Why ‘Data-Driven Decisions’ Are a Myth

I grew up in a mining town, and I don't recommend

that actually, as a place to grow up. Got

partially homeschooled by my mom and dad because the quality

of the local schools was not what they wanted to see. I

saw resource production up close and personal. I saw data

coming in up close and personal before anybody knew what to do with it

or how to use it or analyze it at all.

Statistics on operations, on interruptions, what

causes machines to go down and be unavailable for all

these kinds of things. Data can help us understand that.

Hello, and welcome back to Data Driven, the podcast. We explore the

emerging industry of AI, data science, and, of course, data

engineering. You may notice that my favorite data engineer in the world,

Andy Leonard, is not here. However, I brought the most quant,

most quantum curious person I know. I don't like calling her curious, although

one could argue. And Candace Cooley. How's it going,

Candice? It's great. Thank you so much. I'm really excited to be part of Data

Driven today. Awesome. We're happy to have you. And we're also very

happy to have Mr. Dale Nesbitt, who is,

in addition to being at Arrowhead Economics, he's also a

lecturer at Stanford, and I guess he has his first summer class today,

and we're happy to have him. Welcome to the show, Dale.

Thank you. I appreciate it. And thanks for the opportunity to speak with you.

No problem. No problem. What exactly does

Arrowhead Economics do? It's implicit in the name

we do economics. We named after

ourselves after the second Nobel laureate in economics,

Kenneth Arrow, who was riding his bike around Stanford

happily until his 96th birthday. And then. Then we

lost him about five years ago. So we do

economics in the energy patch, critical materials

patch, any commodity that's. That's

produced or traded. Interesting. And that's a.

That's a very large field. I don't think. If were it

not for Eddie Murphy's movie Trading Places, most

people probably wouldn't know a thing about it. I was a kid when that came

out, and I was just fascinated. So sorry. I'm sure that's not the first

time you've heard that, and I hate to tell you, it's probably not gonna be

the last time you've heard that. Well, yeah, I've heard that.

I had an auspicious start. I grew up in a mining town.

Okay. I don't recommend that, actually, as a place to

grow up. Got partially homeschooled by

my mom and dad because the quality of the local schools was not what

they want. I saw resource production up

close and personal. I saw data coming in up close and personal

before anybody knew what to do with it or how to use

it or analyze it at all. Statistics on

operations, on interruptions, what causes

machines to go down and be unavailable for all these kinds of

things. Data can help us understand that.

Interesting. So you were the data. Life really found you? It sounds.

Yeah. Yes and no. Yes. My background is actually in

probability and decision analysis. That's where I did my research.

And what. One of the things that you're going to find the data is really

good for is developing probability distributions over

phenomena that you don't understand the uncertainty about.

We want to understand the uncertainty about certain phenomena.

Data is a way to do that. Interesting. Sorry,

Candice, I cut you off. No, I was curious. So many organizations, they

collect enormous amounts of data, but why

do so few of them seem to not be making better

decisions based upon their data? As a professor of decision

analysis, you need more than data to make decisions.

You need alternatives. You need information, probability

distributions, not data. And you need a notion of your

values, your objectives, what do you like and what do you not like. So if

you're coming like profits, there's a lot of uncertainty

in the middle. If you're producing oil, there's a probability distribution

over oil price, over supply chain costs and those things.

You can't make a good decision unless you have those probability

distributions. I get them in part from data.

The data is out there telling you what these probability distributions

are, if you know how to process it. The reason is because

the notion data driven decisions is a non

sequitur. It really is. Data is not

enough to drive your decisions. You need intelligence,

you need an understanding of uncertainty. What's the on

average value of these variables that you're going to see?

How much spread is there in those variables? You can't make a decision under

uncertainty unless you understand the uncertainty.

So data driven is not data driven, it's analytic driven. And

the data helps you. If you, some of the techniques that, that

we use in practice get what those uncertainties are

to feed into a decision analysis

framework, then you make good decisions.

Yeah, I mean, that makes sense, right? Because data would be the raw

commodity, if you will, the raw oil. But the thing that

you put in your car, gasoline, petrol, whatever you want to call it,

is the refined product from the raw material, you know, Frank?

Absolutely. Data is the crude oil and

the probability distribution of the finished product, gasoline,

distillate, so forth that you put into your machinery that makes it go

absolutely Data is the raw material.

And one of the problems there, a Candace, pursuant to your remark,

is this is Nesbitt's maxim number three, people

only gather data that's easy to gather.

And maximum number two is the process of gathering

data is itself stochastic.

Watch how data is gathered. It's a noisy, noisy process

to observe anything and to write down the

correct, say, operation of machine. We

don't even know what the price of crude oil is because everybody reports

it differently. So data itself isn't

gathering itself is noisy. You have to take that into account when you

analyze it. So this idea that data is. It's like

Rumpelstiltskin, right? Spin straw into gold. You can't.

The data is not straw. It has to be scrupulously

worked. Frank, just exactly what you alluded to to get

proper decision making, to get proper

forecasting and those things.

Well, then doesn't that explain where a lot of data projects or

AI projects go wrong, is because they don't take into account the random.

The randomness. Because you said data collection is stochastic. Right. So

it is. It's inherently unpredictable.

Right. And doesn't that kind of echo throughout the training of these models

that we have and cause weird outliers?

Or does is there magic that can happen that can cancel that

out? Yeah, there's. Well, there's magic. You try to measure

through observation the intrinsic uncertainty in the data, like

the difference between predicted and actual, maybe due

to measurement error, just observing the data. And

the great statistician Fisher told us that data was

generated by a stochastic process. And what

you're trying to do is figure out what that stochastic process must

have been very subtle. So you,

you, at the same time you're estimating, say, coefficients for your

model or something you want to estimate, you're also estimating what

stochastic process was at work when the data were

gathered. And you can infer that

it's very sophisticated. AI doesn't do any of that.

Regression analysis does that. AI

doesn't deal with uncertainty at all. No,

it actually gets really wonky when it's uncertain.

Yeah, wonky being a technical term, of course.

But I think it's really important when you think of data analysis. You want to

be careful to try to measure intrinsically what sort of

stochastic variation, what sort of process by which people

observe data. What was it? When we

do regression analysis, simple linear regression is a term,

sigma in there, and that sigma has to do with the random nature

of the data. So

interesting. So, yeah, it really goes back to the fundamentals of

statistics, doesn't it? Yeah. I mean, think of the, you know,

the clock's always right. Suppose you're measuring the time

and you fail to notice that the clock broke.

And it gives you the same time all the time. Your sensor

broke and it gave you the same reading, no matter what the state of the

machinery was. Your data set ain't too good with

that in there. And yet, wow. No matter what, I get the

same answer. That's really predictive. Fact is, it's a broken

data machine. Right. So you get just

industrial measurements. Your, your process of measuring,

be it manual or sometimes be it automated, you have to

be very, very cognizant of

that and the stochastics of that

and take that into account when you try to understand what that data

implies about your probability distribution over something

interesting. Candice looks like she's

going to say something. I'm thinking about uncertainty,

and I'm also thinking about risk. And I'm trying to understand

what's the distinction that matters between risk and

uncertainty. Well, it's, it's kind of interesting. So suppose

that. Let's take the price of gold. You're going to open a gold mine. Price

infinity minus three. It's really, really high.

Wars and things like that caused that to happen.

But you're uncertain about it. But if you're going to open

that gold mine and it's profitable at a gold price of

fifteen hundred dollars an ounce, there's uncertainty, but there'

risk has to do with loss. So suppose you were

uncertainty. One day I grew up in Reno. One day I walked into the

nugget, I was underage, and I inadvertently pulled the handle on the

slot machine and it played. You didn't have to put money

in it and it would just play. It was broken.

That's a pretty good lottery. I played it until the security guard came and

kicked me out. Right. So there was

uncertainty. I didn't know if it was going to give me cherries or clowns or

whatever, but there was no risk. I didn't have anything at risk.

Risk happens when you have losses that have to be balanced against

gains. Now all of a sudden, your elementary tech

gets a lot tighter when you have losses. I had to put

a quarter in that machine. And so if it didn't come up, cherries

are better. My quarter was gone. So risk has to do

with the probability of loss and how you,

you trade that off with Your preferences against probability of

gain and uncertainty just has to do with how

sure are you. If you had to do an over under on your

profitability, what's the 50, 50 point

risk estimate? That makes a lot of sense. Yeah, that makes a lot of

sense. That's interesting. But

what's your take on. Obviously the commodity that everyone

is most impacted by, at least in the obvious sense, is oil.

Crude oil. Yep, Crude oil. So

I've noticed obviously there's been some, you know, the elephant in the room, there's been

a lot of instability in that, that space, but somebody had said something and

I didn't quite get it, is that the oil

contracts are down a decade out or five to ten years

out. Yeah. So the fluctuation in

prices that we see at the gas pump have more

to do with

speculation. Is that true? No, Did I mishear it?

I don't think it's true. This has to do with the short term supply demand

balance. When the strait of Horamuz was closed, you had 10

million barrels a day less supply at the gate of

the straight of horror moves. So draw yourself a supply

curve and a demand curve and shift that supply curve 10 million

barrels of the day to the left. Your price is going to go high.

Yeah. No matter what we have. And I do this and

there's a different kind of data I'd like to chat about. We have a world

oil model, multi regional world oil model. We forecast this stuff for

the industry short term and long term. Okay. And

so short term is different from the long term. All the oil that's going to

be produced is sitting right there at the wellhead and you have to have

logistics to get it to market. So the supply curve and the demand curve and

the short term are fixed. In the long term they're not fixed.

People can invest capital and go produce some tar sands in

Venezuela or produce some more oil in the Middle East.

And so longer term price effects tend to look a lot different

than shorter term price effects. But both are uncertain.

Both are. So if we gather data on some

phenomenon, if it affects long term prices,

then that data will have a stochastic effect. Right.

It helps you understand them better, helps you reduce your, your

uncertainty of what those prices are going to be. The more data that you gather,

the more definitive, I. E. The lower variance your look.

So a lot of these comments that you hear, they give me

The stupidity in these comments, number

one, the biggest stupidity in the world, is that the future price

is an extrapolation of the past price. And people do

that statistically, it's dead wrong. Economics teaches us that

the future occurs because of future

supply and demand has absolutely nothing to do with the past.

So AI is going to have a really hard time with any economic

problem because the future price that you're going to

see, say right at the gate of the straight of horror moves, it's going to

be a function of what's happening then, not what happened now.

People will speculate into the future. They will solve. They will sign

a long term buy or sell contract. I'll give you a million

you'll sign a contract that I'll buy it. They sell, buy and sell

futures. That just makes the price more and more available

and visible to everybody.

Speculation is a good thing. It's a very, very good thing.

The best thing is speculation trading.

Because it takes risk away from the producers, right? Or no,

it shows you the price. If you watch how people make these

trades, if all of them are made at an implicit price of $80 a

barrel, there's your best guess at the price. Speculation

shows us the price. So this idea that speculation

is bad is really stupid. It's so stupid.

You want more and more and more and more trading. So everybody knows the

price. So the price in the fair, the free market is known to all.

Interesting. That's why people

trade, right? So you don't have to, you

don't have to be data driven about the future price. It's transactional.

Right? Okay. We have short

term disruptions, there's not enough trading. And

until a point, because you're trading on the probability

that the straight remains open versus vibrates

between open and closed versus closed.

And people will speculate that. Thank God for

speculators. Interesting.

What's your take on these poly market type

marketplaces? What's your take on that?

People are speculating. You do? Okay, I love it, I love it.

I think people are betting their probability assessment against yours. They

think they got a better probability distribution of something over yours and

they put their money where their mouth is, they trade on it. And if you

look at the volume of trades that people lay down, that gives

you some notion of the cons, consensus probability of those events

occurring. So the what like on the election?

Speculating on the election. I love those things. I don't play them, but I love

them. Sportsbook is that sports books are

great. They're so much fun to play too.

Bad they rake off so much of the pot as their fee.

But you can watch the data on those and I think

the data suggests that they're very pretty. Pretty

predictive of what happens in a probabilistic sense.

No, I remember one of the. There's a paper from a long

time ago and basically discussed how if you remember the show

who Wants to Be a Millionaire, the crowd, when you ask the crowd a question,

the crowd was more accurate than any of the other mechanisms

that they had, which was phone of friends pass

and something else. But the

crowd actually got the answer right. Far and above.

Yeah. That book, the Wisdom of Crowds. That was it.

Yeah. And it's very good. See, the market is the ultimate wisdom of

crowds. The crude oil market is so heavily traded. There's

no misinformation in that price. There's even a

theory which I subscribe to, that there's nothing the matter with insider

trading. It gets more information into the price and

more correct information into the price. It's

illegal because people say it's unfair. It's.

It's capitalizing on. On insider knowledge and all that.

I'm not so sure I buy that. I want the. When I buy somebody's

stock like SpaceX, I want all the insider and

outsider knowledge involved in my decision whether or

not to purchase that st. And it wasn't all

the insider knowledge was concealed from the market by law.

Yeah, no, that. It's funny you mentioned that because insider trading

has only been illegal since the Great

Depression. Yeah, I remember one of the

most interesting books. It was called the Patriarch and it was

basically about Joe Kennedy. Oh yeah. And

FDR put him in charge of that. But he was a notorious insider

trader. Oh yeah. As you would point out later on, it wasn't

illegal then. Right? It wasn't illegal. It's not even. It's not

illegal in Congress today. Well, yeah,

exactly. I mean there are people who are financial

geniuses. This is crazy. I mean, really. Yeah. These

guys have confident insider information, it's

confidentially delivered. Confidentially. And you can go off

and trade on it. Come on, guys. It's

not a practice that be condoned. Yeah. Shutting down insider

information is about fairness. Then they should be included here.

But we're a little far afield. Getting back to data science, there's a couple of

other things that. One of my. One of my favorite

themes. I saw this during COVID go back to the beginning of COVID

How much data was there on what was curative

or what was ameliorating of symptoms about any treatment

we had, there was no data. Not a lot, but

there was a lot that was suppressed. If it's suppressed, it

doesn't exist. Okay. If you can't analyze with it,

then there is no data. It's like if I flip a coin and put a

piece of paper over it and ask you to call the coin,

it's still 50, 50 to you. I can look under that paper, it's 100 to

me. I know what it is. Your state of information is still 50,

50. You don't have any information when I have it.

So you're going to have to take that into account. So

people were coming out and saying, oh, hydroxychloroquine is going to have

prophylactic effects and curative effects. That's my

subjective judgment. And my brand of data analysis is called

Bayesian, where you come to it with a judgment in

advance and then what data does is

revise your judgment. So,

so people were treating people with hydroxychloroquine and they were

on TV to the extent they could get on. This is great.

60 year old medicine is going to cure it all. As the data began

to came in, come in. If you'd have been a Bayesian,

you'd have said, my initial statistics on

the curative powers of high gloxychloroquine are being

totally not supported by the data. I would have changed

my judgment. That's what data analysis is for,

to get you to change your judgment.

So valuable. You know, you start off an AI

program with judgment as to what all the parameters are. Data

stops the judgment process and makes those parameters

consistent with observations. That's powerful, isn't it?

Yeah, very powerful. And so

hydroxychloroquine fairly quickly started

the, the curative and the, and the amelioration

properties of that drug were pretty quickly

contradicted by the data. Okay, Same with

Ivermectin, the horse tranquilizer. All right. Horse

tranquilizer is going to cure everybody. And, and

pretty quickly contradicted when they tried a little bit of that in small

case studies. It, you know, you might as well have been taking a sugar pill.

And so then we get to the shots and all the people said,

these shots, man, they're going to be magic. They're going to stop transmission.

That's my judgment. We're going to stop transmission. We began to gather

data. The shots were extremely valuable. They didn't stop

transmission, but they cut the severity of the symptoms by three

orders of magnitude. And so they changed our judgment.

The point of data science is get the data to give you

the best possible judgment of probability. That's

what data science is all about. Gather the right data that

can have the most effect on your judgment. That's what we want

from data, is our judgment as to how these

processes work. That's cool, isn't

it? So very valuable raw

material to making me a smarter person.

Having better judgment is consistent with a gazillion observations,

each of which is stochastic in its own right, but they're not

completely stochastic. It's not random. There's a lot of systematic

stuff hidden in there. So the idea of data science is, is

gathering data. If you don't just go out and gather the data, that's easiest to

gather. Gather data that you really need to have

to figure out whether the shot was going to

be strictly prophylactic or whether it's going to cut the

transmission. Well, we figured it out after a while, got enough data that

said it's not going to change the transmission. And you've heard all the political

bitching and moaning about that. Oh God, it's a lousy shot. It doesn't cut

the transmission. It saved lives by cutting the severity.

And that's statistically valid. And so

big success. The data gathering in places like Israel where shots were

mandatory was really interesting. Really showed the prophylactic effect.

God, that's valuable data, isn't it? That's

trillion dollar data. Is it allowed people to

make better decisions? Candace, you're right on the mark. Once you know

that, you can make a lot better decisions both collectively

and individually. I decided to take the COVID

shot because I did a decision analysis of it using my

subjective probabilities. It was a no brainer. How could you be an

anti vaxxer when you made that kind of assessment?

I've done that with allergy shots. Great example. I take

allergy shots, I'm a severe allergy patient. I went in to take

these things. Well, you know, they know what you're allergic to and they stick it

directly into your bloodstream. That's what an allergy shot is. If you're allergic to

peanuts, you get peanuts stuck in your body. And so I asked,

well, what are the, what are the problems here? Well, you know, you can die

of anaphylactic shock. You can drop off the bench and be dead in 30

seconds. Good. That's really a good thing to do. And

they might not work. So I looked at the statistics

and as near as I could tell, about 70% of the time

they Provided significant relief and it was about

1 in 10 to the 7th. So one 10

millionth probability that was going to drop dead and so

more likely to win the lottery or find another. You know

I looked at my values and I rolled back my decision tree. It was a

no brainer. 50 years I've been stepping up to the lady and

having her stick peanuts in my arm which I'm deathly allergic to.

And so you there also other things too that you're

in the doctor's office when that happens. So like if you do like they could

administer isn't that kind of alter the risk profile too? It's not

like you're. I carry an amputee. You wouldn't. I carry. Yeah, like so

you wouldn't go on a rowboat in the middle of the Ocean without your EpiPen

and then take that shot. I did but you know you're not supposed to. All

right, but, but I mean that would also. Yeah but in that

simple example the statistics of 10 to the -7

and the statistics of 70 probability of

amelioration of symptoms, that's the billion dollar data

right there. Simple minded data but

it helps you make can the right decision. I got that out

of the data but I also had to analyze it to find out

how that interacted with my preferences. So it

data analysis is critical to getting the information to

make a good decision. That's what I think. I didn't even

do any analysis. I just looked in books and saw those probabilities ain't

that great. So every time I go in there, hey, I might never walk out

of here again but the odds are pretty low.

That's a great point. So we want to be respectful of your time. It's

almost 40 after the hour. Where can folks find out about more about

you? And I know Stanford does a lot of these lectures. They post them online

or any of yours online. They haven't

been. No, they haven't been but that's why I'm talking to you. Maybe I'll do

that. And if you want to continue. I haven't told you much about data

analysis per se. If you want to fire up tomorrow, I'm

glad to do that. Yeah, let's. Let's talk again soon. We

want to know the uncertainty we're looking at. And the more data you

get, the less uncertainty you have. It gets pretty definitive if you

got 60 terabytes of data but we don't have that

candice. The other thing we'll talk about next time is in real decisions you don't

have any or very little data that's relevant to the decision you're

making. You get the data that somebody gathered on a government

grant, like crime statistics. There's

nothing in it. It's just data that's easy to gather. Okay.

You are awesome. I look forward to speaking with you

again. Yeah. With that, we'll let the

outro music play.