Latest / Leaders in Lending / The Upstart Macro Index (UMI)

Transcript

The Upstart Macro Index (UMI) · Leaders in Lending

0:02
You're listening to Leaders and Lending from Upstart, a podcast dedicated to helping consumer
0:07
lenders grow their programs and improve their product offerings. Each week, here decision
0:13
makers in the finance industry offer insights into the future of the lending industry,
0:18
best practices around digital transformation, and more. Let's get into the show.
0:27
Welcome to Leaders in Lending. I'm your host, Jeff Keltner. This week's
0:31
episode features my conversation with Paul Goo, the co founder and the head of
0:34
product and data Upstart. Actually this is our second conversation with Paul, and
0:40
his first episode is actually still our most listened to episode of all time,
0:43
so we're excited to bring Paul back. It is a really interesting conversation delving
0:48
into some recent innovations. Upstart is made around measuring the impact of macroeconomic conditions
0:54
on loan performance and really separating out individual performance of loans from the acro economy's
1:00
impact on those in a very quantifiable way, which I thought was a fascinating
1:03
topic. It's a really interesting area of research, so we kind of dive
1:07
into how it works, why, how that can be utilized in what up
1:11
Start is working on in the future. Towards us. It's a really interesting
1:15
topic, one obviously highly relevant in the current environment, and so I appreciated
1:19
having a chance to dive deep with it with Paul. I will say there's
1:23
one correction to make upfront, which is that Paul mentions what he called the
1:26
Upstart Macro Forecast or UMF, and the correct term is Upstart Macro Factor still
1:33
UMF. So just a quick correction and without further ado, please enjoy this
1:37
conversation with Paul Goo. Oh, welcome back to the podcast. Thanks for
1:41
joining us again. Oh, I'm happy to be back. It's grown quite
1:45
a bit since I was last year. It is and yours is one of
1:48
the most popular episodes the original that we've ever had. So apparently alreadience likes
1:52
technical discussions. We talked about AI and lending in a more technical way last
1:57
time, so we're gonna try and give them a little bit more technical stuff.
2:00
Great, let's do it well. The topic of discussion today that I
2:02
really wanted to dive into is the topic of what we call the Upstart Macro
2:07
Index or u am I, because that's been I know, we've talked about
2:10
it a bunch and earnings we've talked about it publicly. We've talked about it
2:14
with our partners, but we haven't disclosed a lot about it or gone in
2:16
depth, and so ill I know some people have gotten one on one,
2:20
you know, introductions or explanations, but I would to spend the day kind
2:23
of diving into it. So maybe the best way to just start is like,
2:25
what is you you am I? Yeah, you AM I stands for
2:30
Upstart Macro Index. And before I describe what it is, I might take
2:36
a step back and just describe how we got here and why we started working
2:38
on this thing that it became known as the u AM I. So if
2:44
you if you take a step back and look at where we were in call
2:47
it mid twenty twenty one. At that time, very loosely speaking, almost
2:57
everybody was doing a good job of repaying their loads. Obviously not everybody.
3:00
There were always defaults, but repayment rates were extremely high compared to historical norms,
3:06
and it had looked like at the time everybody was a genius in lending.
3:10
You could have lent to pretty much any type of consumer and probably you're
3:15
outperforming your expectations in terms of repayment rates and therefore rates of return on debt
3:21
capital deployed. Then you fast forward another year to mid twenty twenty two,
3:27
and the story had flipped. Instead of every type of borrower outperforming, now
3:32
you had almost every type of borrower underperforming. And the change from mid twenty
3:39
twenty one to mid twenty twenty two was dramatic, and you could see that
3:44
in a variety of metrics, but it wasn't totally clear how you would precisely
3:49
tease that out and quantify what had happened in the quote unquote macro environment from
3:53
mid twenty one to mid twenty two. And we started to have conversations with
3:58
a lot of our lending partners of the forum, Well, what's going on?
4:02
I you know, I all I see is that delinquency rates have gone
4:04
through the roof. I don't know if it's because the up Start decided that,
4:10
you know, has upstarts model has gone bad? Has Upstart just sort
4:13
of open the floodgates to low quality borrowers? What you know, servicing degraded?
4:17
What what in the world is going on? And when you see a
4:21
very dramatic change two x three x, they weren't talking large multiple in a
4:26
short period of time. People naturally get worried and they want to know,
4:30
and of course we wanted to know too, and we had a pretty good
4:32
sense from the indirect metrics that we we had, of course kept our eye
4:38
on, but it wasn't precise in the way that we would like to have
4:41
had it, and so we said, well, we really want a precise
4:44
quantitative way to tease apart what we call the macro and the micro effects of
4:49
underwriting. And more precisely, what I mean by that is, if you
4:54
think about what an underwriting model does, the one hand, it has to
4:59
separate risk, meaning it has to say this particular applicant is more likely to
5:02
default than this other one, and they're more likely to default by two times.
5:06
But another thing it has to do is it has to figure out the
5:10
absolute levels of default risk. So that first baseline borrower, what does a
5:15
nearly risk free borrower look like, what does a super risky borrower look like?
5:18
And what is the average population and default rate going to look like?
5:23
And the reason that those are such separate problems is that for that second problem,
5:28
you really have in some sense very limited amounts of historical data to train
5:34
on in the way that you would like to train an mL model. Whereas
5:38
for the first problem, the risk separation problem, you have an enormous amount
5:42
of data that you could look at. Basically, every single default that's ever
5:46
happened is mostly a story of relative default. It tells you that, within
5:48
a given period of time, how likely was a particular borrower characteristic to default
5:54
compared to some other borrower characteristic. Whereas to look at the average level of
6:00
default in say a year, Well, you only have so many years on
6:01
record, and you certainly don't have the thousands or millions of data points that
6:06
you would like to have. I mean, human history barely goes back that
6:10
far. And certainly we weren't recording our loan repayments, you know, going
6:13
back to those days. So that makes it very challenging. And so we
6:16
sort of have are just like a handful of meaningfully different periods recessions and non
6:21
recessions. You can think of it as like, but maybe like twenty data
6:24
points in recorded history, and twenty is just not enough for any kind of
6:28
serious machine learning training problem. And so we separated this sort of micro question
6:33
and this macro question, and we said, okay, how can we just
6:36
precisely measure the macro component and know whether a period of time is seeing a
6:44
increase or decrease and of how much compared to pass periods in its average level
6:49
of default across the population. So that's that's that's what we wanted to know,
6:54
and about eighteen months ago we started pouring a lot of our a lot
6:59
of our research science attention to this problem and the end result is something that
7:03
today we call the u am I the Upstart Macro Index, Yeah, the
7:08
Upstart Macro Index. Essentially, it takes a It aims to tell you if
7:14
you had constant borrowers, if you're just constantly making loans to the exact same
7:19
group of people across different periods of time, what would the delta in their
7:26
delinquency rates be in different days, weeks, months, years, whatever time
7:32
frame you want to compare. And the way it does this is by making
7:36
use of what we call a reference model, because of course you can't actually
7:40
find exactly identical borrowers every single year to make loans too, and no matter
7:46
how you control it with one variable two variable, you might say, well,
7:48
let's just look at you borrowers who are six eighty to seven twenty on
7:53
the FICO scale, and that would be a very simplified way to do this,
7:56
but of course there are a million other things to know about these or
8:00
is other than their phycho score, And so the way to do that in
8:03
a more holistic and continuous way is with what we call a reference model.
8:07
So you say, what if we underwrote all of these these loans with exactly
8:13
the same model, because of course, in reality, every loan you have
8:16
on your books was underwritten with probably a different model at a different period of
8:20
time. But if you instead underwrote them with exactly the same model then made
8:24
predictions for every single month that you've subsequently observed, how do the actual delinquency
8:31
rates and default rates compared to what a constant reference model would have predicted.
8:35
And what that tells you the reference models doing the work of controlling for the
8:39
consumer. And then what you can observe is, of course the actual loss
8:43
rates in real calendar months compared to what the model would have predicted on this
8:48
sort of abstracted, hypothetically constant kind of calendar time period. And so the
8:54
deltas in those observations compared to a sort of now controlled expectation of default or
9:03
delinquency is going to tell you how much more or less risky twenty twenty two
9:09
was compared to twenty twenty one, compared to twenty nineteen, compared to twenty
9:11
seventeen. And so that's that's what we've developed. We call it the Upstart
9:16
Macro Index, and it does a really good job for us of both explaining
9:20
how sort of past delinquency patterns have played out and why they've played out that
9:24
way, and also actually in being able to make forward looking statements of the
9:31
forum. Well, if the macro environment gets fifty percent worse, then this
9:37
is what will happen to your loss rates. And being able to make statements
9:41
of that form, of course, then allows us to make useful statements about
9:46
how we should be building in buffer and conservatism to the underwriting for any given
9:52
lending partner, you know how aggressive or conservative their parameters are set in how
9:58
much lending they wanted to. So a really useful both backwards looking and forwards
10:03
looking. So I got a number of questions to dive off this one of
10:05
the things you said, But first I kind of want to be really clear
10:09
when when you produce the index, it's a it's a singular number for any
10:11
given month, as I understand it, what is how should still want to
10:16
look at the number and say, what is this? What does this tell
10:16
me? If it's a one point zero, a two point zero, a
10:18
zero point file, like, what what's the actual index? And how do
10:22
I think about reading it? And then I want to dive into a bunch
10:24
of things that you said about this. Yeah, so the index is is
10:31
condensed down to a single number, as you just said, and that single
10:35
number can be can be applied to any any length of time. So usually
10:39
we talk about it in one month periods because most of our loans have one
10:43
month repayment period, so it makes a lot of sense to look at it
10:46
that way. But you could look at them in days, weeks, years,
10:48
quarters, however you like. But four a period of time, you're
10:52
going to have a number, and that number is going to range called between
10:54
zero and three generally, though of course there's no theoretical upper meant to to
11:01
how high the um I could go. But what we've state, what we've
11:05
normalized the numbers to mean, is that is a one dotto generally is meant
11:09
to correspond to a what we think of as a sort of historically normal time
11:13
or at least historically normal time in our recorded history, but you can generally
11:18
think of as being the sort of five years before covidum, and those those
11:24
five years there was relatively sort of a stable macro environment. We did not
11:30
see large changes. You know, again, taking that reference model approach,
11:33
if you're just underwritten all the loans using the same model, you would basically
11:37
see the same sort of results across across the different calendar years, whereas u
11:43
oras after that. Of course, once you get into the COVID world,
11:46
you see both large positive and negative shocks to that, which is exactly you
11:50
know, what this methodology is hoping to tell us about. Yeah, I
11:54
imagine that the COVID period, particularly the very serious economic impact than the rapid
12:00
an input of stimulus and then the rapid withdrawal of stimulus, was like changing
12:05
of macro conditions on steroids, or at least at fast forward speed in terms
12:09
of a normal cycle, which I think is kind of what was some of
12:11
the impetus for this is that kind of rapidity of change. What we usually
12:13
think of is maybe not a super slow but but not not as rapid of
12:18
a of a changing environment as we experience over the last twenty four months.
12:20
Yeah, totally. I mean, it was almost a perfect couple of years
12:26
to showcase just the difficulty of the macro prediction problem, which you know,
12:31
we've we've always a felt was not really a good a good prediction problem in
12:37
the sense that these you have so few of these events, as we talked
12:39
about, and each shock looks really different, and so trying to predict them
12:43
based on you know, pass shocks looks kind of silly. And I mean
12:48
April twenty twenty was probably the most extreme example of this, where April twenty
12:52
twenty hits And of course we have we have we've built over the years plenty
12:56
of what you'd call macro prediction models that look at things like unemployment rates and
13:01
UH and and and UH and various sort of macro indicators that unemployment generally has
13:05
in past shock's been the single biggest driver. But what happened in April twenty
13:11
twenty while unemployment rates went through the roof a sort of um never before seeing
13:15
a rapid rise in unemployment, and yet what happened to loan repayment, Well,
13:20
it turns out that within a very short time, loan repayment did not
13:24
only did not get worse. It became the best period of loan repayment probably
13:30
ever in all of human history. Um and uh. And it was because,
13:33
of course, the you know, the government responded very forcefully to the
13:37
increased unemployment the pandemic and um did a lot of stimulus in various ways.
13:41
And and people had a lot of extra cash to make their loan payments and
13:48
uh and and not as many places to spend it because they were locked up
13:50
at home um and and so you you've got exactly the opposite effect of what
13:54
sort of historical precedent would have told you. And then if you look at
14:00
twenty twenty two, you get almost the mirror image of that, where you
14:03
have extremely low unemployment, basically pushing records in terms of how low unemployment gets,
14:09
and yet you have this rapid rise in delinquency rates. And it's for
14:13
exactly the same reason. The stimulus ends. People still had jobs, But
14:16
it turned out that printing money and giving it to people had a bigger impact
14:22
on whether they would repay their loans and seeing if they were working or not.
14:26
So I want to ask about kind of how traditional lender. I mean,
14:30
this is certainly the topic of how is the macro economy playing into loan
14:35
performance? Every time I talk to, lenders are looking at inflation rates and
14:39
interest rates and unemployment rates and trying to understand what that's going to mean or
14:45
is doing to a given portfolio. So how does how do you think traditional
14:48
lenders have looked at the kind of problem you're trying to model, or at
14:52
least the data point of, like what is the macro environment doing to the
14:56
loss rates in a pool? In? What are the as in the way
15:00
it's traditionally been done that you were trying to solve for when you thought about
15:03
building out the U and my methodology. Yeah, so I think there's there's
15:07
spectrum of approaches, but they share a couple of things in common. The
15:11
first is that because many traditional lenders don't change their underlying models very much,
15:16
or they have very simple ones that are have a very limited number of variables,
15:22
they can kind of make an assumption and it doesn't seem as crazy as
15:28
it does in our world, though we'll get to why it still suffers from
15:31
something the same challenges, but it doesn't seem as crazy to make the assumption
15:35
that, hey, we are underwriting with a constant model and so the barrow
15:39
wars that we produce across different years should roughly be the same. And if
15:43
they're roughly the same, then I can understand where the macro is just by
15:46
looking at the average default rate in my own portfolio. And this is this
15:52
is sort of okay. It's sort of okay because if you actually don't change
15:56
your model in any way, then actually you do have a fairly sort of
16:00
constant model. But of course it also assumes that your barrower is your barrower
16:06
mixes unchanging across time, that your sort of pricing approaches have not changed across
16:11
time, and of course that your model actually hasn't changed in any way across
16:15
time. And of course we know that even if you just were using say
16:18
the Phyco scored down right, of course the Phyco score distribution that are changing
16:22
across time and inputs to it, you know, are are are getting changed
16:26
by you know how medical you know, bankruptcies are handled in sort of different
16:30
things like this. So there're actually a bunch of these changes happening under the
16:33
hood. And maybe they're not that extreme, but but put together, they're
16:36
probably each you know, ten twenty percent type effect, and soon you could
16:38
be in this world where you might be seeing default rates that are fifty percent
16:42
higher and now you have to you have to ascribe that to macro, when
16:48
in reality it might just be that your barrower mix change twenty percent and your
16:52
your underwriting model changed twenty percent, and and there you are with the fifty
16:56
percent. So that that's that's that's pretty inherently challenging to measure macro in And
17:03
then I think on the sort of response side to macro, I think it's
17:07
it's also when you don't have h we don't have sort of a precise factor
17:14
that's able to separate what your underwriting models is doing and what what is happening
17:19
in the macro. There's always this temptation to say, well, we can
17:22
fix the problem just by changing the underwriting model. So if we only focus
17:27
on higher phycho barrowers or higher income borrowers, then then we don't have to
17:32
worry about the effects of a recession or the effects of a macro shock.
17:37
And I think that that is an approach that really can get you into trouble
17:41
where you know, you look at a lot of pass recessions something like two
17:44
thousand and eight, and you look at what happened to credit card delinquency rates
17:47
in two thousand and eight or mortgage delinquency rates in two thousand and eight.
17:49
And what you see is actually that the largest multiple increases to delinquency and default
17:56
rates were in high phyco buckets. Maybe it's a little counter tuitive, but
18:00
then if you think about what was happening in two thousand and eight, of
18:02
course, one thing is that the biggest stress for on people who were buying
18:04
large homes and had big, big mortgages, and and of course that's going
18:08
to be a primer consumer um. And another thing is that, of course
18:12
it matters quite a lot, you know, the kind of shock in the
18:15
economy and whether that's something that's going to impact more prime people or less prime
18:21
people. But but some of some of the less prime people actually m in
18:26
some sense are like always in a recession, they are kind of living paycheck
18:30
to paycheck in a way that means their lass rates are are normally much higher,
18:33
but it's also harder for them to get so much higher. Whereas you
18:37
know, when you have someone who goes from being comfortably sort of like you
18:41
know, putting money in the bank every month to suddenly loses their job.
18:45
It's it's sort of a fairly tremendous change in their ability to pay things.
18:48
And of course the dead oppligations are much larger. There's sort of less kind
18:52
of ability to easily make up for it just by um by you know,
18:57
uh going to backup sources of cash and so there's these different effects that play
19:03
out. But but it's it's very hard, I think nearly impossible to solve
19:11
for macro problems by changing the borrower level characteristics that you want to write too,
19:17
because there are no set of characteristics that are inherently a recession proof or
19:21
macro shock proof and um and even if even if there were, frankly,
19:26
I think we don't realistically have the historical data to identify what those are today.
19:32
The other thing that occurred to me is I was reading through the approach
19:34
to you, m I is the sense that particularly to see things in real
19:38
time, you know you not only mean what you've really said is to do
19:41
this right, I need a way to hold the bar or based constant,
19:44
some sort of a constant reference model, the credit model constant. But I
19:48
also need really an individualized risk prediction and ideally a month based risk prediction,
19:52
because if I want to see how it happened, what happened in this month,
19:55
I want to see that I expect this some months or every you know,
19:59
every given month of se reasoning that could be for your loan portfolio exists
20:02
in the portfolio somewhere, and you want to know is this loan that went
20:06
delink when a month thirty two? Was that unexpected or expected? And so
20:08
you need really a pretty precise model of risk to be able to see this
20:12
in something like real time, I mean not real time, but to see
20:15
it rapidly and quickly and to be able to adjust. And so it occurred
20:18
to me that the kind of individual risk prediction and loan level and month level
20:25
risk predictions in the reference model are really critical to being able to see not
20:29
just with granularity but with speed what's actually happening in the environment and how it
20:33
might be changing than the underlying from what the underlying predictions might have expected.
20:37
Yeah, that's that's totally right. If you only have lifetime default probability predictions,
20:44
or you only have those for certain buckets of borrowers, and you're not
20:48
going to be able to do this then, and you won't be able to
20:49
control for the borrower mix at all, well with a reference model that doesn't
20:52
consider enough individual characteristics. And if you if you only have a lifetime prediction,
20:59
then of course you can't compare what's going on in different periods of time.
21:03
So you really need a model where you have a lot of confidence that
21:06
you know, the month seven prediction for this particular individual really is a good
21:11
prediction, at least in relative terms relative to your predictions in the other months
21:15
and relative to your predictions for other borrowers in that month. And if you
21:19
have confidence in that, if we think of as you know, if there
21:23
are months on the sort of X axis and borrowers on the Y axis,
21:29
you need to feel like both your axes are really doing a lot of work
21:32
for you, that it's sort of an unbiased estimate. And if that's the
21:34
case, then you can actually take that point on the matrix, if you
21:38
will, and compare it to other points on the matrix, and it tells
21:41
you something meaningful. And of course if you don't have that, then you
21:44
can't do that comparison. Yeah, so I did want to ask you call
21:48
this the macron actually talked about the macro and micro prediction levels of the model.
21:55
And it's interesting because macro can be a reference to the macro economy,
21:57
the kind of broader world, but it's also kind of size metric macro and
22:00
micro, And so I kind of want to just ask the question, like,
22:03
how do you think about the relative importance in any given moment or overall
22:07
for the micro portion of what you might call the core risk modeling that up
22:11
starts being focused on and this kind of macro level that's maybe harder to predict
22:15
but now somewhat more measurable through you am I how do those kind of stack
22:19
up in terms of how important they are to a lender into a lender's portfolio
22:23
performance. Yeah, so there, I might first say they're both really important.
22:30
They're both important in the context of you know, fixed income of any
22:36
kind has this property that you can lose a lot more money than you can
22:38
make. That's that's the nature of fixed income, and that means, you
22:42
know, generally, anything that can meaningfully increase your default rates is something you're
22:48
really going to care about. And certainly both the macro and the micro here
22:51
fall into that bucket. If you get those really wrong, it's going to
22:55
really affect the return you were expecting. Now, having said that, if
22:59
you were to just force a competition between the macro and the micro, how
23:03
much can losses vary based on macro and how much can loss is based on
23:07
micro? I mean fatly, then it's not even close. Because the macro,
23:11
you know, you can get you can get a period of time that's
23:14
three or four maybe in an extreme pace, five times as bad as another
23:18
period in micro. Of course, you can get borrowers, and not just
23:25
theoretically, but even identifiably. For example, in our model, you can
23:29
identify borrowers who are like eighty percent likely to default, and you can identify
23:33
borrowers who are zero point eight percent likely to default. So you know,
23:37
you're talking about a one hundred x spread between the sort of highest and lowest
23:42
risk in micro, and you're talking about a five to one spread in the
23:48
highest and lowest risk in macro. And so, of course it's so much
23:51
more important to get the micro right than the macro right. There are so
23:53
many more ways that you know, even in the best of times, you
23:57
know, if you try to go to market and underwrite with with an extremely
24:02
poor performing model, just everything will go wrong. And because not only because
24:07
your model will be wrong. You know, you face adverse selection because other
24:11
lenders will pick off at lower aprs the better borrowers. I mean, just
24:14
so many, so many things can can go wrong in micro, whereas in
24:18
macro things can also go wrong, such as you know, when you go
24:22
from a really really unreasonably good period like twenty twenty one to really a sort
24:29
of sharply bad period like twenty twenty two, that's that's going to significantly and
24:33
negatively impact lender returns. But it's still it's still going to be on you
24:37
know, you know, if we're talking about annualized return impact, I mean
24:42
depending on what your risk mixes. I mean you're still talking. You know,
24:45
you can count it on your two hands, like the number of percentage
24:49
points of impact that's going to have. WHEREA if you get the micro wrong,
24:52
I mean, there's there's really no limit to how much money you can
24:55
lose doing this. And and of course on the flip side, you know,
24:57
when things are going well, micro is what enables you and has enabled
25:02
the lenders that use our platform over the years to be able to make dramatically
25:07
more more income running a lending business. Because micros what enables you to confidently
25:12
expand your approval box while holding your risk levels constant. So if you can
25:17
improve fifty percent or one hundred percent or two hundred percent as many as many
25:22
borrowers as you previously could, but you get the same risk levels, then
25:25
of course you know suddenly now you're able to meaningfully change the amount of lending
25:30
income being general. Yeah, interesting, Well, I want to dive in
25:33
a different topic you made this comment about, like if you can model out
25:37
if the world the macro got fifty percent worse or I'd love to say fifty
25:41
percent better. Sometimes I worry we talk about the macro, we only talk
25:45
about bad, talk about macro getting worse and protecting, and not the fact
25:48
that, particularly from where we sit as we're recording this, there's a lot,
25:52
probably more upside for the macro getting better than the substantially worse over the
25:56
next at least twelve or eighteen months. But the interesting thing is you haven't
26:02
talked really about what I would consider traditional macro variables. Right, Like most
26:04
people when they talk about how is the macro doing? You know you have
26:07
a um I you'll talk about where it's at, But most people look at
26:11
inflation unemployment, interest rates, maybe savings rates, things of this nature that
26:15
are very clearly tied to what we think of as a macroeconomy GDP or something
26:19
like that. What is there a relationship or those part of the U and
26:23
my model. Talk to me a little bit about how like traditional views of
26:26
what is happening in the macro relate to the Upstart macro index, or how
26:30
you think about the impact of macro and the lending portfolio. Yeah, yeah,
26:36
they're they're they're super correlated. In short, the upstart measured you and
26:41
I. We like it because it has the nice property that it's directly applicable
26:48
to the math you want to do on lending um, and because it's it's
26:52
really specifically designed to tell you this period is a period in what you can
26:57
expect twice as many deligencies the same borrower pool as that period. You know
27:02
that it lets you do math in a very precise and direct way. But
27:06
of course, UM, it would be really concerning if this index that's meant
27:11
to measure the macro is totally unrelated to the sort of traditional macroeconomic variables,
27:18
which, of course you know, though each one maybe only tells part of
27:21
the story. They are collectively, you know, a pretty good description of
27:23
the things that we care about in the economy and UM and so we have
27:30
that that's been an important part of developing the UUMI is actually building a sister
27:33
model that tells us what we We haven't found a nice, sort of externally
27:38
marketable name for it yet, but if you think about there's sort of um
27:42
I as measured in the sort of collective repayment rates of the of all Upstart
27:48
platform originated loans where we have, you know, access to the data.
27:52
But then there's there's uum I as you would predict based on what you're seeing
27:59
from the the sort of official macroeconomic data, the unemployment rate, the inflation
28:03
rate, the personal savings rate, and and what you want is you want
28:07
to see that these two measures, they sort of internally measured UMI and the
28:11
externally inferred UMI are largely moving together, highly correlated and kind of moving together,
28:19
and UM and that's that's exactly what we see. We've been able to
28:22
see that, and actually you can UM you know. The maybe short and
28:26
snappy version of this is if you just want what single variable released by the
28:32
FED can tell you roughly where the UMI is going, and the answer would
28:37
be the personal savings rate. That of course is a bit cheating because the
28:40
personal savings rate itself is a bit of an amalgamation of a bunch of variables.
28:45
It's, you know, looking at personal and broadly looking at personal income
28:48
and personal expenditures. But of course what makes up personal income. You know,
28:52
people have to work, so there's employment rates baked into that, there's
28:55
wage wage growth baked into that, and of course on the personal expenditures side,
29:00
there's sort of your real rate of spending, and then there's there's the
29:03
inflation rate. And so you've kind of got this one variable that the Fed
29:06
publishes, you know, the end of every month, and it tells you
29:10
roughly the things that really matter unemployment, inflation, spending, and wage growth.
29:15
And the net of that sort of is like how much money people have
29:18
left over, which is a really good proxy for, you know, how
29:22
much money people are going to have available to service their their debt. And
29:26
if you just look at that one variable, that's probably like you know,
29:30
mostly mostly correlated with how we've measured you and I over the past years.
29:36
Yeah, it does seem like a i won't say a cheating variable, but
29:37
it is an amalgamation that that takes into account a number of the things that
29:41
people traditionally think of as pretty impactful to at least particularly unsecured credit, where
29:47
there's you know, it's kind of the I want to say the riskies,
29:49
but the bottom of the payments to act in any ways, where the first
29:52
thing to take take a hit when one times get tough. So one thing
29:56
we haven't talked about today at all, and I'm sure our lending partnership seems
30:00
to be very upset with me to be this long to ask this question,
30:03
But you know, I think there's obviously when you and I goes up,
30:08
macro gets riskier. There's this question of is that a direct indication of loans
30:12
under or overperforming? And how does the sense of the macro play into how
30:18
you actually work with a lending partner to assign a level of risk to a
30:22
borrower at time of underwriting and understand the returns produced by that portfolio. So
30:26
I think the thing that I hear a lot is like does a U am
30:30
I above one? Meaning all my loans are underperforming because the macro is high,
30:33
and therefore I'm losing money and that's a bad thing and I should stop
30:36
lending or whatever. So talking about how you leverage this in the context of
30:40
partnerships with originators to kind of calibrate to what's going on and help them understand
30:45
the level of risk in the environment that's going into their originations. Yep.
30:49
Yeah. So in short note, it does not there's no direct way to
30:56
say you am I is a certain number, therefore my loans are either we
31:00
are underperforming. To know that, you need to know one more thing,
31:02
which is essentially what you and I were you assuming when you made the loans
31:07
in question, and that that particular thing we call the um the macro forecast.
31:14
And if you very roughly speaking, if you just for any batch of
31:19
loans, if you take their UMF and you then look forward and say,
31:26
okay, I made this batch of loans in January of twenty twenty two,
31:30
and at the time we had a UM of one dot O and and now
31:34
the UUMI is one point five. Then that sort of implies that you're going
31:40
to have actual losses that are fifty percent higher than forecasting. In that case,
31:45
sort of the example you gave as valid. Now fast forward to to
31:48
more recent times, and many of our lenders actually have a much higher UM.
31:55
Imagine your UMF is two doto, but the uum I that's in observe
32:00
sense is only one point eight. Well, then, actually, even though
32:02
one point eight is a really high number by historical standards, UM it's actually
32:07
twenty or it's actually ten percent lower than the UM that you assumed. And
32:13
and so that means that your loans were priced as if you were operating in
32:16
a two doto type world. But actually you've been seeing a one dot a
32:20
type world. And now you're going to see defaults and delinquencies that are running
32:24
on average ten percent lower than than than you were expecting, which of course
32:29
means uh, you know, more more sort of net interest, more more
32:32
excess returns, and so returns will be higher than than expected. And of
32:37
course you know it, Actually it matters how this plays out over the whole
32:40
sort of term of the loan. But um but you know, just roughly,
32:45
as an approximation, if you think about a single point in time,
32:46
this kind of tells you. This, this ratio between the um I and
32:50
UM tells you how things are going to be. And so if you underwrite
32:54
with a fairly high UM, then of course there's more room for um I
33:00
to positively surprise against that. On the other hand, if you want to
33:05
write with a high UMF, you're going to be passing up a lot of
33:07
borrowers and pricing a lot of other ones out of out of out of converting,
33:12
just because you're assuming that you know, the world is in a really
33:15
bad place and you want to, uh, you know, lend to people
33:19
who, uh you know, who are going to be going through a really
33:22
hard time. And of course if if if they aren't, or if you
33:27
know other lenders are not seeing it that way, then those borrowers are going
33:30
to go elsewhere. How how can you help a lender think about I mean,
33:37
what a UUMI or UMF of two or one point eight or one point
33:40
five, Like, I feel like that's not something that at least as of
33:43
yet to be great for you. If that became something about which people had
33:45
an intuition and a made a natural understanding, it was like a number they
33:50
were used to kind of thinking about where it may trend and where it may
33:52
fall. But how do you think about, like what causes a um I
33:57
or UMA to get to a certain level? Can you I mean, I
33:59
think that scenario plan. You know, if you think about just the Dot
34:01
Frank stress test kind of thing, if unemployment goes here and somebody goes here,
34:05
like, how bad does that get? What will have my portfolio?
34:07
How do you think about connecting the dots between those that the concept of you
34:12
know what might happen and what that'll mean to you know, where these loans
34:15
could be performing. Yeah, I think they're two good ways to build intuition
34:17
around at these UUMI index values. The first is to go back to that
34:22
personal savings rate, which is largely correlated, not perfect but pretty good and
34:25
again includes the unemployment and inflation, all all the pieces, and and what
34:30
you can basically see is that you can think of it as a historically average
34:34
level of the personal savings rate. Let's just be round numbers. Call it
34:37
on the order of ten percent, roughly corresponds to a sort of normal UUMI
34:43
condition of around one point zero. And then you know what happens, of
34:47
course, if the personal savings rate goes up. You know, if the
34:51
personal savings rate goes up to something like twenty percent, which you know basically
34:55
never happens except in in twenty one when you know people are given lots of
35:00
money. Yeah exactly. Um, but you know when you when you get
35:06
something like that, then you see you and I basically get cut in half
35:09
down to about zero point five. And again I'm just speaking in round numbers
35:14
here and then um uh. And then on the on the other hand,
35:16
UM, you know you can see you and my values go uh, you
35:21
know, go from one point zero to two point zero. If you basically,
35:23
you know, roughly speaking, cut your personal savings rate in half,
35:28
which is, you know, on the order of of what we've seen over
35:31
over the past eighteen months or so. And then the other sort of way
35:36
to build intuition is just to look at different periods of time. So again
35:38
you and I one dot O roughly corresponds to the five years before the COVID
35:43
period. UM in the sort of um, happy days of free flowing stimulus
35:49
money. Um, you know, you've got you of mys that go down
35:52
to nearly zero point five and then um and then uh, you know,
35:54
stimulus running out producing that large shock takes you, you and I over one
35:59
one one one point five. And then if you were to sort of look
36:02
at historical values of personal savings and you and I sort of asked the question,
36:06
well, how bad can it get? At least how bad have we
36:08
seen it get historically? Probably the answer is something like, you know,
36:13
in the worst times ever that again in recent recorded history, we have all
36:16
the relevant economic data, maybe something like the worst month of the Great Procession.
36:20
We think you and I would have been in the high twos. Interesting,
36:23
So I get two more things I want to hit on, I think,
36:27
and then I think we've kind of at least gotten a good intro for
36:30
people into this concept. One is, we didn't really talk about this,
36:34
but You and I as you've built it today is really based explicitly on unsecured
36:38
installment loans, particularly the kind that Upstar partners with banks and credit unions to
36:43
originate. Are there intentions to build something like You and I for other asset
36:47
classes? Is it applicable to other assets? Like? How do you think
36:51
about the usefulness of it as a construct or the limitations of it as a
36:54
construct given that it's really a measurement specifically on a particular category of consumer credit.
37:00
Yeah, great question. Um so we don't we don't see um I
37:04
as something that you just develop methodology once and it's done. I mean,
37:07
nice if it was like that. But really it's a very sort of rich
37:10
um surface to explore, because there's not only the question you mentioned, which
37:15
is a question of you and my for different products, but there's actually you
37:19
a mid for different types of consumers. And while maybe on average they will
37:23
all move together, they don't, of course exactly all move together. Their
37:27
slopes are a little bit different, their sensitivities are a little different, and
37:30
um, and so what we really are. You can think of the progression
37:35
of you and my methodologies as going from first you sort of assume there is
37:38
just one UOM quote unquote working off one reference model. Then you have one
37:46
UUMI that's produced actually by an amalgamation of reference models, and that sort of
37:52
gets, you know, more and more continuous as you go from you know,
37:53
one to two to end and then and you really get to a methodology
37:58
that that's sort of much more continuous with the back to your reference model.
38:00
And then you have a sort of second dimension of you and my methodology improvement,
38:05
which is really about going from assuming sort of quote unquote the average borrower
38:08
to really making it something that where the you and I is interacting with the
38:14
borrower characteristics, not unlike how we see all the sort of the other work
38:19
we do in in in the sort of prediction realm is there's just interaction effects
38:22
going on between almost any given characteristics of a borrower, and of course those
38:29
also interact with macro and so those should those should be considered. And then
38:32
the last sort of dimension is is the type of loan it that it is,
38:36
and that's you know, true of it, of its size, of
38:38
its duration, of its type, and so that could be it being an
38:42
auto loan, a personal loan, a three year loan, a one year
38:45
loan, or um or any sort of combination of these. And and so
38:50
there's there's there's a lot of sort of continuing research into you and I.
38:53
That's that we're doing. And having said all that, the good news is
38:57
that we think even this first of you and I, as the first approximation
39:01
if you just want to answer the question of under what conditions will my loans
39:07
over and underperform and by roughly how much, you're going to get a very
39:12
good approximation just from you and I V one. That is is true,
39:16
it's built on unsecured personal loans, but you know, we look at it
39:20
in the context of our other products like our Autorefi products, and again it's
39:23
it's not quite as precise, but it does give you a really close first
39:28
order approximation of your level of over and underperformance, again looking at those uom
39:32
I and uu F ratios of the changes across time. So it does a
39:37
good job explaining, and that gives us more confidence in the methodology because it
39:42
tells us that there's something very general happening in addition to the sort of many
39:46
smaller but more specific things going on within each product, within each type of
39:51
borrower, and so on and so forth. Interesting. I guess my last
39:54
question for you is it sounds like this is early days in the development of
39:59
you am I like a very solid foundation, but compared to the work you
40:02
put into understanding micro risk is still a pretty pretty immature in comparison and has
40:07
a lot of room for improvement. And you also mentioned if people want to
40:09
get an intuition, they should look at the numbers. So if people are
40:13
interested in looking at the numbers or following along as the methodology and kind of
40:19
the insights coming out of the work into macro and the u m I evolve,
40:23
is there a place they can go to see those or they publish somewhere,
40:25
is there are place they can follow along what's happening, Like where can
40:29
people go to learn more and keep up to date with you know, what's
40:31
happening to um I and what's happening to the work into looking into this macro
40:35
concept. Upstart dot com slash You am I just out and we're really excited
40:42
about it, and it's going to get better and better. But it's going
40:45
to have the core data there and there's going to be more ways to engage
40:46
with and play with it and really understand the numbers. So yeah, go
40:51
go there, check it out, see how those values have changed across time,
40:53
what they were in different periods of time, And I think that will
40:57
really build a good, good level of intuition for this. Certainly, it's
41:00
our hope that this would become a tool of general interest beyond those who are
41:06
either making loans through our platform or even making loans in general. But really
41:09
we think it's a it's a really it's a really good and useful tool to
41:15
just understand what's going on in the world. Upstart dot com Slash You am
41:20
I. It's not a very original you are l very straightforward and easy to
41:23
remember, so I guess that's probably better than original, and I look forward
41:29
to see that. I'm actually really excited that there's we're going to be publishing
41:31
the data set and so kind of a historical data set. I think there's
41:35
a lot of rich surface area for innovation on top of it, so I'll
41:37
be excited to see. And if any of the listeners have questions or feedback
41:39
on what's useful or what you'd like to learn more about the concept, please
41:44
do right to us or right to just me and Jeff at upstart dot com,
41:46
because I think it's actually a really interesting space and we'd love to have
41:51
real input from all of you in terms of where where we take the research
41:53
and what we learn next. All right, Paul, thanks for taking the
41:55
time to joyces Is. I think a great first dive, and I appreciate
41:59
your back for your second time on the podcast. Awesome, great, thanks
42:02
for having me. Upstart partners with banks and credit unions to help grow their
42:07
consumer loan portfolios and deliver a modern, all digital lending experience. As the
42:13
average consumer becomes more digitally savvy, it only makes sense that their bank does
42:17
too. Upstarts AI lending platform uses sophisticated machine learning models to more accurately identify
42:23
risk and approve more applicants than traditional credit models. With fraud rates near zero,
42:30
upstarts all digital experience reduces manual processing for banks and offers a simple and
42:36
convenient experience for consumers. Whether you're looking to grow and enhance your existing personal
42:40
and auto lending programs or you're just getting started, upstart can help. Upstart
42:45
offers an end to end solution that can help you find more credit worthy borrowers
42:50
within your risk profile. With all digital underwriting, onboarding, loan closing,
42:55
and servicing, It's all possible with Upstart in your quarter. Learn more about
43:00
finding new borrowers, enhancing your credit decisioning process, and growing your business by
43:06
visiting upstart dot com slash foward dash Banks. That's upstart dot com slash foward
43:12
dash Banks. You've been listening to Leaders and Lending from Upstart, make sure
43:15
you never miss an episode. Subscribe to Leaders in Lending in your favorite podcast
43:20
player using Apple Podcasts. Leave us a quick rating by tapping the number of
43:23
stars you think the show deserves. Thanks for listening. Until next time,