- | Monetary Policy Monetary Policy
- | Mercatus Original Podcasts Mercatus Original Podcasts
- | Macro Musings Macro Musings
- |
Lars Christensen on AI and its Impact on Monetary Policy and the Broader Field of Economics
As AI technology continues to progress, its effects may forever change the world of economics and economic policymaking.
Lars Christensen is a founding member of the market monetarist tradition, an entrepreneur in the AI space, and is also a returning guest to Macro Musings. Lars rejoins the podcast to talk about AI and its implications for the economy and for monetary policy. David and Lars also discuss the basics and implications of dynamic pricing, AI’s growing use within econometric analysis, how AI will impact the Fed and its policymaking, and much more.
Check out our new AI chatbot: the Macro Musebot!
Read the full episode transcript:
Note: While transcripts are lightly edited, they are not rigorously proofed for accuracy. If you notice an error, please reach out to [email protected].
David Beckworth: Lars, welcome back to the show.
Lars Christensen: Thank you, David. Great to be back.
Beckworth: Well, it's great to have you on. We were talking on Twitter, now known as X. I prefer to call it Twitter, so I'll use Twitter. We were talking on Twitter about some of the activities you've been doing with AI. And so, I want to talk about that later, but you also are an entrepreneur of sorts in AI. You've got your own business. In fact, you're moving, as I understand, from a lot of your focus being on macro work to AI. So, maybe tell us this story, [about] how you are transitioning towards more of an AI focus.
Lars’s Move from Macro to AI
Christensen: Yes. Well, I've been an economist now for nearly 30 years, and the entire period as a macroeconomist— first in government, then later in the banking sector, and running my own consultancy later on, and doing academic work— has been about standing outside of the economy and watching the economy. But there has been this common denominator that data has played a significant role in that. We both, of course, come from that market monetarist tradition, that markets are important, that the markets are telling our stories.
Christensen: And financial markets are interesting, because that's really the best reflection of expectations about the future. But I'm seeing a development where everything is becoming financial markets, that with the development in-- I essentially see it as three parts. First of all, we've had 35 years of the internet, or more, [and] these enormous amounts of data that we all have access to. At the same time, the computing power has gone up dramatically, and the price of that has gone down dramatically. The third factor is that we now can handle that with machine learning, with large language models.
Christensen: And when we get that, we essentially have a situation where all markets begin to behave as financial markets, that the prices become a lot less sticky. If you talk to businesspeople, they now have access to data that only economists used to have access to. So, they can make decisions that resemble the economic textbooks a lot more. They probably did that historically, but now, they actually have the data and the tools to do this.
Christensen: And so, what happened to me is that when I resigned from my banking job back in 2015 to start Markets & Money Advisory, my economic and financial advisory, I had an ambition of not only doing the finance stuff, [but] I also wanted to do more academic stuff. I've been doing that. But I was also interested in sports economics, sports analytics. I've been doing advisory on this side, so to speak, on that and had a fascination on that. And if you look at sports data— and particularly in the US, of course— sports data has now become completely common. You don't see an NBA team making decisions that are very different from what a bank would do. A capital asset pricing model can explain how many three-point shots and two-point shots any given NBA team would take. So, I like that fascination of using the data, and the data that's available.
Christensen: So, what had happened is that when we got ChatGPT out one and a half years ago, I started using it immediately, to begin with, just for the fun of it, but also to do Python coding. And I remember the first day I was playing around with exactly that. I thought, "Okay. I want to estimate-- I want to get ChatGPT to write a code that creates a linear regression on the US stock market explained by different macroeconomic factors.” And I asked ChatGPT to write the code, and it wrote the code in a few seconds, and I took the code and popped it into Python.
Christensen: And I was rather saddened by the fact that it came out showing errors. And I thought, "Okay. I'm not trying to code anything myself. I was just going [through] iterations." And suddenly, after seven, eight hours working on it, it worked. There was a nice graph on my screen, and my model was there, and it was working. And I'm thinking, "My God. This took me eight hours. Using other tools, I could have done it in 15 minutes." But I also knew that something fundamental had happened.
Christensen: So, I started using it much more, and as ChatGPT was improving, suddenly, it was leading me to having very significant productivity gains in terms of doing data analysis [and] econometrics. And I started to write about it and I started using it. And when I had presentations for clients, pension funds, and so forth, they would say to me, "Lars, that graph, that's not macro, but that's not-- What is that?" "Well, that's ChatGPT, or rather, it's Python through ChatGPT.” And they're like, "Can you teach us that?"
Christensen: And so, I ended up being dragged into doing workshops with pension funds and stuff like that, talking about ChatGPT, and doing data analysis from that perspective. Suddenly, everybody was talking about it. Now, I find myself being dragged into the AI revolution because-- I'm not an IT guy, but I do understand economics, and I do understand data. I worked with it. And, obviously, where AI can make a real difference is in finance, in banking, and also in health. But, of course, with my finance background, it's very natural that I have that. So, I have set up an AI advisory company called PAICE, P-A-I-C-E, with a partner, Christian Heiner Schmidt, here in Copenhagen, and we advise primarily financial companies, but, also, other retail companies on, how do you use data to make decisions? What kind of AI tools could be beneficial?
Christensen: It sounds like a bit of a transition, but, to me, it's just the world that has changed, and I'm using my skills, as somebody who's used to work[ing] with data, to work with data on a more micro level. That doesn't mean that I'm not interested in macroeconomics. I do macroeconomics every single day, and I still advise on that. But It's just that the world is changing, and you can say that, in many ways, it feels like we're moving towards this-- It seems like a little bit of a real business cycle model environment that we are moving to, and AI is probably— and we can talk about that— is making prices a lot more flexible, or at least has the potential to do that.
Beckworth: So, you are riding the AI wave up top. You're not waiting to be pulled along painfully or gradually. You're a surfer who's staying ahead of the wave, so to speak, the AI wave, which is fun and interesting to see. And there's a lot to unpack [with] what you just said there. Let's first move to the point you made about flexible pricing. So, you have a post that you wrote titled *From Merchants to Quants: The Digital Revolution in Retail,* and you get into dynamic pricing. Maybe explain to us, what is dynamic pricing and what are some of the big implications? How will it further connect the world and markets?
The Basics and Implications of Dynamic Pricing
Christensen: Well, dynamic pricing is essentially what we have in our textbooks. It's flexible prices. It's prices reacting to supply and demand. But what, of course, we see in the real world is the menu cost. There is a cost of changing a price, and, therefore, you don't change it all of the time. However, there are periods where there is higher demand, there are supply-side problems, and so forth. Markets should react to that, and prices should change. But there is also some aversion towards price changes. We have known the discussion in the US over dynamic pricing or search pricing at Wendy's. That, of course, has highlighted that discussion.
Christensen: But, fundamentally, the reason that prices in the supermarket [are] not changing in the same way [that] stock prices are changing is that there are costs of changing that. And so, f you walk into a supermarket in the US, in most supermarkets, there will be a paper or cardboard price tag, and somebody will come around changing it from time to time; probably relatively frequently, but not a lot, two or three times a day. However, if you go into a Scandinavian supermarket, you would see that— I would guess that half of all supermarkets in Denmark or Sweden have electronic price tags. So, the major supermarkets in Denmark, essentially, are able to change the prices from their headquarters. So, they can change it whenever they want, and the cost of changing that is very low.
Beckworth: This is like a gas station in America. [When] you go to gas stations, you’ve got digital prices. They're changing all of the time. It's just taking that gas pricing model into retail, into the grocery store.
Christensen: Exactly, and that's a very natural thing, as the cost of the whole electronic thing has gone down dramatically. And at the same time, the cost of labor has gone up. So, it’s a very natural thing that we're seeing this. So, this is essentially the same transition we saw in finance 40 years ago, 35 years ago, when we went from open outcry in trading, on trading floors— When I started working on a trading floor in the early 2000s, it was a very noisy place to be. I loved the atmosphere of that. When I resigned from Danske Bank, where I worked for 15 years, in 2015, I was still on the same trading floor, but it had become extremely boring. There was no noise, there was no nothing, and a major reason for that [was] because most of the trading had become electronic. There [were] not a lot of dealers talking to market makers on the phone. There was not a lot of client talk. It was all electronic.
Christensen: Probably, today, 90-95% of all FX trading is algorithmic, in some form. That doesn't mean that it's independent of human decisions, but the actual settlement of the trading is algorithmic. So, that's a result of technological development, and my view is that that is moving further and further through the economy. And so, if you are a major supermarket [and] you have electronic prices, only one thing is stopping you from changing that, and that is plugging in an algorithm, because what they're doing is already doing that. They are tracking their competitors. What do they do? If you talk to retailers in Denmark, for example, they will say, "We're tracking all of our competitors. We're using apps to scrape their prices from their websites, and we send out spies to other shops," and they all say that.
Christensen: And they will change their prices with a rather high frequency. But they haven't plugged in the algorithms yet. But it's not hard to think that you [will] have machine learning that will track all kinds of data. You have seasonalities, you have— Is there a public holiday? Is there a banking holiday? Is it payday? How's the weather? What should the price of ice cream be? Should that be dependent on the weather? Yes, surely. It is likely to become so.
Christensen: And so, in my view, we are only a few years from seeing full dynamic pricing in most supermarkets in Scandinavia. Of course, the technology is there to do it. It's just a question about, who will do it first? And I have this vision of this. Imagine being here in Copenhagen where I am now. I'm walking to my local supermarket while Jerome Powell has a rate announcement in the U.S, and he says, "Forget about it, guys. I said that I would cut interest rates, but I'm not going to do it," surprising the markets. The dollar strengthens, the market break-even inflation rates drop, oil prices drop.
Christensen: But what happens in my supermarket? Well, the price of milk will drop just a little bit in my supermarket in the same way [that] market expectations of inflation will drop, because that is how you would build the algorithm. You were tied up in domestic factors, but, obviously, the Fed determines— [is a] global monetary superpower. You are the guy who coined that phrase, the Fed as a monetary superpower, a global monetary superpower.
Christensen: Imagine that. The Fed [would have] an instantaneous impact on milk prices in a supermarket in Copenhagen, and I don't think we're far away from that, and I think we're actually quite close to that. Then, that, of course, moves us to a situation where a lot larger of a share of prices, in general, will be described by being flexible. And so, menu costs will have just collapsed and the price of information will have gone down. So, this is not fantasy. If we plug these things into an economic model, this is what the economic model would say, and the future is here.
Beckworth: So, that's why you were alluding to a real business cycle model, which is truly a flexible price model, no nominal rigidities, no sticky prices
Christensen: Exactly.
Beckworth: And it's kind of— I don't know if troubling is too strong of a word, but a little unsettling to think that Jay Powell's statements might affect the price of my ice cream or my milk when I go to the grocery store, which means, Lars, that not only will people like you and me be watching FOMC announcements and press conferences, [but] a lot more people will be tuned in to Jay Powell's press conference before they buy their groceries.
Christensen: Yes, and you will get a lot more listeners on your podcast because, suddenly, it becomes very, very clear that the price of milk is determined by Jerome Powell, or, at least, the basket you buy is driven by that. We both agree that inflation is always and everywhere a monetary phenomenon. Now, it will become much more clear to everybody.
Beckworth: Yes, so, that is pretty— I don't know— sobering, to think through all of the implications, that the Fed's reach is going to be even more acutely felt in many places around the world because of dynamic pricing.
Christensen: Yes, and I believe that that is happening.
Beckworth: So, one more point, Lars, on dynamic pricing, so, it both has great potential, [and] maybe some transition cost to this new brave world of dynamic pricing. Something I want to come back to later in the show, but I'll just throw it out there right now, is [that] I completely see your point [that] output prices, product prices will become more flexible. I suspect, however, like wages, some input costs may still be a little more rigid just for other reasons than menu costs, so people— for a number of reasons we can come back to, but I think you can imagine a world where output product prices become flexible, and wages might still be a little bit sticky, and then, what would be the monetary policy implications? But let's come back to that in a minute.
Beckworth: Let's go to another reason why we're talking today and that is some of the Tweets you've had on recently, and you've demonstrated a progression of abilities on Twitter. You showed how to do VARs with ChatGPT. Then, recently, and this is what really triggered this conversation, you showed how to do, with the help of ChatGPT— the new version of it— DSGE models, which I was really surprised to see. So, tell me about that experience.
Using AI for Econometric Analysis
Christensen: Well, it actually came from the discussion about obesity— and it's something that has been on my mind for some time, the obesity problem in the US and the economic implications of that. And I wanted to try to model Gary Becker's old Rational Addiction Model. And I wrote to ChatGPT, "Do you know Gary Becker's Rational Addiction Model?" Can we model rational addiction to calories? If we can model that, how could medicine change that addictive behavior?
Christensen: So, I was interested in modeling that and thinking about that. So, I just asked ChatGPT, "Do you know this?" I said, "Could you do some modeling on this?" And it started to print out equations. I got the consumer maximization behavior, and I'm like, "What the hell is this?" And it kind of worked, and I remember sitting around for a couple of hours trying to find out what the optimal price of anti-obesity medicine would be given the parameters in Gary Becker's Rational Addiction Model. And it kind of happened by chance. From then on, I thought, "Okay." I wasn't totally happy about it. But then, when the newest version of ChatGPT came out a couple of weeks ago, I said, "Let me ask you to do a New Keynesian DSGE model, and I asked it— and you went, actually, further with it than I did, because I asked it just to give me the Python code, because I felt more comfortable [seeing] the Python code, because I wanted to see what was happening under the hood, so to speak.
Beckworth: That's smart. Yes.
Christensen: And that worked well. The interesting thing is, as I was calibrating the model, the model was not converging. It was explosive and some of the shocks were— the model was not stable. I was like, "Okay, I have to find out the parameters,” but there were a lot of parameters, and I wasn't entirely sure what was making it unstable. So, I simply cut out the picture that it has produced, a graph. I just put that directly into ChatGPT and said, "Something is wrong with the model." It replies that, "Yes. I can see it's not stable. It might be because this and this parameter is too much."
Beckworth: Really? Wow.
Christensen: Yes, and so, I was going back and forth, and I often compare ChatGPT to having a good assistant analyst. It makes the exact same mistakes as an assistant analyst, but it's also much better at doing econometrics than you are or I am. And so, you're going back and forth, and I was going [through] iterations on this. Of course, the idea was to build— and this is where it becomes a little bit meta in the sense that I was trying to analyze how the economy will behave in a New Keynesian DSGE model if the Fed failed to understand that natural interest rates have gone up due to a positive productivity shock. So, that was what I was doing. I was playing around and we were going back and forth. And, yes, as you see, it worked quite well. The interesting thing was that when I had built the model, I said to it, "Could you write some paragraphs on the causality in this? What determines what?" And it put that out, and I was then putting that together with my own prose into a blog post. So, yes, it was all connected.
Beckworth: I misread your post. I thought it had done everything. So, I went and ran with it and it did do everything for me. Now, my model, of course, was less precise and had a number of problems, but it was striking and quite surprising to see. I said, "Look, I want to run a standard New Keynesian model. Here are the issues I want to address. Can you give me a model?" And it did. Not only did it give you the model, but it would summarize the equations. It would tell you what they were. You literally did not need to know much about DSGEs to run these things. Then, I said, "Can you run it for me?" Because it gave me the Python code. "Can you run it for me in Python?" It said, "Sure." And it sent back these impulse response functions. My jaw hit the floor.
Christensen: So, that has been the development since I did that first stock price valuation model one and a half years ago in the early version of ChatGPT, to then ChatGPT 4, where I use the GPT called Data Analyst. Data Analyst can do Python stats. So, it will run it within ChatGPT. But now, you don't have to go through that GPT called Data Analyst. It's just where you are, and it can run it. I'm still not entirely certain that it's doing what it's saying it's doing. That's why I was doing the Python thing. And I've been working with ChatGPT intensively for a long time now. And so, you can see that it's progressing so that, okay, it couldn't do that yesterday. Now, it can do that. It's just been amazing.
Beckworth: Now, I think what you're doing is the smart way to do it, is to check ChatGPT, run the Python code yourself. Now, what I did, though, is I ran, as I mentioned, the early versions of these questions I had in the models. Then, I started making it more complicated. Once it started getting a little more complicated— for example, I asked ChatGPT the question, "In what conditions would nominal GDP targeting be better than a simple Taylor Rule, and would it help financial stability?" And it brought up the [financial] accelerator model by Bernanke and Gertler and Gilchrist. But that model started getting a little more complicated. It gave me the Dynare code. It didn't run it, but it gave me Dynare code.
Christensen: Exactly.
Beckworth: So, there are still limits, but still, the fact that it gave me Dynare code, it showed me the equations, showed me what to do. But I guess, Lars, what's even more striking is what you said a few minutes ago, is that you fed a graph back into ChatGPT.
Christensen: Exactly.
Beckworth: And it could see that the model was explosive, that something was wrong, and it was able to figure it out.
Christensen: And I literally just said-- Okay. I just cut the part of my screen out that just came out as Python output and I just plugged it directly in to the prompt and said, "What's wrong?" And it said, "The parameters seem to be wrong. Let's adjust them." It got me a new Python code, and, eventually, it worked.
Beckworth: Yes, so, it's really fascinating. It's not just a simple old-fashioned Google lookup. It's literally thinking through some of these problems with you. That's just mind-blowing. It's really something. And I want to use that as a transition to a related question, and that is, what does this mean for the profession of economics? And I think, in particular, for people going to get a PhD in economics, what do you see as the implications, both for the jobs available, and what skills should they be developing in their PhD program?
The Implications of AI for the Economics Field
Christensen: Well, thinking, because I strongly believe that artificial intelligence has nothing to do with intelligence. It's very advanced statistics. To me, there is a revolution that is completely natural from traditional econometrics and statistics to, let's say, automated statistics, that machine learning is, essentially, trying out different statistical models and choosing the best thing.
Christensen: That's a kind of “wisdom of the crowd” statistics that machine learning is. At least, that's how economists easily can understand machine learning, rather than how data scientists understand it. Large language models follow naturally from that. So, it's natural to work with. But the analysis is only as good as you are as an economist, as a thinker. We have read hundreds and— probably, both of us— thousands of papers on economics through the years, [and] where we'll see from the beginning [is] that this is garbage in, garbage out. They've done all of the econometrics and now claim to have figured out how the world works. That's just a lot of stuff you put in.
Christensen: One of the problems— if you asked ChatGPT to tell you who Lars Christensen, the economist, is, or who the economist David Beckworth is, ChatGPT knows both of us and will give us a surprisingly good description of both of us. But, if you then start to ask about what we think about nominal GDP targeting, it's like, "Oh, yes. These guys favor nominal GDP targeting." Then, when you start to ask further questions, you would see that, then, it ventures into some kind of, I hate the term, mainstream economics, because what large language models are is essentially just a fitted model. It's just an average or median of use.
Christensen: So, if you use it to just ask open questions, you get rather boring, middle-of-the-road answers that seem like something you could read in the Financial Times or The Wall Street Journal from a middle-of-the-road Wall Street economist or World Bank economist. You don't get proper economic analysis. So, to me, this is a fantastic tool to help you think. The picture I use of that, the analogy I use of that, is Iron Man, Stark Industries. This is Stark Industry economics. You are inside of Iron Man, but you are the economist. So, you can do things faster. You can fly. You can shoot. You have this armor, but it's still you. You are the guy thinking. AI does not have a will. It is not creative.
Christensen: So, to me, as a somewhat hyper guy who likes to produce stuff, who likes to write stuff, I just think, "My God." The new version of ChatGPT— you can take out your phone. You can speak to it. And, yesterday, I wrote a post on LinkedIn where I was like, "Okay." I thought, "Okay, I can use ChatGPT on my phone as an intelligent Dictaphone." So, I wanted to write that, and I started to say, "You should write something kind of like blah-blah-blah. Do write this and this." And I said, "Yes, but I need to go to the shower, so you need to continue writing a little bit more." Then, it was continuing, and then it went back, actually, to my DSGE article, and so it had memory of what I had been inputting into it, and out came this piece where I was thinking that I dictated this, but ChatGPT was helping me. The flow of words was there, but it was certainly me. It was not something artificial. It was me, but it had just sped up how fast I could do this.
Christensen: And so, in terms of how we want to educate young new economists, I was never fascinated by the purely technical side of econometrics. To me, econometrics is a tool that we can use to demonstrate things, the same with DSGE models. The reason we use math in economics is to ensure logic. I wrote my master thesis on Austrian Business Cycle Theory, and I got very disillusioned by that, because what I tried to do was to formulate Austrian Business Cycle Theory mathematically, and it couldn't be done. It couldn't be done, because Austrian Business Cycle Theory is not logically consistent. And so, it showed that, but I'm not fascinated by the math itself. It's just a helpful tool. So, in the same way, Python is just a helpful tool if you want to do econometrics, but why should I sit and code when I can just say, "I need a Phillips curve or equation of exchange or uncovered interest rate parity," or, "Maybe you should build this and in this direction."
Christensen: Yesterday, I had been in Sweden doing a presentation. On the way back, I was reading a newspaper article, and I thought, "This was silly. This is a silly article, Danish newspaper article. The problem here is that they don't think consistently logically.” So, I asked ChatGPT on the go, on my phone, to create a small microeconomic model with two sectors in the economy, a fixed stock of capital, and a shock to the price in one of the sectors, and it put it out. Then, I said to it, "Take this model and take this silly article, and write a rebuttal of the article based on the model," and out came the text. And it was like-- obviously, it was not just copy-paste, but was it me, as an economist, thinking? Yes, it was. It was me, but I was sitting, doing it on my phone. I didn't need to get out the laptop. I didn't need to do any coding. I was just doing it on the go.
Beckworth: So, there's still the principle of comparative advantage. We will always have something unique and original to offer. And if anything, this may serve as a way to make you more productive, right?
Christensen: Absolutely.
Beckworth: You have the ideas. You have, "This is what I want to do, but I don't need to spend my time working up the code or plugging in these equations. They'll do it for me." That's a great story. So, you fed it the article. You fed it your ideas for responding and, boom, it pops out. So, I guess, going forward then, what it might mean for the profession is higher expectations. If you look back historically, few people could, at first, use computers. It was expensive, time-consuming, but as time goes on, statistical models get more complicated. Then, DSGE models become more accessible because of computing power. Now, we're at HANK models [which] are very complicated. But this tool is going to make it easy for everyone to do that. Therefore, you, at a minimum, will have to do the more complicated models in your research, but it'll be easier for you.
Christensen: Yes, but I think there is an interesting thing here in this, is that while you're working with these— I think of myself as an illustration of this— is that I'm working as an economist where I'm getting this boost to productivity, and I'm observing this, commenting on what that thing is doing to the world, while, at the same time, being dragged into that world. And from somebody who has been looking [at] the world for 30 years and not participating, you could say— That's not entirely true because, of course, I've been running my own business for nearly 10 years, been a banker, and so forth. So, I've been very much part of the economy. But, suddenly, it's on a micro level. It's not just investment decisions. It's suddenly discussion of, could the price of milk really instantaneously change when the Fed announces its rate decisions?
Christensen: So, my argument has been that that also means that, in terms of our ability as economists, where we are in the job market is changing, because if you go back, I love sports analytics. I'm a great fan of Michael Lewis's book, Moneyball. And if you look at the story of the Oakland A's in the early 2000s, they were essentially doing value investing. They went out and said, "What kind of baseball players are overestimated, overvalued, and which ones are— who are undervalued?" And Michael Lewis describes it perfectly well.
Christensen: But Michael Lewis should write Moneyball 2, how money disappeared or how mispricing disappeared. Because the interesting thing is that if you look at what happened to the pricing of baseball players in the years following the Moneyball book, Paul DePodesta moved from the Oakland A's to the Red Sox. This seems odd, a Dane talking about baseball, which I have no clue about, but I know that story. And I know the story about the pricing of baseball players, that the inefficiencies disappear once they were discovered. So, that is happening more and more.
Christensen: And if you look at the quants in the banks and then in financial markets, the quants came in during the '90s and into the 2000s. If you look at the banks today, there are actually fewer quants in many places. Why? Because markets are efficient. The quants can't contribute there. My point is that, tomorrow, the quants are moving into Costco. The price of milk in Costco will be determined by an algorithm within the next five years. “Yes, call me, I will help you, or I know somebody who will.” That's where we are moving. So, the economist and the skill set of economists will move from government and finance to the broader economy. And I think, that, I'm totally excited about.
Beckworth: That is a perfect segue to the next area that I want to touch on with you, and it relates to AI. What does this all mean for the Federal Reserve? And I want to go two directions with that. First, what you just touched on, the transition of economists from, say, the government sector to the private sector. So, what will the Fed look like? If we have all of this big data, we have AI, will we need a Federal Reserve as big and with as much labor-intensive work as it's currently doing? That's my first question. Then, secondly, time permitting, I'd like to get into a brief discussion on how the Fed could handle the productivity effects of rapid AI growth, [and the] rapid productivity growth that falls out of that.
Daivd: But let's go to the first one. So, the Federal Reserve is the largest employer of PhD economists. It's an important demand engine for people coming out of the pipeline. Getting a PhD in economics has been a very wise investment for many years, in part because the Federal Reserve is there, but also [because] there's lots of jobs in industry as well for PhDs in economics. But what do you think this means? I bring this up because, Milton Friedman, many years ago, talked about a computer doing an algorithm. Now, of course, the algorithm is very simple, just keep money growing at a constant pace. But you can imagine an algorithm, even if it followed a Taylor rule, it could mine the data, it could look for changing parameter relationships, it could do all of that in a second, which now may take economists a lot longer to do. What are your thoughts in terms of where this all lands?
How Will AI Impact the Federal Reserve and its Policymaking?
Christensen: I think that Milton Friedman was right, I think that we were right, and I think, because we were right, the market monetarists— [since] the market reflects all available information, the Fed should just leave it to the markets to decide on monetary policy. Since machine learning and AI is widely used in the financial sector— and soon will be used by supermarket chains around the world— well, there is no need for the Fed to use these algorithms, because the market will still encapsulate all of that. That being said, obviously, we will get much better macroeconomic indicators because of this. We can find new patterns, and there is no doubt in my view that machine learning is a tool that has been underutilized by economists in general. We have stuck to the old econometrics, and, probably, I, among myself, have been very skeptical about machine learning, because machine learning is often a black box. It's good at providing predictions, but it's not very good at telling you why certain things are.
Christensen: As economists, we want to know why. But you can ask yourself, if you conduct monetary policy, do you want to know why? Yes, if inflation is going up and that is what you are targeting, you want to know why, because otherwise, you confuse supply and demand shocks. If you're targeting nominal GDP, it's less important to know why. So, essentially, markets are already reflecting the development in technology in terms of information technology, in terms of large language models, in terms of machine learning. So, in that sense, you could say that the Fed doesn't need all of these employees.
Christensen: I have many friends working in the Fed system, and I have the greatest respect for Fed economists, and it's a wonderful place, all of the districts, to work, undoubtedly, and there have come some very good economists out of that. But if you want to have an institution that educates economists, that's fine. I'm not going to argue with that. But you can also say that one of the things that the Fed system-- There's always been discussions about leaks from within the Fed system, because there's so many economists involved in every FOMC decision.
Christensen: And I have had the view that that's good. We're having that inside information gradually spreading into the economy, so there will be less shocks, because markets will actually tend to adjust faster to the real thinking of that. And obviously, if you have a lot of people talking, writing analysis, I'm sure that if you took all working papers being produced by the Fed, ran them through large language models, [then] you would be able to track the sentiment of FOMC members better than just looking at macroeconomic data.
Beckworth: But back to my question about the Federal Reserve. Will the Federal Reserve, in the future, be much less labor-intensive in terms Fed staff? As I mentioned earlier, they're the biggest employer of economists. And in my mind, I can imagine a world where that changes quite a bit. They still have the FOMC, there are still people at the top. Someone has to take the blame if something bad does happen or be responsible for decisions being made, good and bad, but maybe [there is] far less need for staff economists. Where do you land on that?
Christensen: Imagine that I'm right on dynamic pricing. Then, we're moving towards a world where monetary policy essentially becomes less harmful, even when it's harmful. The cost of monetary policy failure is mostly a result of sticky prices. When I read Fischer Black, he was this complete out-of-the-box financial theorist who suddenly decided to do macroeconomic models. I don't know if you're familiar with his macroeconomic models. From reading it [for] the first time, [I was] intrigued by [it]. What is this? This is really, really weird. He said, "Monetary policy is just endogenous." And I’m this Friedmanite, thinking about the money supply. Now, I'm increasingly thinking that Fischer Black was right, not for the time he was writing about, but [for] the time we are moving towards.
Christensen: Not that monetary policy is endogenous and such, but just that we're living in this world that resembles the world of Eugene Fama. It's not stock prices and bond prices, but it's the price of milk. And so, when you are in that world, monetary policy becomes less harmful. And I'm sure that the Fed will mess up again, but the cost of messing up when prices are much less sticky than they were in the '70s, and there's a lot less inefficiencies, is lower. So, that, in itself, would be an argument [that] you wouldn't need a lot of people to-- Our mutual friend, George Selgin, would say that the reason there are so many people working at the Fed is to clean up after itself, after its own mistakes. And so, if the Fed's ability to make mistakes is reduced by technology, then that maybe removes the argument for the Fed. But again, Milton Friedman would say that there is nothing as permanent as a temporary government program and the Fed is such a thing.
Beckworth: So, you're saying that central banking, in general, will become less consequential, because a world of dynamic pricing, flexible prices, [and] shocks will just have less of an effect?
Christensen: It's a reluctant view I have. It's really a reluctant view I have. But, at the same time, imagine the past 30 years, since the mid-'90s. We moved into the great moderation. Then, we had 2008, and it was a huge shock and this massive monetary policy mistake. Then, you had the lockdowns of 2008, and I, in some ways, look at inflation expectations in the market. I think there was something profoundly important that happened in 1997 when Robert Hetzel wrote his piece in The Wall Street Journal, about TIPS, about inflation-linked bonds. And that gave us a market for inflation expectations. And even though the Fed has never officially targeted that, it became something that everybody could see, and it has had this tremendous impact in terms of creating nominal stability.
Christensen: And had we not had something like that, and had the Fed not taken the actions it did in 2008, even though it had failed dramatically, it would have been much, much worse. We would have been in a 1930s situation. And despite [those] massive lockdowns— and, I believe, a significant monetary policy failure on the easy side in 2021, not in 2020 but in 2021— we returned relatively fast. And I believe that is because markets have become a lot less sticky, prices have become less sticky, and salaries have become a lot less sticky.
Beckworth: Lars, in the time we have left, I want to come to the last question that we touched on earlier, and that is, how will monetary policy deal with rapid productivity gains in a world where AI has taken over and all of these rapid opportunities are there? And, in particular, how do we allow the real gains to be shared widely? How does everyone get to participate and ride the wave of AI? And I'm going to bring up a solution suggested by George Selgin, and I'm going to run it by you and see what you think. So, George Selgin has a book called Less Than Zero, and I think it'd be very fitting for this time, again, should the productivity surge be permanent, should AI really deliver all that we are hoping that it will.
Beckworth: He says this: "Assuming the productivity shock is economy-wide, so it's felt by everyone or most industries, and assuming there's some competitive measures, the best way to share those real gains is through mild, gentle deflation." And he distinguishes that between something like the 1930s, which is a collapse in demand. He would argue that if you keep demand stable, and maybe we still have sticky nominal wages, but you keep nominal wage growth stable, [then] you could still have output prices fall due to these productivity gains, because per unit cost of production will go down. So, if we're paying someone a wage at a firm, and we have, again, a long-term labor contract, which may not happen in this future world, but just for the sake of the example, let's play along. Sticky wages, but the cost of production per unit is falling because of these productivity gains, [so] you could still see output prices fall, and if competition is there, firms compete.
Beckworth: And so, the real gains would be shared widely via lower, not radically lower, but lower prices that— maybe just a small percent over a year, but over time— would accumulate. That's his vision for how you could do that, and I think that could be applied today. Now, some might say, “Well, that sounds wonderful, but has that ever happened?” And I would argue that the postbellum deflation period is one example. [It’s] not a perfect example, but there are periods there where that seems to have been the case. What are your thoughts? Do you think Selgin's solution would be something that we could apply in such a world, or is it more complicated than that?
Deflation as a Response to an AI Driven Productivity Shock
Christensen: First of all, I would say that it's a wonderful book. It's a fast read. It isn't 200 pages, but it's a wonderful book, and it has had a great influence on my own thinking about monetary policy, and it's one of the things that led me to favor nominal GDP targeting, because, essentially, this is nominal GDP targeting where you are targeting nominal GDP growth at the same rate [as] your perception of real GDP growth. More or less, that would bring you to that world. I think that it would be rather dangerous to announce it, going from one to another, because the problem is that our debt is not real, it's nominal.
Christensen: So, debt ratios would go up in a world of mild deflation. That being said, imagine that we had 3% or 4% nominal GDP growth, but you had accelerated real GDP growth due to a productivity shock, or a boost of productivity. Let's say that the US— For the sake of the argument, US trend growth, real trend growth, is 3.5%, that's the 1990s, and we had 4% nominal GDP growth. But then, we’ll end up having 0.5% inflation. Would that be disastrous? Not at all, because we'll still be able to service our debt, because at 4% nominal GDP growth, it wouldn't be a problem. And so, I think that would be my preferred solution.
Christensen: What I would argue, however, is that there's another part of this that is important. It's not what we are targeting only, but also how we conduct monetary policy, what instruments we are making. And in the blog post we talked about, my idea was to analyze what happens if the central bank's perception of the natural interest rate, or the equilibrium interest rate, is lower than the actual natural interest rate, if it takes time to adjust. If I'm indeed right, that AI will boost productivity growth significantly— and the other factors I mentioned, anti-obesity medicine, demographics, and so forth, will help US productivity growth— If the Fed is slow to realize that, then the natural interest rate will increase, while the Fed will maintain the nominal interest rate, the policy rate, at the same time.
Christensen: The only way they can do that is, essentially, by printing more money. So, you have a deflationary impact of higher productivity, but at the same time, you're having an inflationary impact of easier monetary policy. In that scenario, the Fed, if they were focusing on inflation and using the interest rate, they would say, "Oh, we kept the interest rate unchanged, so our monetary conditions haven't changed." Well, both you and I know that that would be monetary easing, because the natural interest rate would be going up, and we would be able to see that in rising stock markets, in rising commodity prices— by the way, what is happening now— and accelerating nominal GDP growth.
Christensen: But you wouldn't see the inflation, because inflation would be kept down by the productivity gains. So, that's exactly what my DSGE model is showing, is that in the calibration I have— and take it with a grain of salt— in the calibration I have, inflation is impacted little by all of these things, because these two factors even each other out. But what you see is an acceleration in nominal GDP growth, in nominal aggregate demand.
Christensen: And you wrote a paper with George about this, I think, back in 2006, where you discussed what happened during the '90s, that what happened during the '90s was [that] we were having a positive productivity shock, and growth [was] picking up. The Greenspan Fed didn't react to that, because they're saying, "Okay, we shouldn't tighten monetary policy, because the pick-up in growth is not driven by demand, it's driven by a positive productivity shock." So, rightly, the Fed didn't tighten monetary policy into that.
Christensen: However, later in that cycle, the natural interest rate had likely moved up, and they had kept interest rates unchanged, and then monetary policy became excessively easy. You could see signs of that, at least, in the US stock market, in commodity prices, and so forth. Towards the end of the '90s and the early 2000s, that's when we got the so-called tech bubble. I'm still not entirely clear, in my head, what really happened there. I'm very cautious when somebody says bubble and can't really tell why that happened. But I do think that there are some elements about what we're talking about here, and that, of course, ties that '90s thing together.
Christensen: It feels like the '90s to me. We are in the '90s with all of the great stuff that happened in the '90s in terms of productivity growth and, actually, inflation being pretty much under control. But towards the end of that, we probably failed to realize that the natural interest rate had gone up. By the way, that was also a period of relatively high immigration into the US. Other elements in the US labor markets were very different. The baby boomers were not leaving the labor market, as they've been doing over the past 15 years. But there was a general— [It was] quite clear that, probably, productivity growth was going up, and that was increasing the natural interest rate, and that led the Fed, probably, to having too easy [of] monetary policy towards the end of the '90s. And I can imagine that becoming a problem one, two years down the road.
Christensen: I don't foresee that as a near-term problem. I think that the Fed, right now, has got it more or less right in [terms of the] calibration of monetary policy, not in the framework. I'm still critical about the framework. But to avoid [making] that mistake, it would be the right thing to switch, in my view, to a 4% nominal GDP growth target. It wouldn't fix it, but that would be better. Then, also, communicate it in terms of market expectations of rates or whatever the tool is, rather than in the absolute level of interest rates.
Beckworth: Okay, well, with that, our time is up. Our guest today has been Lars Christensen. Lars, thank you so much for coming on the program.
Christensen: It has been a pleasure.
Photo Credit: Nordnet Bank