The artificial macroeconomist

Tentative thoughts on my replacement

I am a macroeconomist. For the purposes of this post, I’ll assume that in the not too distant future an artificial super-intelligence (ASI) will surpass my ability to “do macroeconomics”. That may or may not happen, but it seems likely and I’m going to assume it to be true. What then?

One obvious answer is “I don’t know”. If the ASI is smarter than me, how could I fruitfully speculate on its output? But that’s not much of a blog post, so I’ll bravely push ahead in my speculation. I’ll do that in two parts. First, I’ll think about what AIs seem to be good at. And then I’ll discuss the nature of macroeconomics as we know it.

I can easily multiply two single digit integers. I would struggle to multiply two integers that were each 12 digits, an operation that can be easily done by a pocket calculator. I could play Tic Tac Toe against an ASI a hundred times in a row and never lose a game. (All draws.) I could play the game of Go against AlphaGo, and lose all 100 games. It seems like the advantage of AIs over humans grows sharply as the problems become more complex.

Now let’s think about the nature of macroeconomics. Does macroeconomics get more complex as the economy gets larger? Surprisingly, I will argue the exact opposite. In macro terms, the US economy is simpler than the Australian economy, which is simpler than Iceland’s economy. (By simpler, I mean easier to predict.) To explain why, let’s briefly consider how I visualize the field of macroeconomics.

Is see macro as consisting of three major concepts, with everything else being “mere details”.

1. Changes in the value of money (defined as the medium of account) lead to demand shocks, which actually should be called nominal shocks. Monetary policy largely explains long run trends in nominal variables such as inflation and NGDP. Nominal shocks have relatively little impact on real variables in the long run.

2. Because nominal wages and prices are sticky in the short run, and most debt contracts are in nominal terms, unexpected changes in aggregate nominal spending have significant short run effects on employment, output and the stability of the financial system.

3. In addition to demand shocks, the economy can be impacted by what are usually called “supply shocks”, but which should be called real shocks. These include wars, pandemics, trade disruption, political turmoil, and highly distortionary government policy initiatives.

These three factors explain why large economies tend to be simpler in a macroeconomic sense. Both large and small economies are impacted in roughly the same way by demand shocks, which you can think of as monetary policy mistakes. But due to the law of large numbers, large economies are less affected by supply shocks. If the fishing industry in Iceland collapses due to overfishing, that could have a big impact on their economy. The same is true if a volcanic eruption disrupts travel and tourism.

In contrast, volcanoes and fishing problems have only a trivial impact on the large and diversified US economy. This means that a relatively larger proportion of our economic instability is caused by demand shocks, which makes our economy simpler and more predictable. BTW, if there were a difference in the severity of demand shocks between small and large economies, I’d expect the worst mistakes to be made in small economies, on average, due to less expertise in central banking.

If I’m correct, then macro is very different from the board game example discussed above. At first glance, it seems obvious that ASIs would have their greatest advantage in analyzing large complex economies like the US, just as their advantage over humans in the 20X20 grid of the game of Go is vastly greater than in the 3X3 grid of Tic Tac Toe. But I’m arguing the opposite is more likely to be the case.

In the end, I’ll suggest that ASI may well end up being highly effective in improving the science of macroeconomics. (And yes, it is a science.) But first I’ll suggest an important limitation on what ASI might be expected to achieve in the field of macro.

The three major concepts discussed above would all have been quite familiar to a mid-18th century economist like David Hume. Indeed Milton Friedman once suggested:

As I see it, we have advanced beyond Hume in two respects only; first, we now have a more secure grasp of the quantitative magnitudes involved; second, we have gone one derivative beyond Hume.” (1975, p. 177.)

And even that achievement is less impressive than it might seem. Hume lived in a commodity money world, where it was quite sufficient to think in terms of changes in the price level; there was no need to consider changes in the trend rate of inflation. With the onset of fiat money, everything changed. Once again, here’s Friedman:

Double-digit inflation and double-digit interest rates, not the elegance of theoretical reasoning or the overwhelming persuasiveness of serried masses of statistics massaged through modern computers, explain the rediscovery of money.” (1975, p. 176.)

In my view, the fundamentals of macro are more like Tic Tac Toe than they are like Go. As soon as brilliant minds began thinking about inflation and the business cycle, they quickly latched onto the only plausible framework. Over the years, that framework has been substantially refined and embellished, and there are important disagreements about how to best model the process, or the relative importance of fiscal and monetary shocks. But the essence of the model has not changed. I cannot think of a single 21st century macro paper that can be said to have dramatically altered the field. Like Tic Tac Toe, the basic framework for macro is mostly a solved problem, and has been for nearly 300 years.

To be clear, I’m not just saying that, “I think I’m right.” I’m saying that I cannot imagine an alternative framework that is even slightly plausible. It’s analogous to those tough problems like “Why does consciousness feel this way?” or “Why is there something rather than nothing.” It’s not just that I don’t know why there is something; I cannot even imagine an answer to the question. You say there’s something because of X, Y and Z? Then why are there X, Y and Z?

I cannot imagine a plausible macro that didn’t in some way revolve around nominal and real shocks. I suppose you could argue that the Great Depression happened because millions of workers simultaneously decided to take a very long vacation, but does that really seem plausible?

In the real world, debates center around issues such as the question of whether nominal spending collapsed after 1929 because of bad Fed policy, flaws in the international gold standard, or a decline in animal spirits. Those are important debates, but all participants accept the view that NGDP falling in half was an important part of any explanation for the Great Depression. (And please don’t call that a tautology, unless you plan to argue that Zimbabwe’s nominal GDP growth explains its real GDP growth.)

Let’s say I’m correct that the basic AS/AD approach will survive the super-intelligence revolution. What then? How might ASI be able to contribute? I can think of several possible ways. Obviously, there may be a few specific supply side problems that can be addressed by ASI—say pandemic prevention—but I’ll focus on the demand side.

An ASI might be able to devise the optimal monetary policy target, or if policy discretion is best, the optimal instrument setting at a point in time. An ASI might be privy to an enormous amount of “real time” data on the state of the economy, and also might do better than humans at seeing the implications of a wide variety of asset market indicators.

An ASI might be able to tell us whether monetary policy alone is sufficient, or whether stabilization policy also requires assistance from fiscal policy. My own view is that the gains from ASI would be fairly marginal if we adopted my preferred “guardrails” approach to NGDP futures targeting, say with 3% to 5% guardrails. But given that I’m clearly not super-intelligent, I cannot be at all certain on that point.

In other areas, I expect that gains from ASI will be less than might be hoped for. But that’s not because I’m skeptical about ASI, rather it reflects my increasingly downbeat view of the science of macroeconomics. Here are a couple examples:

You might ask the ASI:

Please estimate the fiscal multiplier for changes in taxes and transfers.

Perhaps you hope for a precise figure, like 2.67. But what is the ASI supposed to make of that question? What are you holding constant? Monetary policy? What does it mean to hold monetary policy constant? Stable interest rates? Stable money supply? Stable NGDP? In the latter case, the answer is zero.

And don’t say, “It should be stable interest rates, because that’s what the Fed targets.” No, they target inflation, and they adjust interest rates as necessary to hit their inflation target.

Here’s another example. The PCE price level is up about 8-fold since 1959. So ask the ASI this question:

What is the true rate of inflation since 1959?

Again, the ASI might scratch its head. How are you defining inflation? What should I do with quality changes? How should I think about the inflation rate for “television services”? How about “telephone services”.

If the official government (PCE) inflation figures are correct, my daughter should be indifferent between earning $100,000 today and $12,500 back in 1959. But I don’t even know whether she’d prefer $100,000 today or $100,000 in 1959! She might ask me for some additional information, to make a more informed choice. “So Dad, how much did it cost back in 1959 to have DoorDash deliver a poke bowl to my apartment?” Who’s going to tell her there were no iPhones to order food on, no DoorDash to deliver the food, and no poke bowls even if a restaurant were willing to deliver food.

Your $100,000 salary back then would have meant you were rich, which means you could have called a restaurant with your rotary phone to see if it was open, and then gotten in your “luxury” Cadillac with its plastic seats (a car which in Wisconsin would rust out in 4 or 5 years from road salt) and drive to a “supper club” where you could order bland steak, potatoes and veggies. Or you could stay home and watch I Love Lucy on your little B&W TV set with a fuzzy picture. So which will it be? Do you want $100,000 in 1959 or $100,000 today?

I don’t need an ASI to know how my daughter would answer that question, and even an ASI will never be able to tell us the “true” rate of inflation since 1959.

PS. Why is there something and not nothing? I wonder if an ASI would ask us to define “why”.