The Orthogonal Bet: Complex economics is applying complex systems methods

Danny Crichton:

Hey, it's Danny Crichton here. We take a break from our usual Riskgaming programming to bring you another episode from our ongoing mini-series, The Orthogonal Bet. Hosted by Lux's scientist and resident Samuel Arbesman, The Orthogonal Bet is an exploration of unconventional ideas and delightful patterns that shape our world. Take it away, Sam.

Samuel Arbesman:

Hello and welcome to The Orthogonal Bet. I'm your host, Samuel Arbesman. In this episode, I speak with J. Doyne Farmer, a physicist, complexity scientist and economist, and currently the director of the Complexity Economics Program at the Institute for New Economic Thinking, at the Oxford Martin School and the Bailey Gifford professor of complex system science at the Smith School of Enterprise and the Environment at the University of Oxford. Doyne is also the author of the fascinating new book, Making Sense of Chaos: A Better Economics for a Better World. I wanted to speak with Doyne about his fascinating history in complexity science, his new book and the topic of complexity economics more generally.

Doyne and I had a chance to discuss the nature of simulation, complex systems, the world of finance and prediction, and even the differences between biological complexity and economic complexity. And we even got to chat about Doyne's experience building a small wearable computer in the 1970s, that fit inside a shoe and was designed to beat the game of roulette. This was a fascinating conversation. Let's jump in.

Doyne, great to be chatting with you and welcome to The Orthogonal Bet. Thank you so much. You've just published this thought-provoking new book entitled, Making Sense of Chaos. It connects complexity science to economics and finance, and there's lots of different topics in there. I think the best place maybe to start is with complexity economics. I'd love to learn more about how you define what this field is and why is it so powerful as opposed to more traditional economics?

Doyne Farmer:

Well, so complexity economics is loosely defined as applying complex systems methods to economics. So then you say, well, what is a complex system? And that's something that frankly we can't define very well, but we kind of know them when we see them. A great example would be the human brain, which consists of neurons, 80 billion of them. Each neuron is itself a simple little device. Well, maybe not so simple, but you feed a signal into it, and it may or may not have a signal coming out of it. But yet you hook 80 billion of them together and you get something that's just qualitatively completely different from a single neuron. So it's an emergent behavior. It emerges, and you couldn't have predicted just from looking at a neuron, that you get something like a human brain.

And there are many examples of that. Ant colonies or another great one. Ants are very simple little organisms, yet ants can capture slaves, ranch, farm, do all kinds of things that they've been doing for a long time before people even existed. And economics is a great example because the economy allows us to do all kinds of things we could never do by ourselves. Now to say how complexity economics differs from mainstream economics, I have to go into a longer story about how we both approach the world.

Samuel Arbesman:

Please do.

Doyne Farmer:

So maybe it helps to start with mainstream economics. Mainstream economics, how does it work? And this is something that's evolved over 150 years. In a mainstream economic model, and I want to emphasize model as opposed to data analysis. Economists do a lot of data analysis. They do regressions. They do what's called econometrics, where you just take data and you fit a model to it. So I'm not talking about that. I'm talking about where you try and build a causal model, where you try and really understand what makes things tick. Economists approach that by assigning every agent, take a person or an institution like a firm, you assign them a utility function, that's a scorecard that says what they like. It gives an ordering, I like this better than that. And then you give them a model of the world. The classic one that's been used a great deal is rational expectations, which means that you can reason about the world, think Mr. Spock from Star Trek. You reason about the world, you take all the information you have and you extract all the reasoning you can out of that information.

And finally, they assume equilibrium, which in a simple case just means supply equals demand, but sometimes it means that your strategy is as good as it can be, providing nobody else is changing their strategy, and that's true for everybody. So you take those elements, you write them all down in equations, you solve those equations, you deduce what each agent's decision or decisions are going to be, and then you calculate the economic consequences of those decisions. In a world where everybody makes those decisions, what does that do to the economy? And there's caveats to that. In modern economics, behavioral economics has come to play a big role. But I would say that behavior is not well-fused together with modeling yet, and that nobody really knows how to incorporate behavioral observations, put them into the kind of model I described and get everything to match reality.

Now in complexity economics, we do something very different. Building on studies of other complex systems where we model from the bottom up. And in economics model, we make some assumptions about how the agents make their decisions. We typically assume they use some rules to do that. Those rules could be very simple. They could also be based on learning algorithms or even reasoning. But if they're reasoning, it's not rational expectations, we assume they reason a few steps ahead, not an infinite number of steps ahead. So we make assumptions about the agent's decision-making rules. We give the agents the information or model agents. They make decisions. Those decisions have economic consequences that we calculate, that generates new information because the economy just changed. Other information may also flow in from the rest of the world. Then the agents make a decision again, and we just go around and around that loop.

Samuel Arbesman:

And to be clear, so that part of this kind of dynamic feedback, that's a very clear relaxation of the whole equilibrium assumption of classical economics [inaudible 00:06:37]

Doyne Farmer:

That's right.

Samuel Arbesman:

Okay.

Doyne Farmer:

That's right. Now equilibrium may emerge.

Samuel Arbesman:

Sure.

Doyne Farmer:

We may discover after a while, that the agents are just making the same decision over and over again, or that the economy is in some kind of static state or it's fluctuating, but it's sort of staying around the same place. In which case we would say, ah, an equilibrium emerged. It's an emergent phenomenon in our models, because we typically don't know in advance whether we're going to reach it.

Samuel Arbesman:

Okay. Right, so rather than assuming that equilibrium is already there.

Doyne Farmer:

Exactly.

Samuel Arbesman:

Okay. And then related to that also, how would you think about the scale and maybe the heterogeneity of agents as well? Is that something else in terms of the number of agents, the differences between agents, the features of complexity economics, that they handle in a way that traditional economics is not?

Doyne Farmer:

Yeah, you are right on point.

Samuel Arbesman:

Awesome.

Doyne Farmer:

You hit the nail on the head. Two of the key things that we can do are heterogeneity. We have simulations in the macro models we're building now, where we have millions of agents.

Samuel Arbesman:

Okay.

Doyne Farmer:

And these agents can be quite heterogeneous. In fact, what we try to do at the really micro-macro models, we're trying to model the economy of a country or even the whole world. We create synthetic populations that we try and demographically match the real populations. So we try and match earning power, age, gender, race, anything that is considered important, we can match it. So I think there's a big advantage to agent-based modeling and doing heterogeneity. I mean, heterogeneity is a hot topic in macroeconomics these days, but they are forced because of the constraints of having to write down these equations and solve them and compute the optimal solutions. We don't compute optimal solutions. We just compute solutions the agents make. They may be trying to be optimal, but they're typically not achieving that. And that gives us a huge computational advantage, which means that our models are tractable, which is why we can do millions of agents.

In contrast, standard models, once you have more than about 10 variables in the model, meaning probably less than 10 agents, you can't solve the equations anymore, even with powerful computers. So that really restricts what you can do, and it means you have to keep the model simple. Now, simple is good if it's good enough. But there are a lot of situations where you need to go beyond simple, and we have a big advantage doing that.

Samuel Arbesman:

And so how do you think about... And you mentioned simple. How do you think about the trade-offs in terms of the sophistication of the simulation? Whether or not actually it's mapping onto the heterogeneity of the underlying agents in contrast to certain issues of understandability of the model? Because of course, I mean, you might be able to make a somewhat simple model of how an individual agent acts, but of course, due to all the interactions, there's going to be maybe a certain amount of surprise in what the resulting emergent behaviors are. Is the goal then a little bit different than perfect simulation and prediction? It's more about understanding the shape of these kinds of behaviors. How do you think about those in terms of the trade-offs?

Doyne Farmer:

Yeah. Well, perfect's a big word.

Samuel Arbesman:

Okay.

Doyne Farmer:

We're a long way away from that.

Samuel Arbesman:

Yeah, no, I understand that.

Doyne Farmer:

And we're never going to be there, never going to get there. I mean, models are never perfect, but some are better than others and some are useful. Now in economics, particularly in macroeconomics, models are not very good. Not surprising, it's a hard problem. But the challenge is to make models that are good enough to be useful, and one of the changes that's happened in complexity economics in the last five years is that we are now making quantitative agent-based models. So we're making models of a specific economy at a specific point in time. We try to make quantitative predictions and try and do quantitative policy analysis. If we change the world by implementing a new policy, does the world get better or worse? And if so, by how much? And we try and measure that.

Samuel Arbesman:

Okay.

Doyne Farmer:

That said, there's always a quest to figure out what are the things we really have to have in the model? What are the things we can do without? Because you want to make the model as simple as you can, but no simpler. And so we pay a lot of attention to that question, and we try and measure everything we can, and we try and match it against what the model's doing and figure out where the model is right and where the model is wrong, and then iterate by changing the model to make it better. Now, agent behavior is always a challenge because people are complicated. That's a place where we particularly can't expect to be perfect. What we see is in some cases, agent behaviors are really important and the model is very sensitive to the way you specify them, and in other cases not. We made a model of the Covid pandemic, for example, that made accurate predictions about how the UK economy would respond, before it happened. And believe it or not, agent behavior just doesn't matter in that model.

Samuel Arbesman:

Interesting.

Doyne Farmer:

Why? The key thing in the model was getting what's called the production function right. Well, first of all, predicting what shocks were going to be. How would the covid pandemic shock the model or shock the world? We did that by actually looking at a mixture of things. What we said in our model was that an industry can't produce if there's no demand for its good. Say there wasn't as much demand for airlines as there was before the pandemic. Airlines never shut down, but people just weren't flying. So that was a demand shock. For that, we relied on a survey done by the Office of Budget Management in 2006, [inaudible 00:12:24] pandemic. It was broadly speaking, pretty good.

An industry can't produce what it needs to produce if it doesn't have labor. So for labor, we looked at characteristics of occupations and what's the blend of occupations in each industry? And for that, we had this magnificent database put together by the Bureau of Labor Statistics in the US, with statistics like how close to people in this occupation typically work next to each other? Well, if they're working closer together than two meters to each other, well, they probably aren't going to be able to go to work. More than two meters, we assumed they could go to work. That allowed us to predict the labor shock.

Samuel Arbesman:

That level of fine grain detail was available in the labor statistics?

Doyne Farmer:

Yeah.

Samuel Arbesman:

Wow, that's amazing.

Doyne Farmer:

They had 450 occupations and we could go occupation by occupation asking, are these guys going to be able to go to work?

Samuel Arbesman:

Interesting.

Doyne Farmer:

And then we mapped that into the industry, because they also have a map of which occupations sit in which industry, and then we could predict the effect on each industry, of not having labor.

Samuel Arbesman:

Wow.

Doyne Farmer:

Then finally, an industry can't produce its good if it doesn't have the inputs it needs. There we did something novel, in that if you're the steel industry, you can't produce steel if you don't have iron or if you don't have energy to heat the iron or coke, because you need to blend it with the iron to make steel. But if you look in the input table that you'd see in an economic model, there are other inputs like restaurants. Restaurants well, they have a cafeteria, so they have an input to the restaurant industry. You can get by without that. Restaurant industry closed down, didn't affect steel production. Management consultants. Well, in the long run, that's a good thing to have. But over a two-month or three-month period, you can get by without them.

So we understood through a survey that the company called IHS Market thankfully did for us at lightning speed, [inaudible 00:14:21] all their analysts, what are the essential inputs for these industries, industry by industry? So we could see what the inputs they needed to have. And our model was extremely simple. It went at a daily scale. We initialized each industry with an inventory, which we got from a table compiled by the US statistics gathering, so we could look industry on industry, what are the typical inventories? And then every day we would produce the good. But if there was no labor, that would restrict prediction. If there was no demand, that would restrict prediction. And once that industry had used up its inventory, that would restrict predictions.

And then we hit the British economy with shocks from the supply side and the demand side, and our model effectively allowed us to see how those shocks propagated through the economy and even collided with each other because you'd see the labor shock from the supply side colliding with the demand shock from the demand side. And we see delayed effects. The industries might run out of inventories after a month in a hard lockdown policy.

And we analyzed different lockdown policies because the government at the time, UK government was in a hard lockdown, but they were considering coming out of the lockdown. So we analyzed different policies for coming out of the lockdown. We said, well, the one that looks good is one where you keep all the upstream industries open, all the industries that produce stuff. But the customer facing downstream industries, we suggested they close. Why? Because there was a lot of infection produced by those industries. So we said, live with the extra infection in the upstream industries, but stop the downstream industries, or at least make them operate differently. So you have to do stuff through the mail or pick it up in some COVID-safe way. And so we said it's the least bad solution because yeah, people still die, but not that many more people die than die in a hard lockdown. And the economic impact is way less than under a hard lockdown, which they've been in.

So we predicted the hit to the economy. We said, for example, in the second quarter of 2020, we expect the hit to be 21.5%. When the dust settled and the measurements were made a year later, well, it was 22.1%. So we hit the nail on the head. And we predicted lots of other things, unemployment. We predicted industry by industry. The predictions weren't perfect, but they were pretty good. They also matched their trajectory through time. And we didn't have to assume any agent behavior because all of it was just... It was almost a mechanical reaction. Now that's unusual, but in other models, we do have behaviors and they matter to varying degrees, and they need to be sophisticated or not to varying degrees.

Samuel Arbesman:

In terms of other use cases, do you feel that there are certain kinds of uses where this kind of complexity economics approach is better suited? Is it the kind of thing where it's flexible enough that really can be used in all different sorts of situations? How do you think about in terms of application?

Doyne Farmer:

Yeah, no, great question. There are certainly some situations it's better suited for. Why? Well, COVID for example, was a disequilibrium phenomenon. It was a sudden shock. Wham! The economy got whacked. So our model was really ideally... It was explicitly disequilibrium model. Supply was out of whack with demand. And I explained what I think was really the dominant thing that caused the economic problems, our model is capturing those, I think pretty well. So very well suited for that. We are building what I mentioned is a micro-macro model. It's a micro-macro model because we simulate behavior of individual households and individual industries, even individual firms. So it means we can look at micro questions like what happens to this specific industry or this specific demographic group, or these kind of firms? But we can also make macro predictions by aggregating everything up.

We can answer questions that a traditional macro model can't even ask. But the jury is out whether our performance at the macro level will be the same or better? There are a couple of papers that suggest that it's about the same. They're not as statistically significant as I would like to see, we're working on that. But the answer seems to be, yeah, they're about the same right now, but we have a lot of room for improvement because these are just early prototype models that have been... The amount of human effort in them is 1000th the effort that's gone into traditional macro models.

Samuel Arbesman:

Sure.

Doyne Farmer:

Let me maybe say any kind of disaster, I think we're better suited for because disasters in general are these disequilibrium phenomenon. Any kind of micro phenomenon, we made a housing market model, for example, where again, we had individual houses. We actually simulated the buying of selling in houses, we simulated going to the bank to get a loan and credit ratings and the types of mortgages that people were getting, and we showed how the type of mortgage affected the housing bubble. But again, you can't do that with a traditional mainstream model. For example, the way houses are actually sold, we simulated it in our model. You go to the real estate agent. The real estate agent finds some comparables. You look at what those comparables sold for, you make a guess. You put the house out on the market. If it doesn't sell after a month, you mark it down. Still doesn't sell, you mark it down again and by five or 10% each time.

So first of all, we looked at the data and we saw that's really what people do. We could put it in our simulation and simulate it. Now, when you do this in a model, you see that you can have demand differing from supply by an order of magnitude, and you see it flip as you go through the bubble. Leading into the bubble, there were 10 times as many buyers as sellers. Coming out of the bubble, the opposite was true. And our model, we just see that flip around. It also means prices respond sluggishly. And in a traditional model, you have to assume supply equals demand. And that's just a bad assumption. So that's another good example where our model could work better because we could more realistically capture the actual mechanisms that the economy really works under, at least for houses. By the way, that way of selling is an example of what's called a heuristic. It's called aspiration level adaptation. It's the way people sell used cars, lots of other things, actually. But it's hard to capture in equations. Much easier to capture in a simulation.

Samuel Arbesman:

Is the eventual goal to almost create some sort of massive master simulation of every aspect of economics and the way individual agents interact and respond to various phenomena, and then depending on the situation, almost simplify it for that specific use case? Right now, spinning up bespoke models depending on the specific situation.

Doyne Farmer:

Well, we ultimately want to do what you suggested.

Samuel Arbesman:

Okay.

Doyne Farmer:

We're trying to get there by bespoke models.

Samuel Arbesman:

Okay.

Doyne Farmer:

To state the big vision, and maybe I can take a step back before I state it. I started a company called Macrocosm, that employs now some of my best students and postdocs and so on, ex-students and postdocs. And our goal is to scale up complexity economic solutions, reduce them to practice so that we're making predictions day in, day out, and that we're really modeling things at scale. We're building a global model of the economy that's both a macro and micro model. It now runs for 40 countries individually, we want have that running altogether. But the goal is to really make predictions day in, day out for commercial purposes and use that as the goose to lay the golden eggs, to scale the whole enterprise up and go towards what you were suggesting. In 10 years, we would like to do for economic planning, what Google Maps has done for traffic planning.

So any business could look at themselves in the model, because we'll have all the businesses in the world in the model, and think about their business strategy, run counterfactuals. Well, what if we changed our strategy to do this? Or a government planner could say, what if we implement this new policy? How will that change the way things happen? A company would be able to think, how do we become more sustainable? Because we'd be able to see their environmental footprint and their carbon emissions. And help them find paths that would still make money for them, maybe even as much or more money, but have less carbon emissions than a smaller environmental footprint. So that's where we want to get. Now, that's a huge goal. To get there, we've constructed stepping stones, things that are already pretty useful that are short of that, but are moving us towards that goal.

Samuel Arbesman:

I mean, I almost view this as an... Obviously it's not quite as a game version, but a much more rigorous version of SimCity, but for economics, where it's like you're now given this massive model and then you can try counterfactuals, try different things, see how the system responds. And then use that as a way of trying to understand how to make decisions and how to understand.

Doyne Farmer:

Yeah.

Samuel Arbesman:

Yeah, the way the system bites back.

Doyne Farmer:

It's exactly SimCity but matched to a real city. In this case, we're matching to a real economy. And the big progress that we've made in the last few years is being able to make models like our Covid model, that made what we would call technically a time series prediction. Meaning we predicted next quarter if the government makes this decision, this is what the economy's going to do, this is what production will be in all these industries. Run that forward through time, and it's better both because it's more useful to have a more exact prediction, but also because it gives us a tougher scorecard. We know when we're wrong and we can keep adapting our model to make it better. So that's essential.

Samuel Arbesman:

Well, and certainly, I mean, having a certain sense of feedback I think is very powerful and much better than just having a talking head without much accountability. So that's [inaudible 00:24:46]. And then related to that, what has been the response both from governments and municipalities, and organizations that are interested in these kinds of predictions, as well as economists who are more in the traditional mold to this kind of approach?

Doyne Farmer:

So we're hearing a lot of interest from governments, municipalities, and so on, an interest we're sprinting to try and fulfill. And central banks. There is now a macro model made in this style, that the Canadian Central Bank is using. And is sitting there, it makes predictions alongside the traditional models. And the decision makers aren't slavishly following our model versus the other models, but my colleagues who were doing that model were very excited when they were able to do a better job of forecasting inflation than the traditional models.

Samuel Arbesman:

Interesting.

Doyne Farmer:

So that got the attention of the governors of the bank. There are probably six or eight central banks throughout the world now, that have housing models along the lines that I mentioned. The Bank of Italy is also busy building a macro model that they're going to incorporate and use. So we're breaking through in those spheres, and I think there's huge demand. And the economics community, meh! Been very little reaction to my book so far. I mean, from a few unique individuals like Larry Summers and John Giannakopoulos, and Andrew Lowe, and people that I know that are open-minded, I've gotten very positive reactions.

Samuel Arbesman:

That's good.

Doyne Farmer:

But I would say in general, the academic economics community is just not paying attention. And I've seen many examples now where graduate students go to their advisor and say, "Well, I'd really like to build one of these complexity economics agent-based models." And the advisor says, "If you want to have a career, you shouldn't do that." That's still, I think, the status quo.

Samuel Arbesman:

Okay. I was going to ask you why there's been that sort of response? And is it one of these kinds of things? What is it, Max Planck, the Planck's principle of you have to wait for a certain generation of scientists to pass away before the certain ideas are ready and whether not... And it could be one of those kinds of features there where it's, as long as the powers that be are in charge of granting tenure and reviewing certain papers, those kinds of things are not necessarily going to be the things that make your careers.

Doyne Farmer:

That's exactly right. There's just a view that this is not the right way to do things. It's not the way we economists do things. And if you want to do that well, you're in a different club. I do sense a lot of interest particularly from younger economists. I think there's more open-mindedness than there was, and I do think that things are going to start shifting, but by and large, we're still blocked from publishing in the mainstream journals that have to be published in to get tenure at the prestigious institutions.

Samuel Arbesman:

Related to that, your background is not originally in economics. You've kind of had this interesting trajectory which perhaps has allowed you to sidestep some of that academic infighting or whatever the term might be. Maybe you can share a little bit of your trajectory. I believe you started in physics, and you can kind of talk about your trajectory of going from that through complexity science, now into complexity economics.

Doyne Farmer:

Yeah. Well, I started in physics. I was actually doing physical cosmology.

Samuel Arbesman:

Okay.

Doyne Farmer:

I dropped out of graduate school for a year because actually my best friend from when we were children, who was also a graduate student at UC Santa Cruz where we were, and I led an effort to beat the game of roulette.

Samuel Arbesman:

There's actually a great book about this.

Doyne Farmer:

Yeah. Written by our friend Thomas Bass, The Eudaemonic Pie in America, the Newtonian Casino, in Britain.

Samuel Arbesman:

I want to continue the trajectory, but maybe you can kind of pause there and talk a little bit about that effort and how it went.

Doyne Farmer:

Yeah. So we used physics to beat the game of roulette. We took advantage of the fact that there's about a 10 to 15 second interval between when the croupier releases the ball and when the croupier closes the bets. We've done an extensive study of roulette wheels, it's just a ball rolling on a circular track that's doubled. So we solve the physics of that. So measured initial conditions using a computer by clicking with switches in our shoes, that were with our toes, and measuring how long the ball took to complete a circuit on the wheel, made a prediction, laid down bets, and we beat the house with about a 20% edge. We had a lot of hardware problems, we were really pushing the envelope. And we were doing this at the same time that Jobs and Wozniak were making the very first early, early Apple computers.

Samuel Arbesman:

Okay, so this is early to mid 70s, right? So there were no personal computers at this time?

Doyne Farmer:

Yeah, 1977, '78. And we were also a bit chicken. There were well-documented stories of people getting beaten up in the back room, and so we didn't want to go there. So statistically we did great, we didn't get rich.

Samuel Arbesman:

But it was a powerful proof of concept.

Doyne Farmer:

It was a great proof of concept. Then I worked at Los Alamos for 10 years, did a lot of stuff with prediction. When I would talk about my prediction work, somebody would say, "Well, have you tried applying this to the stock market?" So I finally got fed up with hearing that question and joined back together with Norman. We then did apply our stuff to the stock market, has a big advantage that you can make the stakes as big as you want, and they don't beat you up for winning.

Samuel Arbesman:

Right, there's no back room where you're getting hit.

Doyne Farmer:

Yeah.

Samuel Arbesman:

Neat.

Doyne Farmer:

And it took us a while, but we did eventually break through and we did quite well, we made a good amount of money. Enough that I could quit and do whatever I want.

Samuel Arbesman:

That's great.

Doyne Farmer:

So that's essential for actually this story because as a result of that, I'm independent and if I don't have a job, well, so what? Since then, I was at the Santa Fe Institute for 12 years. I've been at Oxford for 12 years, and I was in the math department at Oxford, and now I'm in the department of geography and the environment. I've attracted a lot of brilliant young students who come to me saying, "Well, I really want to do economics, but I don't like the way the economists do it." And I read them the Riot Act, say, well, I'm happy to work with you, but you're not going to get a job at Harvard even though you're probably brilliant. They go, "Okay." So I've got a cadre now of ex-students who have come through that, and I think we're all really, really good.

Samuel Arbesman:

And do they end up going to different places elsewhere in academia, or do they also... You mentioned some of your students are now at your company, Macrocosm. Do they end up in the tech world or startup world, or work for less traditional research organizations?

Doyne Farmer:

Yeah. Some of them are in academic departments. I have one student who's in the informatics department at University College London, another student who's in the logistics department in the Vienna Business School. I have a ex student who's at the World Bank, another ex student who's at the International Monetary Fund, another one who's in the Central Bank of Spain. A few of them are doing postdocs and sort of complex systems, interdisciplinary departments. Some of them are working for Macrocosm. I'd love to just hire them all.

Samuel Arbesman:

Do you suggest that any of them take some time off, beat the stock market and then go off and go back to academia?

Doyne Farmer:

Well, some of them are working for hedge funds. To be honest, I try to discourage that because okay, it's a great way to make money, but it doesn't do much good for the world. I have mixed feelings about my own time doing that. I did it for eight years and quit, so I made some money. Great. But this time with Macrocosm, we really want to produce tools that can help us make better decisions and guide us in particular, through the green energy transition so that we can stop emitting as much carbon as we're doing now.

Samuel Arbesman:

And so was that the impetus of beginning to focus more directly on complexity economics, like certain things around climate change? Was it saying, okay, I want to apply the same kinds of approaches that have been done in finance investment world, but for specific understandings of entire economies? Was there a specific moment that made you think you want do this applying complex systems to economics, or was it a slow shift?

Doyne Farmer:

I mean, I knew when I started prediction company with Norman, that I didn't want to do this for the rest of my life. My original goal was five years, $5 million. It was eight years, and I did significantly better than the $5 million.

Samuel Arbesman:

Congratulations.

Doyne Farmer:

So I was very happy with that. But I knew it was time to leave, and I knew I wanted to go back to doing basic research. And I was tempted to do bioinformatics because I think it's really cool. But I realized because I've been reading the finance literature and actively working in it for eight years, I now knew a lot about the financial world, and I'd also been reading the papers in finance and felt the paradigm they were using was not one that I thought was correct. And so I thought I should fuse my domain knowledge of finance with my experience and intuition about complex systems, and that's the direction I moved. So I started out at Santa Fe Institute, studying what's called market microstructure, which is nice because there's lots and lots of data, so you can make models that really have traction. On the other hand, it's not the most relevant problem for making the world better, and so I've drifted over time into macroeconomics and actually micro-macro economics. So that's I think more or less the evolution.

Samuel Arbesman:

Okay. No, that's great. And you mentioned that at one point you were tempted to explore biology and bioinformatics. When you think about simulation and complexity science and prediction of different systems, you mentioned before that the human brain has 80 billion neurons or whatever it is, and of course the earth has about 8 billion people. Obviously, it's not like, oh, the world is 10 times simpler than the human brain. There's a lot of different things going on here in terms of scale and how to understand this. How do you think about... And maybe this is even just ill-posed, but how do you think about different aspects of complexity of different kinds of domains? I mean, because in certain ways there's ideas from biology that are, I guess, portable or exportable to the world of economics. In terms of thinking about, I think in your book you kind of talk about almost like the metabolism of the civilization. You have ecological ideas and systems evolve over time. There's lots of similarities, but these systems are also very different. How do you kind of think about similarities, differences, differing levels of complexity of all these different domains?

Doyne Farmer:

Yeah, good question. On one hand, I think metaphors can be very useful because the way you think about the problem, it changes, then the way you ultimately solve it is different. On the other hand, you really have to roll up your sleeves and acquire the domain knowledge that you need. I took one economics course in my life, I wasn't very happy with it. Didn't learn mainstream economics very well. I spent a lot of time teaching it to myself. I've had the luxury of having some excellent friends who became my tutors. So when you have somebody who really understands it and you can just ask them questions, you'll learn it really fast. And domain knowledge actually is also not just the theories in mainstream economics, it's the nitty-gritty of how stuff works.

We made models of what's called the limit order book where you actually make trades, you have to really understand how the limit order book works. Put the rules in there very explicitly, understand how traders use it. So that's another kind of deep domain knowledge that we've drawn on a lot. So you have to fuse these two things together. On one hand, new ideas. On the other hand, you have to master the mechanics. It's often very useful to collaborate with people with other knowledge sets, who in some cases have been economists. In other cases, we'll have a team where I'll have one student who's an expert on model calibration, Bayesian estimation, and so forth.

Samuel Arbesman:

Sure.

Doyne Farmer:

Another student who's an expert on computer programmer. Others who really know a lot about ideas about macroeconomics. And we fuse it all together and create a final model with a synthesis of knowledge from different people.

Samuel Arbesman:

It sounds like once you assemble a team, the level of complexity of the system, it's less relevant because you're drawing on different domains. But I mean, in some ways though, I imagine the economy is a lot more complex than the bouncing of a ball in a roulette wheel-

Doyne Farmer:

Exactly.

Samuel Arbesman:

... given that you were able to spend a year and basically solve this problem. And in many ways, I mean physics is... And people talk about this, physics is much more simple because atoms and molecules can't think. Maybe is biological level simulation, is that the worst of both things, where it's like you can't abstract way all the details, but you have to deal with the messiness of economics, but in biology, so it's like somewhere halfway between, but also maybe even worse. How do you think about other domains like that?

Doyne Farmer:

Well, maybe you can think of two axes. One is how much thinking is going on or how much information processing, because even a bacteria processes information from its environment and makes decisions about which way to swim based on that. But obviously that's a lot less than human brain. And the other is, how complicated is it? Economics rates at the hard end of the scale in terms of thinking, because it involve people who think, but it rates at the lower end of the scale in terms of how complicated it is. Okay, a ball bouncing, that's really simple, that's less complicated than the economy, but biological organisms are pretty complicated. And the brain is really complicated. And I think understanding the brain is a harder problem than understanding the economy. The weather would be another example where it's really complicated, no thinking going on.

Samuel Arbesman:

So it's just a lot of sophisticated calculations at a very, very high level in order to understand what's going on.

Doyne Farmer:

Yeah, yeah. You might say, well, eight billion people in the world, that's a big number, but actually we can simulate eight billion entities making decisions in the cloud. Take state-of-the-art computing, but it's doable. There's only the order of 200 million firms in the world. Okay, that's a lot, but Orbis collects data on every one of those firms. So we know stuff about the balance sheet of every firm in the world or almost every firm. Now, we don't know as much about how those firms are connecting together, which firms are getting inputs from other firms. We know only about 1% of those connections. So there's a lot of uncertainty there. So that's the other problem.

Maybe a third axis is data. How much data do we have? And economics is pretty good. It's quite hard to gather the information in biology or neuroscience. Economics is easier. Now we're fighting the battle of confidentiality because we don't want to violate anybody's confidentiality, but on the other hand, we want all the data we can get. Economic sits in this middle spot. Eight billion is still a lot smaller than 80 billion. And we probably don't need eight billion people to really get it right, we think we can do this with synthetic populations of a few million.

Samuel Arbesman:

Interesting.

Doyne Farmer:

But economic interactions are... Markets are fairly well-defined things. Human decision making is the hard part, but as I said, oftentimes you can get a pretty good answer. Sometimes you find it doesn't even matter too much how people make decisions, as long as they make some kind of reasonable decision. And sometimes you find it doesn't matter too much, occasionally it matters a lot. So you really have to home in on those cases where it matters a lot. I think the jury's out of how many cases are there, where it really matters a lot.

Samuel Arbesman:

Interesting. Putting complexity economics into this larger framework, it might be a perfect place to end. Thank you again. Doyne Farmer is the author of the new book Making Sense of Chaos. Thank you so much for chatting. This is fantastic.

Doyne Farmer:

Thank you, Sam. It's a pleasure to be on this show, and you have such great questions.

Samuel Arbesman:

That's very kind.

‍