00:00:00
Well, the finance sector is taking a
00:00:01
much closer look at how technologies are
00:00:04
going to be playing a potential role in
00:00:06
their operations in the years ahead.
00:00:08
Things like artificial intelligence,
00:00:10
cyber risk are more on the agenda of
00:00:13
many experts in these areas. And so the
00:00:16
Wharton School is going to be hosting a
00:00:17
conference this week about some of those
00:00:19
potential concerns. Wharton finance
00:00:21
professor Itai Goldstein joins us to
00:00:23
talk about the conference and as well
00:00:26
some of these issues and why the
00:00:28
conversations are more important in this
00:00:30
day and age. Great to see you again,
00:00:32
Itai. Thanks for coming up.
00:00:33
Great to see you, Dan.
00:00:34
So, let's start about just the the the
00:00:37
element of AI in your area in finance
00:00:41
and the potential impact you think that
00:00:43
it might have.
00:00:45
Yeah, great question. So, AI is now a
00:00:47
very hot topic. I think hot topic across
00:00:50
different fields and certainly has a big
00:00:52
effect on finance. If you ask where
00:00:55
finance is going to be affected, it's
00:00:57
probably going to be affected across the
00:00:58
board. So, different areas of finance
00:01:01
will be affected. I don't know that
00:01:02
there is an area of finance that will
00:01:04
not be affected. So, it's it's really
00:01:07
a systematic effect across the board.
00:01:09
And you know, when we put together this
00:01:11
conference in collaboration with the
00:01:13
International Monetary Fund, the IMF, um
00:01:16
we really tried to cover all of it. So,
00:01:20
there will be discussion on how it
00:01:22
affects financial markets. There will be
00:01:24
discussion on how it affects firms. And
00:01:28
when it comes to firms, you know, those
00:01:29
that are probably more exposed are those
00:01:32
that have labor
00:01:34
that is working on tasks that are
00:01:37
potentially replaceable by AI. So, all
00:01:40
this will be affected and will come up
00:01:41
at the at the conference.
00:01:43
And so part of this I understand is out
00:01:46
of some research that you have done
00:01:48
looking at the component of AI and at
00:01:52
times how it could potentially impact
00:01:53
something like trading.
00:01:55
Yes, absolutely. So, we've done work
00:01:58
together with Winston Du, who is a
00:02:01
in the finance department with me, and
00:02:03
Yan Ji, who is in the Hong Kong uh uh
00:02:06
University of Science and Technology.
00:02:08
Um basically, what we did was to think
00:02:10
about how AI traders are going to affect
00:02:14
the equilibrium in financial markets. Uh
00:02:17
you know, everyone is talking about AI
00:02:19
algorithms replacing humans uh in uh
00:02:23
trading. Um and one thing that is nice
00:02:25
about AI is that you can actually go out
00:02:28
and check uh by just doing the
00:02:30
experiment. Uh so, you know, when you're
00:02:32
thinking about humans,
00:02:34
uh then if you want to get an answer of
00:02:36
how will they behave in this situation,
00:02:39
you can either do an experiment, but
00:02:40
doing an experiment with humans is not
00:02:42
that easy. You have to uh recruit 20
00:02:45
students, put them in a lab, give them
00:02:47
instructions, see what they do, and
00:02:49
there are all these concerns about who
00:02:51
is going to participate in these
00:02:52
experiments, and whether it mimics the
00:02:55
real world environment or not. With AI,
00:02:57
you can actually do the experiment
00:02:58
fairly easily because there are no
00:02:59
humans involved. So, basically, what you
00:03:01
do is you just program these AI uh
00:03:04
traders, uh you program the financial
00:03:06
market environment, and then you let
00:03:08
them act, and you see what happens. So,
00:03:10
you know, the AIs that we are looking at
00:03:12
in this paper are the reinforcement
00:03:14
learning uh AI. Uh you know, the sub uh
00:03:19
category of that is known as Q-learning,
00:03:21
which uh is very prominent in in the
00:03:24
industry. Basically, what they do is uh
00:03:27
they are just trying to maximize their
00:03:29
payoff, and they don't know anything
00:03:32
about the environment, who they're
00:03:33
trading against.
00:03:34
Right.
00:03:34
Uh the only thing they know is uh they
00:03:37
have a set of options.
00:03:38
Uh every round of trade, they pick one
00:03:40
option, and then they update, you know,
00:03:43
uh this was the state of the world, this
00:03:45
is what I did, this is what I got. And
00:03:47
they have these huge metrics of state of
00:03:49
the world, action that I took, and then
00:03:51
they keep updating, what do I get for
00:03:53
every combination of state of the world
00:03:55
and action that I took. And and they
00:03:57
kind of learn as they go. So, over time
00:04:00
they become better and better, and they
00:04:02
know how to choose the right actions.
00:04:04
Now,
00:04:05
you know, the system is such that they
00:04:06
have to experiment, so occasionally they
00:04:09
would try something new and see if if it
00:04:10
works, and then they update, but
00:04:12
eventually they converge to what they
00:04:14
learn is is the best outcome. So, we did
00:04:17
that. We run We ran this experiment, and
00:04:19
we found very interesting results that
00:04:21
in some cases, and we identified those
00:04:24
cases, they converge on collusive
00:04:28
behavior. So, you know, what you would
00:04:29
like is traders in financial markets not
00:04:32
to collude. You know, when they get
00:04:33
information, they just trade on the
00:04:34
information. The information immediately
00:04:37
shows up in the price, and then the
00:04:38
system is very efficient. A collusive
00:04:41
behavior is they get the information,
00:04:42
but they don't trade so quickly on it
00:04:44
because they realize that there is
00:04:46
bigger long-term gain if they're not
00:04:48
trading very aggressively. Um, and you
00:04:51
know, there are all sorts of theories
00:04:52
about when humans will collude and so
00:04:54
on. What is interesting is these very
00:04:56
basic algorithms that I just described
00:04:59
to you end up in what looks like a
00:05:01
collusive behavior, where they just
00:05:03
don't trade very aggressively, and they
00:05:05
end up with higher profits in the long
00:05:07
run as a result.
00:05:08
Right. Should we be surprised that there
00:05:10
is this element of collusion potentially
00:05:12
involved in AI? Because I think most
00:05:16
people would think that that term really
00:05:19
associates with how human beings act.
00:05:22
Right, yes.
00:05:23
Um,
00:05:24
so I I think that's a that's a great
00:05:26
question. You know, why do we see AIs
00:05:29
ending up in this more sophisticated
00:05:31
behavior that we tend to expect from
00:05:33
humans.
00:05:35
Um, and people have looked at AI
00:05:37
collusion in other contexts. So, there
00:05:40
are sort of simple experiments that have
00:05:42
been done in much simpler context.
00:05:45
Um, and people have seen this kind of
00:05:47
behavior emerging. I think what was more
00:05:50
challenging and really the motivation
00:05:53
for our paper is to see whether it can
00:05:55
also emerge in a financial market
00:05:57
environment because a financial market
00:05:58
environment is much more complex.
00:06:00
Right.
00:06:00
Uh, you have noise traders, right? You
00:06:03
have all these noise shocks
00:06:06
hitting the trading process and
00:06:07
affecting price. So, collusion could
00:06:09
become more difficult. You have a market
00:06:11
maker. So, some someone out there is
00:06:13
observing all uh, the trades and decides
00:06:16
on the price and sort of acts
00:06:18
rationally. So, when you let all this
00:06:20
interact, the question is whether
00:06:21
collusion still still emerges. And and
00:06:23
we identified two types of collusion.
00:06:26
One is sort of based on a price trigger
00:06:29
punishment if you want.
00:06:30
Right.
00:06:30
So, the idea is that if I am an AI
00:06:33
trader and I see the price going beyond
00:06:36
a bound that I expected, then I realize
00:06:39
that, you know, maybe I should also act
00:06:41
aggressively. And and this punishment is
00:06:43
what maintains the collusive behavior in
00:06:45
equilibrium. Uh, we call that artificial
00:06:48
intelligence in the sense that they are
00:06:50
acting
00:06:51
intelligently, but this is really
00:06:53
artificial. Uh, but then there is also
00:06:55
another type of collusion which we
00:06:57
coined the term, you know, artificial
00:06:59
stupidity, which is that uh, they really
00:07:02
end up colluding because they don't
00:07:04
realize that if they deviate from
00:07:06
collusion, they can actually make a
00:07:08
higher profit in in the short term. So,
00:07:09
they just got used to this kind of less
00:07:12
aggressive behavior and that's what they
00:07:13
end up doing.
00:07:14
So, the importance of doing the
00:07:16
conference right now and you obviously
00:07:18
you talked about the partnership with
00:07:19
the IMF. Uh,
00:07:22
I guess we're coming to a kind of a a
00:07:25
critical mass point here on a lot of
00:07:27
these issues around AI and obviously
00:07:30
things like cybersecurity risk and how
00:07:31
they can all factor in, correct?
00:07:33
Yes, absolutely. So, you know, we
00:07:35
started the collaboration with the IMF
00:07:37
last year. Uh, so this is a
00:07:39
collaboration between Wharton Initiative
00:07:42
on Financial Policy and Regulation that
00:07:44
I'm uh, the director of
00:07:46
and we're doing this conference with the
00:07:48
IMF. Last year we did a collaboration,
00:07:50
they wrote the global financial
00:07:52
stability report on non-bank financial
00:07:54
fragility.
00:07:55
I worked with them on that and we said,
00:07:57
you know, let's just do this conference
00:07:59
to highlight the issues that are coming
00:08:01
around non-bank financial stability.
00:08:03
Conference was a big success and we
00:08:04
said, you know, we can think about it as
00:08:06
an annual event. Every year the IMF is
00:08:08
producing this global financial
00:08:10
stability report and they identify new
00:08:13
issues that are on the agenda of
00:08:14
financial stability and we can do the
00:08:17
conference around that and this year it
00:08:18
happened to be cybersecurity and AI and
00:08:21
how they affect financial stability. So
00:08:23
they wrote the report on that. I think
00:08:24
it's getting
00:08:25
worldwide attention
00:08:28
and and that is a good opportunity for
00:08:30
us to team again with them and do the
00:08:32
the conference on that. So this these
00:08:33
are certainly issues that are high on
00:08:36
the agenda of policy makers.
00:08:38
I I
00:08:39
we talk so much I I think in our
00:08:41
perspective about what's going on here
00:08:43
in the United States, but from a global
00:08:44
perspective on the IMF, these are issues
00:08:47
that they are
00:08:49
dealing with and have to be focused on
00:08:51
on a daily basis.
00:08:53
Focused on
00:08:54
different financial systems all across
00:08:56
the globe.
00:08:57
Yes, absolutely. So that is the mandate
00:08:59
of the IMF. You know, the IMF is kind of
00:09:01
an international organization who's
00:09:03
trying to coordinate
00:09:05
financial policy, macroeconomic policy
00:09:08
around the world. I mean that their
00:09:10
direct activities are basically giving
00:09:12
money to countries in trouble, but they
00:09:14
also follow countries, follow emerging
00:09:17
risks and are trying to provide advice
00:09:20
and and guidance as to what policies
00:09:21
should be undertaken. So AI is certainly
00:09:24
a global issue. I mean there there there
00:09:25
are no borders to it.
00:09:28
Is there while there's a concern about
00:09:31
the level of risk, there obviously has
00:09:33
to be a level of optimism about what AI
00:09:36
could bring to the development of
00:09:39
financial systems as we move forward.
00:09:42
Yes, I think that is absolutely true.
00:09:44
You you know, I mean when we think about
00:09:46
AI, I think we have to start from the
00:09:49
opportunity. Because AI was not
00:09:52
developed to destroy the world or make
00:09:55
the world unstable.
00:09:57
AI was developed because there is an
00:10:00
opportunity to make things better, to
00:10:03
have things work more efficiently. And
00:10:06
we see that I think across the board
00:10:09
that AI is very powerful and can do
00:10:11
things very effectively.
00:10:14
And and you know, when it comes to
00:10:15
financial markets, yes, the idea that
00:10:18
there will be this algorithm out there
00:10:20
that can process all this information
00:10:21
very quickly and trade certainly can
00:10:24
lead to improvement. But at the same
00:10:27
time as we see that developing, I think
00:10:29
it's critical to keep an eye on the
00:10:31
emerging risks.
00:10:33
And you know, what one risk that
00:10:34
everyone highlights when it comes to AI
00:10:37
is that AI is just going to replace
00:10:39
humans, maybe cause massive
00:10:42
unemployment, maybe when they take over
00:10:45
from humans, they start controlling the
00:10:47
system and then who knows where we end
00:10:49
up. So this is a major risk. You know,
00:10:51
the thing I talked to you about is is
00:10:53
not as extreme as that,
00:10:55
but also something to take a look at.
00:10:57
You know, if you have all these AI
00:10:59
algorithms out there that are running
00:11:01
the trades in financial markets, are
00:11:03
they going to end up without directly
00:11:05
communicating with each other? Are they
00:11:07
just going to end up colluding and
00:11:09
limiting competition? So that is another
00:11:11
form of risk that we have to think
00:11:12
about.
00:11:13
What what do you hope that that people
00:11:14
will take away from from the conference?
00:11:17
So I I think it's an opportunity for
00:11:20
people to come together and talk about
00:11:23
issues that are now on top of the
00:11:26
agenda.
00:11:27
It happens to be AI and cybersecurity
00:11:30
and how they affect financial markets. I
00:11:31
think it's really about understanding
00:11:34
what new dimensions of risk are there
00:11:36
that that we have to to look at. Um the
00:11:39
the conference features six papers on on
00:11:43
these topics, AI and and cybersecurity,
00:11:46
but there are also panels. Uh there is
00:11:48
an academic panel, there is a policy
00:11:49
panel, so people are going to talk about
00:11:51
what they think are the main issues
00:11:54
right now and what we should uh look uh
00:11:56
look for. I I think this interaction
00:11:59
between academics, policy makers, and uh
00:12:02
people working in the industry is very
00:12:03
important. It's really uh an opportunity
00:12:06
to exchange ideas and get informed about
00:12:08
what is coming uh on these dimensions.
00:12:11
Great to see you again.
00:12:12
Yeah, good to see you.
00:12:13
Thank you. Itai Goldstein, Wharton
00:12:14
finance professor.
