

And then if we can do that, maybe then we

can start to talk about, is this something

that you can develop in a player?

Is it something that you have to identify

in a player?

That's sort of the, in my head, I mean, I

don't know, I could be wrong.

This is not.

Nobody, think everybody's trying to figure

this out, but I could be wrong.

But in my head, that's at least the

process that I would, you know, I try and

think through when I think about these

things.

Yeah.

Yeah.

It makes a ton of sense.

mean, and it seems like, yeah, that, and

there are so many areas, open areas of

research on all of that stuff.

That's just, just fascinating.

I'm

I'm already thinking, that'd be amazing to

have a huge patient model where you have

all of those topics that we've talked

about.

Basically, it could be a big patient model

where you have a bunch of likelihoods.

And yeah, that'd be super fun.

I'm guessing we're still a bit far from

that, but maybe not too far.

Hopefully in a few years, that'd be

definitely super fun.

Yeah, no doubt.

Yeah, I mean, and that's, definitely

doable.

But yeah, you need you need really good

data and you need really good structure in

your model.

Yeah, that's the part too, is getting

getting good data, know, player tracking

data is fine.

I mean, it has errors, you know, people

who think that it's like a panacea, you

know, it's like, have you really worked

with it?

I mean, there's

Sampling at 10 hertz for humans that move

really, really fast.

Acceleration is a derivative of speed.

At 10 hertz, people who are moving really

fast, that data gets noisy pretty quick.

I think one of the things is as we

progress, as the technology keeps

improving, things get better.

you get better data and maybe that helps

you also answer some of these questions a

little bit more specifically.

yeah.

And then we'll be able to have our huge

patient model with a lot of different

likelihoods in there that fit into each

other.

And then we don't even need to play the

game.

We don't have to play the game.

They just let the computers play the game

and it's over.

We're done.

Yeah, no.

No, you still have to play the game

because you still have randomness.

Then you're like, yeah.

mean, because otherwise the model is kind

of like if you want kind of a quantum

state, right?

Where the model can see the probabilities

of things happening, but then you have to

open the box and see what is actually

happening.

So you can have the best model.

In the end, you still have to play the

game to see what's going to happen because

it's not deterministic.

Yeah, thankfully.

yeah, that's right.

Yeah.

Yeah.

But I mean, it's definitely I always love

doing these these big models.

And that's definitely doable.

I've done that for election forecasting,

for instance, where you have several

likelihoods, one for polls, for instance,

and one for elections.

So yeah, that's I know that's definitely

doable in the Bayesian framework, because

I mean, why not?

It's just part of the big

of the same big model in a directed S

-secret graph, if you want.

But yeah, I'm curious to see that done in

spots.

Maybe we'll get back together for another

episode, Patrick, where we talk about that

and how we did that.

That'd be cool.

Yeah, there you go.

Yeah, actually, I wanted to ask you to

close us out here.

about, you you've started talking about

that right now, like some emerging trends

in sports analytics that you believe will

significantly impact how teams manage

training, performance, drafting in the

near future.

And also if there are any spots you see as

more promising than others.

well, mean, yeah, trends.

Yeah.

We talked a lot about that stuff and I

think, you know, better data and better,

you know, better technology.

all of those things will, will, I think

will help us.

I think also it's getting, you know,

getting the decision makers comfortable

with the utility of some of this stuff,

you know, baseball, has always been a game

of numbers.

And, I think early.

seven, you know, releasing data kind of to

the public, really the first sport to get

player tracking data, things like that.

I think that opened up a lot of

opportunities for people to do really

interesting work in the public space,

which then sort of got

teams interested and then sort of a, you

know, more of a shift in people in the

front office where, maybe historically it

was ex players who kind of played out

until they retired and then became scouts

and managers and things like that.

I think that, you know, that happening in

baseball was a really good thing for that

sport.

And I think slowly for the other sports,

that's really

probably needs to happen because the more

that these things are open and sort of

curbside, I think the more the decision

makers become comfortable with them and

can say like, I can see how I would use

this.

I can see what this might help me with.

so I think that's never underestimate the

work that you do in the public space

because I think there's an opportunity to

always.

you know, help things evolve,

crowdsourcing, guess.

Yeah, mean, preaching to the choir here.

Yeah, for me, a lot more of these data

would be open sourced.

Yeah, I mean, there is also an extremely

interesting trend right now towards open

sourcing more and more parts of large

language models.

I think that's going to be extremely

interesting to see that develop because

At the same time, this is very hard

because these kind of models are just so

huge.

You need a lot of computing power to make

them run.

So I don't know how open source can help

in that, but I know how open source can

help in the development and sustainability

and trustworthiness and openness of all

that stuff.

So that's going to be super interesting.

And I'm also going to be very interested

in

the different spots evolve.

Now that basically the nerds are they are

much more right than before.