# Behavioral Economics Anomalies:Then & Now with Alex Imas & Richard Thaler | Markus Academy | Ep. 151

https://www.youtube.com/watch?v=YUtNAYvxfRI

[00:08] So, welcome back everybody for another webinar organized by Princeton for everyone worldwide.
[00:11] We're very happy to have Richard Thaler with us and Alex Imas.
[00:16] Hi Richard and hi Alex.
[00:18] Hi Marcus.
[00:20] Today we talk about the new book Behavioral Economics Anomalies Then and Now and here's the cover of the book.
[00:28] It's a second edition.
[00:29] I think there's an earlier edition on the winner's curse which I read several years ago or many years ago now.
[00:35] I guess the new part of it is then and now you can test all the anomalies to what extent they're still holding after more than a decade later.
[00:46] And uh let me just before we I pass the phone on to to Richard and to Alex, let me just classify some of the uh behavioral economics literature.
[00:57] In a sense there's two big strands.
[01:00] One is biases.
[01:03] Bias can be in preferences.
[01:03] There's some reference dependencies, loss aversion, some anchoring, status quo bias, hyperbolic discounting and so
[01:08] forth.
[01:09] And they can instead of preferences they can also be in the beliefs.
[01:13] You might have probability weighting, I need to ask Kahneman, overconfidence, confirmation bias, gambler's and hot hand fallacy and so forth.
[01:20] So, the first traditionally I think there was a lot of emphasis on biases.
[01:25] And more recently there's a lot of emphasis on noise.
[01:26] It's like a cognitive origin where the biases might be coming from.
[01:32] And people might be aware of this noise on their own.
[01:35] And it's often modeled with a drift diffusion models.
[01:38] So, you learn the drift of these models.
[01:41] It's like an um information acquisition uh problem that you everything is noisy and get the information.
[01:47] Richard of course did also important work and is famous for uh the world of nudging in terms of paternalism.
[01:53] And uh you can do you know overcoming biases where you have some biases in preferences or beliefs.
[02:01] And you know how just exposing the agents or the citizens to the default option makes a big difference.
[02:07] And I still remember
[02:09] very vividly when he first time presented save more tomorrow.
[02:13] uh which you know you just sign up for something tomorrow and you can overcome some biases this way.
[02:21] If you want to overcome some bias some noise distortions, uh what you do you nudge essentially people to get certain expose them to certain information to overcome some of these uh noise components.
[02:32] You learn at a different speed on a different direction.
[02:36] So, with this I think we will today we will learn more about uh the anomalies which I think Richard and Alex have studied all their life and uh and I think they will show us today how well they hold up over the decades.
[02:53] So, we're looking forward to it and Richard I think you're starting and then Alex will take over subsequently.
[03:00] So, yeah one one small correction.
[03:02] The original version of this book came out in 1992.
[03:07] So, um
[03:09] and it was based on a series of columns I wrote in the then new Journal of Economic Perspectives.
[03:20] So, um yeah the book was published in '92 and Alex and I started working on this new version at least 3 years ago.
[03:30] So, it it's like looking back 30 years and saying uh you know, should we throw everything away or did we actually know something?
[03:42] Okay.
[03:46] So, uh this is you know Nobel week as our president seems to be keenly aware of.
[03:53] And uh let me cite uh two of my heroes who have that award.
[03:59] One is Danny Kahneman.
[04:02] Uh you know, in his book uh Thinking Fast and Slow.
[04:07] It I'm not sure he ever said this but
[04:09] I'm putting words in his mouth.
[04:11] Uh maybe before you do something rash
[04:13] check in with system two
[04:16] uh which is the slow part.
[04:18] But uh Bob Dylan has a famous song Don't
[04:22] Think Twice It's All Right.
[04:24] So, who's right?
[04:27] And um my message to economists is always uh
[04:31] listen to Danny.
[04:33] And um
[04:35] I think we need system two and economists need some system two.
[04:41] So, what do I mean by thinking twice?
[04:46] We've made a lot of progress.
[04:49] It used to be like when I first started causing trouble in the '80s, economists didn't think there was any other way to do things.
[04:59] And as Marcus alluded to earlier, there's now a large literature on a large number of topics like loss aversion, fairness, salience, hyperbolic
[05:12] discounting.
[05:16] These topics now appear in top journals.
[05:20] It It It used to be you'd have to justify why are you introducing this?
[05:23] Now you can just do it and nobody really complains.
[05:28] On the other hand, uh textbooks have barely changed.
[05:30] If you open up a standard principles book, it looks pretty much like the ones that um Marcus would have read when he was a student or even I when I read when I was a student.
[05:51] So, here are some I'm going to give a few examples of things I need I think economists need to stop and think twice about.
[06:02] And the first, this goes back to the first paper I wrote that is a behavioral economics paper uh published in 1980.
[06:14] bef- before Alex was born.
[06:21] it's think twice about the model you're writing down is a normative model or a descriptive model.
[06:30] And I think the big big problem economists have is they've never clearly distinguished between these.
[06:37] So, take expected utility theory.
[06:42] That was explicitly a normative model.
[06:47] And we had John von Neumann, smartest man on Earth or contender for that title, saying, "Look, here's a set of axioms that make sense to him and you want to behave according to those and if you do, this is what you've got to do and you have to maximize expected utility."
[07:07] That's a normative model but economists just also use it as their workhorse
[07:14] model of how people make decisions.
[07:18] So, what what's an example of the difference?
[07:22] Take Google Maps.
[07:25] That that is a model of the fastest way to get from one place to another.
[07:33] Um and that's a normative model but we might also be interested in well, what do actually people take?
[07:41] And let's show the next slide.
[07:45] That's a descriptive model.
[07:48] Right? So, that's what people do and we might be interested in that.
[07:54] Maybe we want to put the sidewalk there or a hedge preventing people from doing that.
[07:59] So, but I think So, that's my first point is that when we're writing down models, let's think about what we're trying to do.
[08:10] Are we trying to say what's the smart way to do something or what is the way that people do it?
[08:15] Okay, next slide.
[08:18] Now, a related point is think twice before you write down max.
[08:29] Maximize.
[08:30] So, I think economics really is distinguished from other social sciences because virtually all models start by agents, we don't even call them people, agents optimizing.
[08:50] That and that and the study of markets I think is really what distinguishes economics from sociology and psychology.
[08:59] There's no underlying optimization going on.
[09:03] Um but you know, we assume all agents maximize.
[09:10] And uh behavioral models might be called max but maybe you know,
[09:18] um and so how seriously should we take the maximization part and what's the alternative?
[09:27] And what one first question people should ask is,
[09:32] "Well, how hard is the problem?
[09:34] Are we studying tic-tac-toe where a bright first grader can solve it or chess where no human has yet solved it?"
[09:47] Obviously maximization is a good model of tic-tac-toe and a lousy model of chess unless you've got the an app open in front of you.
[10:00] Um but if we're going to go down that path, then we need a some way of measuring difficulty.
[10:06] So, we know chess is harder than tic-tac-toe, but there are a lot of things we don't really know how to measure difficulty.
[10:15] So, take the life cycle savings model,
[10:19] which is still the workhorse model of how people allocate their income over their lifetime.
[10:28] That's a really complicated problem.
[10:31] The only person I know who might have solved that problem for himself is Doug Diamond,
[10:38] who is one of the smartest people I know and is completely fascinated with these kinds of problems.
[10:48] So, you know, if I have a finance question, I go ask Doug.
[10:50] But for people who don't have access to Doug, it's completely preposterous that they're doing anything like figuring out the present value of their future income and simulating all the investment possibilities and then figuring out their life expectancy and updating that all the time.
[11:13] I once presented a talk at the Cornell psychology department and described this model uh
[11:19] And the audience just started laughing.
[11:23] And thankfully, my friend Bob Frank was there and could assure them I wasn't making this up.
[11:28] It's not a caricature.
[11:31] It's a real model.
[11:34] Modigliani won the Nobel Prize for this.
[11:36] That's our model and it's too hard.
[11:38] Nobody could do that.
[11:39] So, uh how You could say that we have to have simpler models.
[11:45] So, we can't go with this complicated models.
[11:46] Let's try to represent reality with a simple model.
[11:48] And then people can solve it within the simple model.
[11:52] Well, I don't know I I know I think I disagree with that because well, we don't know what the simple model is.
[12:00] So, what what you're saying is, okay, what do people do?
[12:01] Well, what what the Keynesian model is spend your income.
[12:09] Now, if you ask me, if I had Suppose I have to choose between a hyper-rational uh Robert Barro kind of life cycle model
[12:21] where I'm not just doing that for myself, but for my heirs and their heirs.
[12:27] or C equals Y, I'll take C equals Y.
[12:31] But I it's not that I think spend your income is the right model, but it's closer.
[12:38] And so, my main point here is people are not capable of maximizing.
[12:43] We need to adjust for difficulty and we have no idea how to do it.
[12:48] One thing I'm sure of is that how hard a problem is for a computer is not a good measure of how hard it is for humans.
[13:00] That's completely different.
[13:05] so, what would an alternative model be?
[13:07] So, Herb Simon, who was a behavioral economist before the term existed, like to talk about satisficing models, which is you work you think about it and
[13:22] you choose something.
[13:25] And when you have something when you have an al uh a decision that feels good enough, you quit.
[13:31] He calls that satisficing.
[13:34] Uh but I don't think we have good models of satisficing.
[13:39] And uh therefore, we don't really know how satisficing models differ from maximization models.
[13:47] And I think uh at the end of this book, we talk about some of these things and I'm talking about this now because we're not going to get to it, but I think that is the frontier right now or one of the frontiers, which is if maximizing models are clearly wrong, well, then what?
[14:11] And I don't think we have a then what yet.
[14:15] What is what is this meh? m e h
[14:19] That Okay, that I'm going to get answer it and then you
[14:22] can play the the laugh track, Marcus.
[14:28] Okay.
[14:28] So, um meh is uh you know, uh it's not bright, it's not wrong, it's kind of meh.
[14:36] And that's the way, you know, if if I'm deciding how much time to spend trying to understand some paper I'm reading, I'm not maximizing, I'm doing something and that's that's meh.
[14:54] So, uh we don't have models of meh.
[14:58] I see.
[15:01] So, you don't criticize rational expectations.
[15:02] That's maximizing models that happens in models with deterministic.
[15:08] Um not on this slide.
[15:09] I mean, I with rational expectations, I knew you were going to talk about that.
[15:14] Okay.
[15:16] before I started.
[15:18] So, um and in a longer version of this talk, I talk about expectations.
[15:24] I I think to a.
[15:27] Look, you know, the Fed is preoccupied by expectations.
[15:32] Yes.
[15:34] I I ask everybody I know in that area,
[15:38] "Whose expectations are you talking about?"
[15:42] If you walk down the street in Princeton or on Michigan Avenue and you ask people what their inflation expectations are, they have no idea what you're talking about.
[15:56] They can answer the question whether they think prices are high or low and whether they have gone up and they'll all say yes, but most of them won't know the difference between a rate and a level, uh much less what it means to have an expectation.
[16:16] Yeah, they think prices are going to go up because prices go up.
[16:21] Um uh so,
[16:24] you know, I think bond traders have expectations,
[16:30] but who else?
[16:34] I don't know and I don't even know whether we have a good framework for saying what loosey-goosey expectations are and whether they meet what we would call an expectation.
[16:53] But would you agree that it's a useful benchmark?
[16:55] It might not represent reality, but it's one benchmark and then we can describe the deviations from that.
[17:03] As part
[17:03] Yes, absolutely.
[17:07] Look, I think expected utility theory is a benchmark.
[17:10] Yes.
[17:10] Efficient market hypothesis.
[17:14] You know, uh Gene Fama always teases me that behavioral finance would be nothing without him.
[17:24] Uh um he takes com complete credit
[17:29] for the existence of behavioral finance.
[17:31] and none of the blame.
[17:33] I see. uh for the mistakes we've made,
[17:37] but I'll give him that.
[17:39] You got to give him something, right?
[17:41] So, I give him that and it's like prospect theory couldn't exist without expected utility theory.
[17:48] And uh if we didn't have the efficient market hypothesis, we wouldn't have a null hypothesis.
[17:56] So, yes, benchmark models are essential, but then if we claim to be behavioral economists, then what are what are we doing?
[18:06] And writing down a model, the rational expectations model, which is assuming that the market expectations are the same as those of the best econometrician, I think is preposterous.
[18:24] Yes.
[18:27] There's a question coming in from Bill.
[18:29] He would like to know whether an AI assistant will overcome all of these
[18:31] behavioral limitations.
[18:33] once you can consult your private AI agent to help you do the life cycle smoothing of your consumption.
[18:40] Um.
[18:41] Alex.
[18:44] Um.
[18:46] I think.
[18:47] All hard questions go to the younger team member.
[18:53] I mean, I think I think it's really hard to to answer that question without knowing what that AI agent is.
[18:56] If it's just an LLM that you got, you know, from just log in and start do using chat GPT, uh the answer's no.
[19:07] Um the LLM is trained on human data, which is biased.
[19:10] So, it's going to give you a biased answer.
[19:13] Um if you prompt it and say, "Give me the unbiased answer," it'll give you some approximation that's closer to some benchmark, which again, you you don't know if it's correct or not.
[19:24] There's the the black box nature of these agents uh makes it very, very difficult to to to answer this question in a in a concrete
[19:31] way.
[19:33] Even with the open weights models, the uh it's just so complicated uh like how the neural network is actually spitting out the output that it is.
[19:42] Uh, there's just no way to verify that that the agent is actually going to be, you know, solving the dynamic optimization problem in a way that we think it should be solving the problem.
[19:53] Um, you could say like, "Hey, is it is are the forecasts obeying fire?" or something like that.
[20:00] And mo- what you probably realize is they don't.
[20:02] Um, you know, I've tested this in in in in my own research.
[20:06] Uh, a lot of the time they're they're even kind of like asking them to do something very simple like maximizing expected utility.
[20:11] Unless you prompt it exactly to do the calculation, here are the numbers, here's what you have to do, they're going to give you something different.
[20:19] And the problem is unlike with humans or unlike with a mathematical benchmark, um, you have no idea where that answer's coming from.
[20:27] So, I I'm not particularly optimistic about this.
[20:30] And even going
[20:33] into the future
[20:35] Yeah, let me make a
[20:36] uh
[20:37] a a different point, which is
[20:40] there are results that are 50 years old
[20:44] that simple linear models do better than
[20:48] experts at virtually everything.
[20:52] So,
[20:53] deciding
[20:55] which applicants to a graduate program
[20:58] are going to be best to predicting
[21:02] which
[21:04] uh tumor is likely to be benign or
[21:08] malignant.
[21:10] So,
[21:11] and
[21:12] that's a 50-year-old result that's still
[21:15] true.
[21:16] And in almost all the cases that have
[21:18] been studied,
[21:21] uh, we haven't switched over to models.
[21:24] Um, so, we didn't need AI to get better
[21:28] predictions.
[21:30] Linear regressions
[21:32] make better predictions than humans.
[21:35] And of course, machine learning can
[21:37] improve,
[21:39] but will we use them?
[21:43] Uh, that's a
[21:45] another question. So, Alex is answering
[21:48] the question of whether that model will
[21:50] be
[21:51] as good as our
[21:53] uh assumed
[21:56] hyperrational model. And there's a
[21:58] second question, which is will we bother
[22:00] to use it? And I would say,
[22:04] yes, hedge funds are going to use it.
[22:08] Um, but and and we use it
[22:12] when we're driving, the GPS is
[22:15] is optimizing
[22:17] and telling us what the traffic is. But,
[22:21] uh I've yet to
[22:23] be at a university that used models to
[22:27] predict students.
[22:28] But, can you imagine in the future we
[22:30] outsource our decision-making to 90% to
[22:33] some AI agents and we don't really make
[22:36] so many decisions anymore? We Yes, I can
[22:41] I do I let the GPS
[22:44] decide the best way for me to get home.
[22:48] Yeah.
[22:48] Um, but I don't have it decide
[22:53] uh uh
[22:54] what movie I should watch. Now, But,
[22:58] some people might even let decide whom
[23:00] whom to marry in a sense. Uh, yes. And
[23:04] for good or for bad, I mean, it's hard
[23:07] to imagine that AI would be worse.
[23:12] Right? I mean, marriage
[23:14] uh
[23:16] It
[23:17] uh we seem to be
[23:18] about 50%.
[23:21] Yes.
[23:22] Uh on choosing spouses optimally.
[23:26] Um, I don't you know, it might have been
[23:29] that the old model
[23:32] of having your parents choose a spouse
[23:34] was better. I see no evidence that it
[23:37] was worse.
[23:40] But, summarizing, could it be that in 10
[23:43] years behavioral economics dies because
[23:45] all the decisions are made by some AI
[23:48] components? Or do you think so?
[23:49] >> I just I just don't see that. Mhm. I
[23:52] mean, again, it's hard it's it's hard to
[23:54] make predictions, especially about the
[23:56] future.
[23:57] Uh as a as a smart man once said,
[24:00] uh but
[24:02] there needs to be a kind of a revolution
[24:04] in the architecture of the AI that's not
[24:07] based on what we know now as in the LLM.
[24:10] Um, you know, I have work with Sanjog
[24:12] Misra, who's a who and Kevin Lee, who
[24:13] are colleagues at Chicago, showing that
[24:15] even with AI agents, you know, AI agents
[24:17] are supposed to be
[24:19] uh these kind of hyperrational things
[24:21] that you know, you you just prompt it,
[24:22] they go out there, they negotiate for
[24:24] you. Everything's going to now be
[24:25] homogeneous. Everybody's going to be
[24:27] getting similar outcomes based on you
[24:29] know, the preferences that you input
[24:31] into the agent. Turns out that these
[24:33] agents, because they're built on text
[24:35] data and all
[24:36] anytime you're going to use the
[24:38] architecture of a LLM, that's what it's
[24:40] going to use. Um,
[24:42] regardless of how you scale it. Um,
[24:44] you know, what we find is that basically
[24:46] you get the same sort of heterogeneity
[24:47] as you get with human negotiators,
[24:49] because the prompt is literally kind of
[24:51] injecting the preferences and biases of
[24:54] the humans into the agents. So, but now
[24:56] again, you have this black box nature
[24:58] that the heterogeneity looks kind of
[25:00] alien. You don't know where it's coming
[25:01] from. So, you're not getting to like
[25:02] this hyperrational
[25:04] uh result from the standard model where
[25:06] like, look, we're going to get just get,
[25:08] you know, heterogeneity based on
[25:10] explicit preferences that are fed in
[25:11] through the prompts. You're just kind of
[25:13] getting alien heterogeneity that you
[25:14] can't really explain where people are
[25:16] getting outcomes that they might not
[25:17] really be happy with, because they don't
[25:19] have no idea why that the agent
[25:21] negotiated the outcome that they did.
[25:23] And you know, these are this is the top
[25:24] flight models that we're using. And
[25:28] um, I think this is going to be a real
[25:31] uh impediment for the current
[25:33] architecture as far as getting anywhere
[25:35] close to
[25:36] uh being these universal decision aids
[25:38] where people kind of sit back and
[25:41] uh let them decide what to watch on on
[25:43] Netflix, let them, you know, pick their
[25:45] partners and things like that.
[25:47] So, just to summarize,
[25:49] these are two great questions and and we
[25:52] don't know the answers because partly we
[25:55] don't know whether those agents
[25:57] will meet those models and be we don't
[26:00] know whether humans will choose to use
[26:02] them.
[26:04] And
[26:05] we'll see.
[26:07] Good questions.
[26:09] Uh, if we ask who has good expectations,
[26:13] weather forecasters, perfect.
[26:17] Uh, CFOs of US firms
[26:21] were are asked
[26:23] to predict the return on the S&P 500 in
[26:27] 80% confidence limits.
[26:30] And how do they do?
[26:33] So, this is the percent of their
[26:35] forecasts that lie within the 80 80%
[26:41] limits.
[26:42] And they should all be up at that dotted
[26:44] line at 80%. And you can see
[26:48] they managed to get up there four times
[26:52] in the sample. And in many cases,
[26:56] the uh none of the
[26:59] the this is the percent of the
[27:01] forecasters that had a wide enough
[27:03] limit. And in some years, it's none of
[27:06] them.
[27:07] And uh this is like
[27:10] during the financial crisis. What? Uh,
[27:13] for years, the worst-case scenario was
[27:17] zero.
[27:20] And uh obviously, we've had periods
[27:22] where the return on the stock market was
[27:25] below zero.
[27:26] So, next slide.
[27:30] Uh, all right. Now, this is one where
[27:34] economists don't even really think
[27:35] there's anything to think about. So,
[27:39] we write down uh some utility function,
[27:43] and we say,
[27:45] "Do people maximize that?"
[27:48] And and people think that Well, that's a
[27:50] tautology.
[27:52] And if if you believe revealed
[27:54] preference, if that's kind of your
[27:56] approach, then
[27:58] um
[27:59] we choose whatever it is that maximizes
[28:03] our utility. How could you disprove it?
[28:06] Well, there's another agent literature
[28:11] documenting something called the
[28:12] preference reversal phenomenon,
[28:15] where it's very easy
[28:18] to get people to say that they prefer A
[28:21] to B and B to A.
[28:25] And um
[28:27] the you know, this work was done in the
[28:29] '70s, and then
[28:32] uh David Grether and Charles Plott tried
[28:34] to disprove it
[28:36] and failed and published that in the AER
[28:41] in around 1980.
[28:44] But, that hasn't stopped economists from
[28:47] studying utility maximization problems
[28:51] and assuming that
[28:54] that they're correct.
[28:56] And um
[28:58] you know, a
[29:00] a question that uh
[29:02] Danny Kahneman asked, and we have a
[29:05] piece Danny I have a piece in this book,
[29:09] which is
[29:10] do
[29:11] um
[29:12] are people even capable of maximizing
[29:15] utility?
[29:16] And
[29:17] in order for them to be capable, they
[29:21] have to be able to make forecasts.
[29:24] So, if I'm going to choose the
[29:28] dinner option
[29:30] that will make me happiest, I have to
[29:32] make a good forecast.
[29:35] And
[29:36] there's lots of data suggesting
[29:38] that we're not good at that.
[29:41] And so, even I would say the
[29:46] the idea that people maximize utility,
[29:48] which seems like it doesn't require
[29:51] thinking twice, I would say it does.
[29:55] So, but what's your alternative? So, if
[29:56] we have these inconsistencies that
[29:58] people prefer A to B and B to A,
[30:01] um what do we do
[30:04] instead, in a sense?
[30:06] Um
[30:09] Wha- Well, that
[30:11] that's a good question.
[30:13] And
[30:16] I think
[30:17] that is the entire behavioral economics
[30:19] agenda.
[30:21] Okay. Is to say, what do people do?
[30:25] And and
[30:27] and you know, that my little theme of
[30:29] thinking twice is
[30:32] even
[30:34] e- even on questions that
[30:38] economists would think, well,
[30:40] come on.
[30:41] We can't
[30:43] It It can't be problematic
[30:46] to assume that people maximize utility.
[30:48] Isn't that what utility means?
[30:52] And you know, we we can disprove that
[30:56] quite quickly.
[30:58] The next slide is a good joke, and then
[30:59] I'm going to turn it over to Alex.
[31:01] In case you don't speak Spanish, Marcus.
[31:06] So, this is a violation of the law of
[31:08] one price.
[31:10] Yes.
[31:11] Okay, I'm going to turn the reins over
[31:14] to
[31:14] >> but what is the explanation for
[31:15] non-Spanish speakers or something?
[31:17] >> Exactly.
[31:19] Exactly.
[31:20] That includes Germans, Marcus.
[31:25] What should I do?
[31:26] I'm happy to pay $5.
[31:28] Choose the cheaper one.
[31:30] Probably even better.
[31:36] So, okay. Uh
[31:39] Skip over to whatever
[31:42] uh Oh, well, yeah, I guess this is
[31:44] another joke.
[31:45] Um
[31:47] on the law of one price.
[31:50] The
[31:51] There's a closed-end fund called the
[31:53] Cuba Fund. That's
[31:55] a ticker symbol CUBA. Has anything to do
[31:59] with the country Cuba?
[32:01] No.
[32:03] Okay. It It It's invests in the
[32:05] Caribbean.
[32:07] Okay.
[32:08] But it has a ticker symbol CUBA. Mhm.
[32:12] And it normally sold at a discount.
[32:18] Until there was this one day
[32:21] when it shot to a 70% premium.
[32:26] Now, remember, it has nothing to do with
[32:28] Cuba. Mhm. And there are no securities
[32:31] in Cuba.
[32:32] And if there were, it would be illegal
[32:34] to invest in them.
[32:37] But any guesses as to what happened on
[32:40] December 18th of whatever year that is?
[32:44] That's the day that Obama announced his
[32:47] intention to relax relations
[32:51] with a country that happens to be
[32:52] spelled the same way I see. as this
[32:57] s-
[32:58] closed-end fund. So, I'm done with my
[33:00] jokes. Keep skipping until Alex starts
[33:03] talking.
[33:05] Okay, um so, that's a that's a good
[33:09] segue
[33:10] uh to kind of uh talk about what we what
[33:14] we basically did with the book as far
[33:16] as, you know, there was these this
[33:17] original Winner's Curse in in published
[33:19] in 1992, which basically took these
[33:21] anomaly columns that Richard was writing
[33:23] with his co-authors and put them
[33:25] together in a volume. And it's been o-
[33:28] almost three decades since that
[33:29] publication. So, what we wanted to do
[33:32] with the book was to say, okay, so it's
[33:33] been three decades, what have we
[33:34] learned? Can we look back at the field
[33:37] and see where it is today? And one of
[33:39] the things that we add to each uh kind
[33:41] of an uh topic, which is kind of
[33:43] anchored around this anomalies column,
[33:45] um is to show that, look, a lot of the
[33:48] research that was done in the past as a
[33:51] bedrock for behavioral economics, a lot
[33:53] of it took place in in the lab,
[33:55] essentially. So, student samples, low
[33:58] stakes, sometimes even hypothetical. So,
[33:59] if you read the original prospect theory
[34:01] paper, all of the experiments were
[34:02] hypothetical choice.
[34:04] And a lot of the pushback by economists
[34:06] at the time, including in '92, was,
[34:08] look,
[34:10] we don't really care about what students
[34:11] in the lab do at these low stakes. Smart
[34:13] people are the are the ones who survive
[34:16] market conditions, and these smart
[34:18] people are just not going to display the
[34:20] biases that you're talking about and
[34:21] telling us that we need to be worried
[34:23] about. So, we're just going to say we're
[34:24] not worried about it.
[34:26] So, the part of the what Richard and I
[34:28] argue is that part of the reason
[34:30] behavioral economics has become a
[34:32] successful subfield of economics, uh is
[34:34] that the last 30 years or 25 years have
[34:37] shown that a lot of these anomalies show
[34:40] up in real-world settings amongst
[34:42] exactly the type of people who
[34:44] economists should say they should These
[34:47] are specifically the people who
[34:48] shouldn't have these anomalies. So, the
[34:50] last three decades have shown that the
[34:51] anomalies appear amongst professional
[34:53] investors, amongst professional
[34:55] athletes. Richard showed a graph on
[34:57] CEOs.
[34:59] So, you think the field experiments were
[35:01] the key new element. What's about the
[35:04] neuroeconomics
[35:06] finding neuro correlates? Do you think
[35:07] it was also important to make the field
[35:09] mature?
[35:10] Um so,
[35:12] it depends on what you say about
[35:14] neuroeconomics. So, I think uh there was
[35:16] a lot of excitement in the mid-2000s, I
[35:19] think 2005, 2006 about neuroeconomics.
[35:22] And uh about kind of going to that uh to
[35:26] the point where you can use neural data
[35:27] to do like that uh what Edgeworth was
[35:29] talking about, this hedonanometer, where
[35:32] like you can actually measure utility in
[35:33] the brain. Um I'd say I I was actually
[35:36] at the Society of Neuroeconomics
[35:38] conference this past weekend. It's a
[35:39] It's a It's still a huge organization.
[35:42] Um there were hundreds of people in the
[35:43] room. Um I'd say that the idea of using
[35:48] non-choice data to to think about
[35:51] economic decision-making, that idea is
[35:53] alive and well, and actually, I'll talk
[35:55] about this at the very end, that is kind
[35:58] of where the for- the the frontier of
[36:00] behavioral economics is now, is to think
[36:02] about the cognitive foundations for the
[36:04] types of things that we're going to be
[36:05] that that make up these behavioral
[36:07] economic anomalies. So, what is loss
[36:09] aversion? Is it really just a parameter
[36:11] we're adding to the utility function, or
[36:13] is this some sort of result of people's
[36:16] natural cognitive constraints being
[36:18] limited? So, for example, the fact that
[36:20] we don't have infinite attention, that
[36:22] we don't have infinite memory.
[36:24] If we put these constraints on people's
[36:26] attention, memory, and all of these
[36:28] other resources, can we get what looks
[36:30] like loss aversion? Can we get what
[36:32] looks like narrow bracketing? And I
[36:34] think the most the exciting frontier
[36:36] research in behavioral economics is
[36:37] showing that actually, yes, that's the
[36:39] case, that we can think about almost um
[36:43] you know, thinking about behavioral
[36:44] economics through the lens of
[36:46] constrained maximization in some ways.
[36:48] But it in a different way than the sort
[36:50] of you max utility that would that that
[36:52] we've been thinking about before.
[36:54] >> And I thought this I don't I can do
[36:55] still this with models. I don't need
[36:57] neuro correlates or certain parts of the
[36:59] brains are active or not active. Is this
[37:01] correct? Yeah, so what what I meant by
[37:03] in the beginning is that the
[37:05] we're still using neural correlates, but
[37:07] we're not necessarily using fMRI data.
[37:10] Mhm.
[37:11] So, people are still using fMRI data,
[37:12] but that data has become pretty
[37:14] controversial even in the neuroscience
[37:15] community. Because just because
[37:17] something lights up, it turns out that
[37:19] that's not actually where it's
[37:20] happening,
[37:21] necessarily. W- W- One other thought
[37:23] about Mar- that, Marcus, we
[37:25] you know, Colin Camerer has a been a
[37:28] friend of mine for 40-some years.
[37:31] My challenge to him
[37:34] has always been,
[37:35] show me either
[37:38] uh
[37:40] something you've discovered that I
[37:41] didn't think I already knew,
[37:44] like that people are loss averse,
[37:47] or show me something I thought I knew is
[37:50] wrong.
[37:52] And
[37:54] I don't think he's been able to
[37:57] do either of those.
[37:59] But he's giving a talk in Chicago on
[38:01] Monday, maybe I'll have to revise.
[38:06] Okay, good. Let's go back to behavioral
[38:07] finance.
[38:09] Yeah, and I just kind of wanted to end.
[38:10] I think the the on that note, I think
[38:12] the the what that line of research has
[38:14] been doing is is allowing us to put more
[38:17] structure on something that behavioral
[38:18] economists and economists have been
[38:20] ignoring for a long time, which is
[38:21] context effects.
[38:23] Um behavioral economists have kind of
[38:25] like not really
[38:26] tried to touch those formally for a long
[38:28] time.
[38:29] But what the empirical evidence shows is
[38:31] that they're actually super important.
[38:34] Right, you can get lo- If you look at
[38:35] data on even loss aversion, that data
[38:38] there's a lot of heterogeneity on that
[38:39] data. So, putting structure on when you
[38:42] will see some level of loss aversion
[38:43] versus a different level of loss
[38:45] aversion. I mean, the we're talking
[38:46] about magnitudes. Putting structure on
[38:48] that is where I think the meat of this
[38:50] of this research program is going to be.
[38:53] Um So, going back to behavioral finance,
[38:55] behavioral finance is one of the most
[38:57] successful subfields of behavioral econ.
[39:00] Um and I think the reason for that is
[39:01] the fact that these are this is an
[39:03] environment where people are uh you
[39:06] know, spending time on these decisions,
[39:08] getting feedback on these decisions.
[39:10] This is real money on the line, often
[39:12] thousands of dollars.
[39:13] And a lot we've seen a lot of kind of
[39:16] behavioral biases being documented,
[39:17] sometimes even in the field before it
[39:19] gets documented in the lab, such as in
[39:21] the case of the disposition effect,
[39:23] uh
[39:24] and overconfidence, and things like
[39:26] that.
[39:28] But the pushback against behavioral
[39:30] finance has uh which started in with
[39:33] Richard's paper in '85, and then uh
[39:36] Terry Odean's papers in in the late
[39:38] '90s, has also been kind of similar to
[39:40] the subjects in the lab. Look, a lot of
[39:42] the data comes from retail traders who
[39:43] might not know what's going on. They're
[39:45] a tiny portion of the market. Um it's
[39:48] getting larger with Robinhood, but
[39:49] they're still really small. They're not
[39:51] They don't have a lot of money. So,
[39:53] there's been a dearth of studies on
[39:54] smart money, and as a result,
[39:58] finance uh academics have still been
[40:00] able to kind of push back on this
[40:02] research saying like, "Look, you have
[40:03] incentives, you have uh you have
[40:05] feedback, but this is still not the
[40:07] people who we're we're really concerned
[40:09] with."
[40:11] Um so, I wanted to give an example of
[40:12] kind of where where some of the some of
[40:14] the research is is today in behavioral
[40:16] economics. So, I'm going to use my paper
[40:18] as an example uh
[40:19] uh just because I I know a lot about it.
[40:22] Uh it was published in the JF in in
[40:24] 2024, and I think um
[40:27] where this paper kind of
[40:29] uh where this paper's kind of placed in
[40:31] behavioral economics is that we're
[40:33] actually able to look at financial
[40:35] market experts. So, we're actually able
[40:37] to get data on daily trade decisions,
[40:40] looking at the entire portfolio of these
[40:42] financial uh experts on a daily basis.
[40:45] So, I have their entire access to the
[40:46] entire portfolio over a 13-year period.
[40:49] I know exactly what they can trade, and
[40:51] I know exactly what they do trade every
[40:54] single day.
[40:55] Um and because of this access to data,
[40:58] we can actually say, "Look, are there
[41:01] markers of rationality in these
[41:03] decisions, or are there markers of the
[41:05] same sort of behavioral economic biases
[41:07] that we've been documenting amongst less
[41:09] sophisticated populations?"
[41:12] And the question that we ask in this
[41:14] paper is a very basic one, which is that
[41:16] our buying and selling decisions
[41:19] They're They're pirate part two sides of
[41:20] the same coin in standard economic
[41:22] models. Do they look similar in uh in
[41:25] this population?
[41:28] So, here's like a summary slide of the
[41:30] type of people we have in the data set.
[41:31] We have about eight 800 uh folks over a
[41:34] 13-year period. They have very large
[41:36] portfolios. The Odean data set Terry
[41:38] Odean's data set had about um
[41:41] you know, the average market portfolio
[41:43] was about $10,000. In our case, it's
[41:45] $573 million.
[41:47] Uh they hold a lot more assets. Uh they
[41:50] have longer holding periods, and they
[41:52] they and because we have such a long
[41:54] span, we have a lot of we have a lot of
[41:56] trade data.
[41:57] Uh next slide.
[41:58] Um and just to to give you kind of
[42:00] background, uh Matt Levine uh who may
[42:03] some of you may might be reading covered
[42:05] this paper, um and he basically said
[42:08] that, you know,
[42:09] we In our economic we think these should
[42:11] be the same, but actually, if you kind
[42:13] of talk to traders, there's some hints
[42:15] that these might be different. So, he
[42:17] said, "If you hang around financial
[42:18] markets, you'll get the sort of advice
[42:20] about how to decide what stocks to buy."
[42:22] And almost all of this comes down to
[42:23] some sort of fundamental analysis.
[42:25] You're actually doing research about
[42:26] what to buy.
[42:28] You also get some advice about what how
[42:30] to sell, but this advice tends to be
[42:32] heuristic. So, something you hear is cut
[42:34] your losses and let your winners run.
[42:36] People say that. If you double your
[42:38] money, uh sell and take the profit.
[42:40] That's what Barron said
[42:42] uh in their handbook. The basic folksy
[42:44] rule of wisdom for buying is about
[42:45] fundamentals, but for selling, it's all
[42:47] about price action.
[42:49] Is it called the disposition effect?
[42:52] Uh not quite.
[42:54] Not quite. I'll show you what that looks
[42:55] like. Um
[42:57] Uh half of it looks like the disposition
[42:59] effect.
[43:00] So, the question is So, we looked at
[43:02] buying and selling, and we we've we
[43:05] because we have access to their entire
[43:06] portfolios and we have 13 years of data,
[43:09] we can actually construct
[43:10] counterfactuals for for their buying and
[43:12] selling
[43:13] uh performance. So, on buying, we can
[43:15] say, "Look, I can say what they see what
[43:17] they bought. We could I can also see
[43:20] what they could have bought. Let me buy
[43:22] something that is matched on all of the
[43:24] characteristics that we can think of in
[43:27] terms of what they actually bought, and
[43:28] say, "Is there some sort of signal that
[43:31] they're acting on in terms of
[43:33] counterfactual returns?" And what we see
[43:35] is that there seems to be some sort of
[43:37] skill in buying. They're doing better
[43:38] than this random counterfactual that
[43:40] we're
[43:41] We We threw the boat at it. We have a
[43:42] lot of different counterfactuals in the
[43:44] data. All of it points to the same
[43:46] picture that it seems like they're
[43:48] they're they actually know what they're
[43:49] doing for buying.
[43:51] Um next slide.
[43:53] Um but in order to buy, because these
[43:55] are long-only portfolios, the portfolio
[43:58] managers actually need to sell.
[44:00] Uh it's unfair to compare the the
[44:02] counterfactual to the benchmark because
[44:03] they can't necessarily buy or sell the
[44:05] entire benchmark that the client
[44:07] provides for them. So, what we do is we
[44:10] use a very conservative counterfactual,
[44:12] which is I see you sell a dollars worth
[44:15] of shares. I'm going to I know you need
[44:16] to sell this dollars worth of shares.
[44:18] What I'm going to do is I'm going to
[44:19] throw a dart at your portfolio and
[44:21] randomly sell a dollars worth of
[44:23] something else. And we have versions of
[44:25] this where we match to risk
[44:26] characteristics,
[44:27] matched industry, all of these sorts of
[44:30] things. So, we call this a random
[44:31] selling strategy, and we see, "Do you do
[44:33] better than this random selling
[44:35] strategy?" Uh next slide.
[44:39] And the answer is no.
[44:40] They're horrible at selling.
[44:43] Uh
[44:45] Yep.
[44:46] So, are you saying that I'm just selling
[44:47] to free up some funds in order to really
[44:49] buy what I want to buy?
[44:51] Yeah, they're they're they're selling
[44:53] funds to That's exactly what they're
[44:54] doing. Uh but even if they were selling
[44:57] randomly, they should be That should be
[45:00] at zero, essentially.
[45:02] So, they're somehow doing worse than
[45:04] random.
[45:05] So, what are they doing?
[45:07] Uh next slide.
[45:09] Uh the natural
[45:11] uh natural candidate for thinking about
[45:12] what they're doing is to think about the
[45:13] behavioral economics literature on
[45:15] limited attention. That people can't pay
[45:17] attention to their to
[45:19] all of the things that they're that
[45:20] they're paying that they need to be
[45:21] paying attention to. These guys
[45:23] technically can because they can slow
[45:25] down, they have resources for analysts
[45:27] and things like that. But what we're
[45:28] arguing is that that is exactly what
[45:30] they're doing for their buying
[45:31] decisions, but for their selling
[45:33] decisions, they're using some limited
[45:35] attention heuristic. So, how do we find
[45:38] evidence for limited attention in the
[45:39] data set? Well, we can use uh
[45:42] uh
[45:43] So, here's the slide. We can use proxies
[45:45] for limited attention, namely, are they
[45:47] buying or selling salient positions in
[45:50] their portfolio?
[45:52] Salient meaning extreme returns on the
[45:54] winner or loser side. On the on the
[45:57] buying side, we find absolutely no
[46:00] correlation between prior returns and
[46:02] what they're buying for their portfolio.
[46:04] So, they seem to be so doing something
[46:06] we can't predict, which is why we're
[46:08] seeing skill.
[46:09] On the other hand, we can really really
[46:11] really predict what they're doing with
[46:13] their selling. They're basically
[46:16] mainly looking at the in the terms of
[46:18] the consideration set, they're mainly
[46:19] looking at these salient buckets of
[46:22] things that have gone down a lot or
[46:23] things that have gone up a lot, which is
[46:25] very salient on all of their screens,
[46:27] and basically choosing things from that.
[46:29] Um the reason that it tends to not it
[46:33] actually this strategy underperforms is
[46:35] because they're not just choosing
[46:37] randomly even from those buckets,
[46:38] they're selling something that that
[46:40] they've they're least attached to, which
[46:42] is the endowment effect.
[46:44] And because the
[46:46] the thing that they're least attached to
[46:48] are the things that they recently bought
[46:49] is actually generating them alpha, this
[46:51] is precisely what you shouldn't be
[46:53] selling.
[46:55] It's still get generating them alpha. We
[46:57] could actually map the alpha curve and
[47:00] show that this is where they're selling,
[47:02] and this is where they they should be
[47:03] selling, which is much later.
[47:06] Um so, that's kind of an Yeah, go ahead.
[47:09] If you were to look at short selling, I
[47:12] I guess the short selling would be like
[47:13] buying. If I explicitly go short a
[47:16] stock. So, we can't look at that in the
[47:18] data because they're not allowed to
[47:19] short. These are long-only portfolios.
[47:22] So, think like pension funds.
[47:23] Conceptually, that's right, Matt. Matt
[47:25] Marcus, that if it's a if it were a
[47:28] long-short portfolio,
[47:30] then the short it would be the
[47:32] initiations
[47:33] Yes. uh versus the closing.
[47:37] Yeah. And and that would
[47:39] That's an interesting follow-up paper.
[47:41] Just
[47:42] Give Alex the data from some hedge
[47:44] funds.
[47:44] >> Yeah, I need some data. That's the
[47:46] problem with all of this. I wanted to
[47:48] also ask, is there some tax management?
[47:50] Did you take the tax management into
[47:51] account? Yeah, we did. So, these are all
[47:53] tax-exempt portfolios. Okay.
[47:56] Yeah.
[47:58] Um Okay, so the the the the So, that's
[48:01] kind of an example of where the field
[48:02] has gone, which is again, you looking at
[48:05] anomalies with in field data. The other
[48:07] part that we wanted to to do in the book
[48:09] is to look back and to say, "Look, these
[48:11] anomalies form the bedrock of behavioral
[48:13] economics."
[48:15] As many of you know in the in the
[48:18] audience, there's been a bit of a
[48:19] replication crisis in the social
[48:21] sciences where, you know, you go out
[48:23] there, you replicate a bedrock study,
[48:26] uh and it doesn't replicate.
[48:27] And this is a big problem if you have a
[48:29] whole field based on uh studies that may
[48:32] not replicate. So, a natural question we
[48:35] wanted to say is, "Do these anomalies
[48:37] replicate?" So, what we did in the book
[48:39] is we took the main experiment for from
[48:41] every single chapter.
[48:43] We created instructions for it,
[48:45] basically mirroring uh the instructions
[48:47] that that we that that are available,
[48:49] and we ran it. And then all of this, as
[48:52] I'll I'll say in a couple slides, all of
[48:53] this is included in the online material
[48:55] of the book, so you can actually do it
[48:56] yourself.
[48:57] >> experiments?
[49:00] Um the the lab experiments.
[49:03] Pardon me? Sorry. So, So, So, it's
[49:04] actually So, as I mentioned before, the
[49:06] majority of the original anomalies
[49:08] columns were based on lab experiments.
[49:10] Yeah. The ones that weren't based on lab
[49:12] experiments used publicly available
[49:15] data, and we did that, too.
[49:16] So, for example, the equity premium
[49:18] puzzle used publicly available data. So,
[49:21] we just took that data and extended it
[49:22] through 2024, and showed that we find
[49:25] very similar results.
[49:28] Um so, we uh this is what I just said.
[49:30] We ran all of these experiments on
[49:32] Prolific.
[49:34] Um and what we found is that the
[49:36] anomalies are remarkably robust. So,
[49:38] this is the preference reversal of an
[49:41] phenomenon. So, people are asked they're
[49:43] they're they're given a choice between
[49:45] two lotteries. The one on the left has a
[49:48] fairly high chance of winning, but a low
[49:50] chance of losing 50 cents. The one on
[49:52] the right has a low chance of
[49:55] winning $40 and a high chance of losing.
[49:58] When people are asked what to do in
[50:00] terms of binary choice, they say,
[50:01] "Actually, I prefer the one on the
[50:03] left." 86 70% participants do that.
[50:08] But when they're asked their willingness
[50:09] to pay, which is what Richard alluded to
[50:11] earlier, they flip. They actually assign
[50:14] a higher willingness to pay. They value
[50:17] the one on the right higher. So, this is
[50:19] the classic preference reversal
[50:20] phenomenon. We replicate that one.
[50:23] The dictator game, all of the social
[50:26] preference experiments essentially
[50:27] showing, "Look, if you're I'm playing a
[50:29] dictator if I'm a can divide money
[50:31] between myself and another person.
[50:33] I have $10. The other person
[50:35] here 100 units, which translated to $10.
[50:40] How much do I give to the other person?"
[50:42] Pure selfish model says I should give
[50:44] zero. This one
[50:46] the prior studies found that people
[50:49] generally give 50%
[50:51] and the minority gives zero and we
[50:53] replicate that
[50:54] perfectly well. The ultimatum game also
[50:57] replicates very well.
[50:59] The endowment effect replicates really
[51:01] well. So, you could say, "Look, half of
[51:03] you are given mugs, half of you are
[51:05] given mini artwork. This is a picture of
[51:07] the mug. Here's the picture of the
[51:09] artwork. Your choice your and then
[51:11] you're asked, do you want the other
[51:12] one?"
[51:14] Basically, do you want to trade? If
[51:16] things are randomly assigned, Coase
[51:18] theorem says, you know,
[51:20] it shouldn't matter what your random
[51:21] assignment is in terms of how much what
[51:23] what's the next thing you want.
[51:25] Turns out it has a huge effect. And so,
[51:28] if people endowed with art, most of them
[51:30] want the art. They want to keep it.
[51:32] Those that are endowed with the mug,
[51:33] most of them want to keep the mug. You
[51:35] also find one result which was not
[51:38] robust.
[51:40] Um Anomalies.
[51:41] >> So, the only result So, the only result
[51:44] that we found
[51:47] that wasn't
[51:49] robust was In prospect theory, there
[51:52] were 14 questions. Mhm.
[51:57] What There are several questions that
[51:58] show that convexity over losses.
[52:01] Diminishing sensitivity. Diminishing
[52:03] sensitivity.
[52:05] The one of the questions was between
[52:07] getting zero. Sorry, the lottery had
[52:10] zero
[52:11] as an outcome and the other one just had
[52:15] two two losses. Mhm.
[52:18] And the lottery with zero had a higher
[52:21] had a probability of a larger loss than
[52:23] the other one. So, selecting it meant
[52:24] that you were risk seeking. We
[52:26] replicated that one. But then there was
[52:28] another question where both of the both
[52:31] of the lotteries were in the in the loss
[52:33] domain, but there was no opportunity to
[52:35] recover. I see. And that one we didn't
[52:37] find support. That was about a 50/50
[52:39] choice.
[52:40] But that was the only one and that's
[52:42] that's been actually shown in in in
[52:44] subsequent studies that have been
[52:45] replicating the the
[52:48] the the prospect theory questions is
[52:50] that essentially 13 of the 14 replicate
[52:52] really well, but that's the one that
[52:54] that that that
[52:56] tends to fail to replicate. That's out
[52:58] of all of the ones that we've tried,
[53:00] that's the only one we found.
[53:02] Okay. So, you made a serious check of
[53:05] all the anomalies.
[53:07] And only this little one, I mean, sort
[53:09] of first order
[53:11] finding what could not be replicated.
[53:13] Yeah. Next slide.
[53:14] >> Yeah, and in fact, that's replicating
[53:17] a finding Eric Johnson and I had in
[53:20] 1990.
[53:24] Yeah, so as I said, like there's
[53:27] there's the the the
[53:29] risk seeking over losses is really
[53:32] driven by the ability to get back to
[53:33] zero. And Makes sense.
[53:36] Makes sense. So, all of this is
[53:39] available online including the material.
[53:43] So, you don't have to trust us. You can
[53:44] download the materials and there's
[53:45] actually instructions on how to do it
[53:47] yourself, how to run these experiments.
[53:49] We also include teaching slides if you
[53:51] want to teach any of the chapters or
[53:53] anything like that. So, you can go ahead
[53:55] and and
[53:56] and and take a look.
[54:00] Um
[54:01] So, kind of taking stock, we kind of
[54:04] tossed we kind of thought about how to
[54:06] frame this
[54:08] and uh
[54:10] Our view is that it's not really that
[54:12] surprising that the anomalies are are
[54:14] are are are replicating.
[54:17] Part of the initial pushback for
[54:18] behavioral
[54:19] against behavioral economics was to say,
[54:21] "Look, we don't think these are very
[54:22] robust." And so, economists like Richard
[54:26] was talking about Charlie Plott and
[54:28] Grether, they they took and they took
[54:30] these instructions and they started
[54:31] replicating these things and they found
[54:33] that they were robust. By the time
[54:35] Richard was writing the writing the
[54:36] anomalies chapters, most of them had
[54:39] been replicated already. And
[54:42] many of these anomalies are used as
[54:44] classroom demonstrations of behavioral
[54:46] economics. If in a classroom something
[54:48] kept being not replicated, that would
[54:51] probably because of frequent replication
[54:53] attempts, chances are that this would
[54:54] come up. So,
[54:56] I think this is a really yeah.
[54:57] >> really good nose to identify in the '80s
[55:00] and '90s already which anomalies are
[55:03] robust.
[55:04] And advanced. There must be some other
[55:06] anomalies.
[55:07] I wasn't picking them at I wasn't
[55:09] picking them at random, Marcus.
[55:12] Exactly. Yeah. So, but that's all all
[55:15] we're claiming is that the topics I
[55:19] chose to write anomalies about in the
[55:22] '80s and '90s,
[55:24] those replicate. Yes.
[55:29] Fair enough. Yes.
[55:32] Yeah, and I think I think the other
[55:34] thing that that we argue is that
[55:38] like behavioral economics is very
[55:40] closely linked to experimental
[55:42] economics.
[55:43] These are two separate fields, but
[55:44] they're closely linked. And the methods
[55:47] that behavioral economics adopted were
[55:48] the methods that were in experimental
[55:50] economics. And experimental economics
[55:52] used methods that are quite different
[55:54] from psychology. One of the main things
[55:57] that they differs is that from the
[56:00] beginning and then since the '77 Vernon
[56:02] Smith paper is that all instructions and
[56:04] all the data were just stapled in the
[56:06] back of the paper.
[56:08] And because of that, you know, when a
[56:11] paper a lot of the time when a paper was
[56:13] published,
[56:15] it built on an existing paradigm. So,
[56:18] one of the conditions was a replication
[56:20] of the previous result. So, if you look
[56:22] at all of the bubble experiments that
[56:24] have come up in the AER Econometrica,
[56:26] they have a previous demonstration as a
[56:30] replication.
[56:32] And behavioral economics has kind of
[56:33] evolved in a similar way where you say,
[56:35] "Here, I want to explain something about
[56:37] prospect theory.
[56:39] Before I, you know, do the contribution,
[56:43] one of my control conditions is actually
[56:44] going to be the condition from the
[56:46] previous experiment. And the reason that
[56:48] I'm able to do that is because the
[56:50] people who had published the previous
[56:51] paper included their instructions and
[56:53] included my their data. And that's been
[56:55] kind of um
[56:56] a a tradition that's luckily been kept
[57:01] up throughout the process where
[57:04] it's it's
[57:06] it's kind of like it has the hallmark of
[57:07] accumulative science
[57:09] in some sense. There's also a big
[57:11] difference between
[57:13] experimental economics and experimental
[57:14] psychology with regard to deception.
[57:17] Yeah. Yeah. Can you say something? Is it
[57:20] experiments which involve deceptions are
[57:22] equally robust? Or is it because
[57:24] economics focused very much on
[57:26] experiments without deception?
[57:28] I honestly don't think the deception
[57:30] part is a big deal
[57:32] for as far as replicability. I think
[57:33] it's a it's a big deal as far as
[57:35] potentially a big deal. This This is
[57:37] actually a controversy in the field.
[57:38] There's papers being written about if
[57:41] people think that they're being
[57:42] deceived, are they going to act
[57:43] differently? So, this like poisoning the
[57:45] pool argument. Most of the papers I've
[57:48] seen have shown very little difference
[57:51] in terms of how people behave.
[57:53] Um
[57:54] And so, I think it's an important norm
[57:57] to have just morally and potentially it
[58:02] has an effect. But as far as
[58:04] replicability, I think the biggest thing
[58:06] in experimental economics has been the
[58:08] tradition of posting your data and
[58:10] posting your instructions. And you still
[58:12] see that, you know, you review a paper
[58:13] and it's very common for a reviewer such
[58:16] as myself to email the editor saying,
[58:17] "Hey, they didn't include their
[58:18] instructions. Can you email the authors
[58:21] and send me the instructions so I can
[58:23] take a look."
[58:26] What's about that the instructions
[58:27] become too complex? What Richard was
[58:29] talking about, you know, immediate
[58:31] attention. If I have five pages I have
[58:34] to read,
[58:36] Yeah, this is actually this is actually
[58:37] come up recently in in in in
[58:40] experimental economics. People worried
[58:41] about noise driving some of the results.
[58:44] This is So, this is actually you brought
[58:47] up an active area of research. This gets
[58:49] me to the next slide about what's next.
[58:52] What will behavioral economics look like
[58:54] in the future? And I think that's
[58:55] exactly right, Marcus, is thinking about
[58:58] um what parts of experimental economics,
[59:01] economic decision making in in
[59:03] observational data are driven by
[59:06] responses to to difficult decisions. How
[59:09] do you model difficult decisions? How do
[59:11] you think about the response to
[59:13] difficult decisions? One approach has
[59:15] been to say, "Look, they're going to be
[59:17] noisy." But that's actually not even
[59:20] predominant response. A lot of the time
[59:22] people switch to using
[59:24] actually super predictable and stable
[59:26] heuristics.
[59:28] So, the data actually looks less noisy
[59:30] as the environment becomes more complex.
[59:33] So, it's a very and
[59:35] it's it's a very difficult problem, but
[59:38] this
[59:39] you know, I'm visiting Princeton and
[59:41] Leif, Pietro, and I are actually running
[59:44] a reading group for graduate students in
[59:45] complexity. Trying to read the reading
[59:48] the literature from across the social
[59:50] sciences and trying to make progress on
[59:52] that. It's a it's a complete it's a open
[59:54] area of research that that
[59:56] that
[59:58] eager students can make a lot of
[01:00:00] progress on.
[01:00:02] So, let me ask So,
[01:00:04] I guess let me ask a question about
[01:00:07] complexity. Do you have complexity
[01:00:08] measure? Because Richard started out
[01:00:10] with that, you know,
[01:00:12] we can't really measure complexity.
[01:00:15] It's also, you know, when the financial
[01:00:16] crisis happened, so some assets were
[01:00:18] more complex than others. CDO squared
[01:00:20] was more complex than a CDO.
[01:00:23] And you know, but Goldman Sachs probably
[01:00:26] is if you were to go from modeling
[01:00:28] bottom-up, would be the most complex
[01:00:30] because they have CDO squared and CDOs.
[01:00:33] Do you have any idea Do you have any
[01:00:34] suggestions how we could measure
[01:00:36] complexity of a situation or of an asset
[01:00:38] itself or
[01:00:40] Is there something out there in the
[01:00:41] literature yet? In your reading group?
[01:00:44] There's a lot of stuff out there.
[01:00:46] There's a lot of suggestions. It but
[01:00:48] Richard actually a lot of them draw on
[01:00:51] computer science.
[01:00:52] Because computer science has a you know,
[01:00:54] 20 years of literature on complexity.
[01:00:57] But as Richard said, you know, these
[01:00:59] measures don't actually line up very
[01:01:01] well with how people perceive
[01:01:03] complexity and I mean, if you if you
[01:01:05] give me a hundred things versus ten
[01:01:06] things, I'll say like the hundred thing
[01:01:08] is more complex. That's those
[01:01:09] definitions align. But you know, you
[01:01:12] have choices like the CRT where people
[01:01:15] get it wrong, which is
[01:01:16] computer program would say like, look, a
[01:01:18] bat and a ball one cost ten cents more
[01:01:20] than the other.
[01:01:22] This is not a complex problem, but yet
[01:01:24] people get it wrong all the time. So,
[01:01:25] there must be some element of complexity
[01:01:28] in that problem. Um
[01:01:30] So, I think there are measures of
[01:01:33] complexity to answer your question. But
[01:01:36] we need a lot more research to show when
[01:01:40] specific measures of complexity maps
[01:01:43] onto the ecologically valid response to
[01:01:45] complexity and perception of complexity
[01:01:47] that people have. Because sometimes
[01:01:49] people view a hundred things and they
[01:01:50] actually don't think it's complex
[01:01:52] because there's like a salient thing
[01:01:53] that they all look at and they're
[01:01:55] saying, I'm just going to go with that.
[01:01:56] This is actually a very simple problem.
[01:01:59] And sometimes I only have to look at the
[01:02:00] problem from a different angle and
[01:02:01] suddenly it becomes non-complex.
[01:02:04] Yeah, exactly. Exactly. And and that's
[01:02:06] that's the problem with trying we end
[01:02:09] the book by saying we think
[01:02:12] a big big problem is
[01:02:15] what is difficult.
[01:02:17] Mhm.
[01:02:18] And we say warning, that's a difficult
[01:02:20] problem.
[01:02:23] That's a nice ending. Let me just ask
[01:02:26] You said going mainstream. So, that
[01:02:28] means, you know, behavioral becomes more
[01:02:30] integrated or part of the main canon.
[01:02:34] Would you also say that experimental and
[01:02:36] behavioral essentially becomes one field
[01:02:39] or where do you see the big differences
[01:02:41] going forward? Yeah, yeah,
[01:02:43] let me answer that, which is I think the
[01:02:45] biggest thing that's happened to
[01:02:46] behavioral is that it's gone Okay, hold
[01:02:49] on. Alex, you answer that. I have to
[01:02:52] go away for one second. Okay. Okay.
[01:02:55] Um
[01:02:56] Uh
[01:02:58] So, we think that you know, we've I say
[01:03:01] we because we've had this conversation a
[01:03:03] lot. Um
[01:03:04] Where I the where I was going to end on
[01:03:07] with going mainstream is what Richard
[01:03:09] had talked about before, which is about
[01:03:10] the textbooks and how economics is
[01:03:12] taught. And in one sense, behavioral
[01:03:14] economics is extremely mainstream in the
[01:03:16] sense that there's, you know, almost
[01:03:18] every top department has behavioral
[01:03:19] economics on the faculty. Behavioral
[01:03:20] economists are publishing in top
[01:03:22] journals. Um there's, you know, that
[01:03:25] we're having this conversation. Uh but
[01:03:27] on the other measure of mainstream,
[01:03:29] behavioral economics has completely
[01:03:30] failed. Like colossally, right? And that
[01:03:34] measure is what's in the textbooks that
[01:03:35] people use when they learn economics.
[01:03:38] And uh
[01:03:39] you know, apart from like little segues
[01:03:42] in some cases and
[01:03:45] you know, Laibson, Acemoglu, and List
[01:03:47] have a micro textbook that has a little
[01:03:50] bit flavors of behavioral economics. But
[01:03:52] on on the whole like the the the the
[01:03:55] the the core of economics that's being
[01:03:57] taught is still very much what you and
[01:03:59] Richard came up with in some ways.
[01:04:01] There's more game theory, obviously. Or
[01:04:03] perhaps it's because you need to
[01:04:05] understand the benchmark first before
[01:04:07] you understand the deviations.
[01:04:10] That's certainly true. But I think in in
[01:04:13] many ways and and I agree with you that
[01:04:15] that we should like this is the earlier
[01:04:17] conversation, which I completely agree
[01:04:19] with. You need benchmarks. Um
[01:04:22] But at the same time like, you know,
[01:04:23] thinking about like moving on to the
[01:04:25] next concept after learning the
[01:04:26] benchmark,
[01:04:28] you might want to know how people
[01:04:30] actually behave relative to that
[01:04:31] benchmark. And that's not done. You just
[01:04:33] move on to the next benchmark and hope
[01:04:35] that somebody takes an an elective that
[01:04:37] may not even be taught at the school.
[01:04:39] So, when I went to Northwestern, there
[01:04:40] was no behavioral economics course. So,
[01:04:42] all you walked away from as an
[01:04:44] undergraduate is perfect competition a
[01:04:46] lot of the time,
[01:04:48] you know, hyper-rational agents,
[01:04:50] rational expectations. You go out and
[01:04:52] then you start thinking,
[01:04:53] you know, that markets are are are
[01:04:56] awesome, which I mean, they are awesome.
[01:04:58] But you know, in they they they there's
[01:05:00] market failure.
[01:05:02] And people are hyper-rational, so we
[01:05:04] don't need any kind of
[01:05:09] regulation on shrouded attributes and
[01:05:11] things like that.
[01:05:12] Marcus, which issue do you want to come
[01:05:13] back to?
[01:05:15] Yeah, Marcus. There's just Sorry, I had
[01:05:17] to step away for a second.
[01:05:18] When I first started teaching a PhD
[01:05:21] course in behavioral finance at Chicago,
[01:05:24] I
[01:05:25] made Gene's class a prerequisite. Mhm.
[01:05:31] Uh I will say he didn't make my class
[01:05:35] a prerequisite. But
[01:05:37] Because you wanted to shoot him down.
[01:05:41] You know, I just wanted people to be
[01:05:43] informed. Very good.
[01:05:46] So, what's next? Is the element of
[01:05:49] artificial intelligence by purpose not
[01:05:51] on this slide? What do you think
[01:05:53] artificial intelligence will play a big
[01:05:55] role
[01:05:56] in behavioral economics down the road in
[01:05:57] the next ten years?
[01:06:00] Um so, I work in quite a bit in our in
[01:06:03] AI research these days.
[01:06:06] I Your your your colleague Tom Griffiths
[01:06:09] is doing excellent work at the cognitive
[01:06:11] psychology department.
[01:06:13] Um
[01:06:14] To be honest, I just I just I I'm not
[01:06:17] quite sure what it's going to look like
[01:06:19] yet.
[01:06:21] I think
[01:06:22] there is a vision where which you
[01:06:26] mentioned earlier that behavioral
[01:06:27] economics ceases to exist
[01:06:31] because AI just kind of does all the
[01:06:32] decisions for us and it's perfectly
[01:06:33] rational. But I don't see that
[01:06:35] happening.
[01:06:37] Um and
[01:06:39] in the near future, sorry. I mean, I
[01:06:41] don't know what's going to happen in the
[01:06:42] in the far future. But if you were to
[01:06:44] advise a PhD student who thinks, oh, he
[01:06:46] wants to work in that space, would you
[01:06:48] say, oh, that's too risky?
[01:06:50] Or that's something
[01:06:51] >> No, not at all. I think I think I think
[01:06:53] the intersection between cognitive
[01:06:54] science and AI is is is an exploding
[01:06:57] field and I think
[01:06:59] one area that AI can potentially help us
[01:07:01] with, particularly if you're working
[01:07:03] with small models which are
[01:07:04] interpretable
[01:07:05] to some extent, is eliciting how people
[01:07:08] perceive a problem. Exactly about the
[01:07:10] complexity issue. So, you can instead of
[01:07:12] like digging into a person's brain, you
[01:07:14] could think about approximately
[01:07:16] representing some part of the decision
[01:07:18] of the of the judgment process and
[01:07:20] thinking, can we elicit something useful
[01:07:22] here? Uh this is an active body of
[01:07:24] research.
[01:07:27] Very good.
[01:07:29] So, I like that we ended up with the
[01:07:30] next elements. I think I will wait for
[01:07:33] the third edition.
[01:07:35] Hopefully it shouldn't take so long
[01:07:36] anymore. Perhaps in ten years.
[01:07:39] You know,
[01:07:40] when I came out with an edition a new
[01:07:42] edition of Nudge, we called it the final
[01:07:45] edition. Oh, is it?
[01:07:48] Uh we didn't do that with this, but
[01:07:50] every 30 years I think I can confidently
[01:07:53] predict that I will not be part of the
[01:07:55] process.
[01:07:57] You never know. Perhaps you made some
[01:07:58] big progress on the next one.
[01:08:03] Okay, thanks a lot. It was a pleasure
[01:08:05] and I have learned a lot. And thanks for
[01:08:07] all of you for attending.
[01:08:09] And I hope to see you soon again for
[01:08:11] another webinar.
[01:08:12] Thank you.
[01:08:13] >> Thank Thanks, Marcus. Thank you, Marcus.
[01:08:15] >> Bye-bye. Bye.