Can robots replace humans in monetary policy?
This article is an on-site version of our Chris Giles on Central Banks newsletter. Premium subscribers can sign up here to get the newsletter delivered every Tuesday. Standard subscribers can upgrade to Premium here, or explore all FT newsletters
Federal Reserve chair Jay Powell and President-elect Donald Trump agree that running the US central bank is the greatest job in government.
Their reasons differ. Shortly before winning the election Trump was characteristically insulting, saying that the main benefit of being Fed chair was the adulation that comes with the role. “It’s the greatest job in government. You show up to the office once a month and you say, ‘Let’s see, flip a coin’, and everybody talks about you like you’re a God,” he said.
Last week Powell responded, rejecting the coin flipping reference, but otherwise in some agreement (at 7:20 in this video): “I do love the work,” he said. “And it’s a special place to be surrounded by people who are so dedicated and to know that your work really matters for people. It’s a very special honour to do that work.”
Forget the coin flipping, but Trump’s comments do raise an important question. How much monetary policy setting and analysis can be automated?
Robots setting policy
Setting monetary policy rules has a long history because, in principle, monetary policy should be easy — you have a coherent theory, accurate data and by applying one to the other, you get your optimal policy path for interest rates (or money supply if you are in the monetarist tradition).
But these rules have never worked well. The most famous monetary policy algorithm is the Taylor rule which links interest rates to the deviation of inflation from its target and the degree of spare capacity in the economy. The Taylor rule therefore states that if inflation is high and all resources fully used, interest rates should be high. Low rates are needed to stimulate the economy when people or companies are unemployed or inflation is well below target. The underlying theory is a new Keynesian approach that assumes output gaps and inflation deviations can be measured accurately.
New research from the Bank for International Settlements shows how poor the Taylor rule is as at predicting interest rates across most advanced economies.
The authors’ intent in the article is to augment the Taylor rule with better economic theory, current central bank monetary policy doctrine and better data, all with the aim of making monetary policy more responsive to demand shocks rather than supply shocks.
In this world, higher interest rates are the right response to demand-driven inflation because they curtail economic activity and address the underlying problem. But where inflation is caused by a supply shock — for example, an oil price increase — the question of interest rates is moot. Central banks should “look through” the initial price effects if they are temporary; raising rates would have effects too late and too large for the shock, hitting demand after the inflation has subsided and ultimately creating deflationary forces.
There is a nuance for a large supply shock, such as the post-Covid inflation period, where the fear was of second round effects generating a wage-price spiral. In this case, higher rates would be needed to anchor inflation expectations and prevent workers or companies seeking to take advantage of the initial supply shock by seeking to raise profits or real wages. But the empirical result is clear, according to the BIS work. There should be “a more muted policy response to supply than to demand-driven inflation”.
The data work here is crucial. Drawing on academic work (largely from Adam Shapiro at the San Francisco Fed) that attempts to split inflation into demand and supply components, the BIS examined whether central bank policy could be explained much better by a “targeted Taylor rule” that responded strongly to demand-driven inflation and mildly to supply-driven inflation, as shown in the chart below. The BIS study was clear. An asymmetric Taylor rule approach could explain policy well.
Given the results, the natural question not examined by the BIS is whether robots can replace policymakers with an asymmetric policy rule. According to Hyun Song Shin, head of research at the BIS, the answer is “no”.
“Monetary policy strategies in practice are a little more complicated than an asymmetric Taylor rule,” he told me. Why?
First, the data is far from perfect. Methods for splitting inflation into demand- and supply-driven components are far from agreed. The chart above looks reasonably sensible with the latest inflation driven mostly, but not exclusively, by supply. Back in July, I highlighted other research that came to the polar opposite result, especially for Europe. These distinctions are themselves model outputs and subject to error and uncertainty, especially when measured in real time.
Data problems continue with the inflation component. Ideally you need an accurate forecast of inflation rather than a recent measured rate to prevent your rule from being backward looking. An approach that relies on an estimate of the output gap uses hypothetical data that cannot be known with any accuracy.
The theory is not necessarily correct, either, with parameters such as the degree of the implied relationship between inflation and spare capacity. In practice, many events happen outside strict model parameters.
Human judgment and disagreement will be needed for some time, even if we can describe better how central banks operate.
Can robots interpret policy?
If robots cannot easily replace central bankers, can they interpret them as well as, or better than, humans?
The BIS quarterly review also has an interesting article on how best to use large language models in economics.
Rather than regurgitate its findings, I want to highlight some ongoing analysis my colleagues and I have been working on at the FT, led by Joel Suss. We have been using a large language model to interpret central bankers’ speeches on a hawks-dove scale.
The results for the Fed are in the chart below. You can click on the chart and see that each dot represents a speech from a Fed governor and includes a key passage extracted by artificial intelligence. The question here is whether this will put central bank watchers out of business.
There is no doubt that after quite a lot of honing, the model produces excellent results with Fed speeches being judged hawkish when rates were rising (or shortly before) and more dovish as the Fed geared up to cut rates.
Let’s be brutal though. There is a bit of a “no shit” element to the results, with speeches deemed more hawkish when rates were rising and more dovish when they were falling, so there is some question about how much value the model adds. The model can also glean information from across the internet and might, for all we know, be using the Fed Fund rate as an input into its assessment.
But let’s not be curmudgeonly about this. The model is very effective in parsing huge amounts of text with impressive accuracy and enables us to “read” speeches very quickly and extract the valuable information.
Central bankers can make their words as dense and long-winded as they like. We now have tools to extract some signal from long prose.
Is Powell programmable?
My computer programming skills are rather dated, having dabbled in BASIC as a child and Modula-2 as a junior researcher. But it does strike me that there is a simple algorithm that can explain Jay Powell’s policy justifications of late.
Remember in September when the Fed cut rates by half a percentage point and Powell said the large cut was warranted because the US economy was in “good shape” and he wanted that to continue.
In an interview last week, Powell said the Fed could “afford to be a little more cautious” given the health of the economy, which he described as the envy of the world.
Powell does therefore appear to be programmable. The following algorithm applies (with apologies to all proper coders).
10 Choose policy according to taste;
20 Pronounce it is appropriate because the US economy is in good shape;
30 Go to 10
What I’ve been reading and watching
-
Andrew Bailey tells me that, barring surprises, the Bank of England is planning to cut rates four times by the end of next year. There are likely to be surprises, however
-
South Korea’s central bank governor Rhee Chang-yong says Donald Trump’s potential trade policies are more of a concern than the country’s domestic political turmoil
-
India has replaced its hawkish central bank governor Shaktikanta Das with Sanjay Malhotra even though inflation is still a problem
-
China reaches for economic stimulus again
A chart that matters
The Fed prides itself on its data dependence. Not only is this backward-looking, but the data that most influences officials — monthly payroll growth — is horrible.
Monthly US payroll data showed last week that jobs increased by a healthy 227,000 in November. But the average absolute revision in this series by the third month of publication is more than a quarter of that at 57,000.
So, what can we say? The US labour market is somewhere between pretty weak and going gangbusters. In other words, we do not know very much and far too much significance is attached to monthly US payroll numbers.
Recommended newsletters for you
Free lunch — Your guide to the global economic policy debate. Sign up here
The Lex Newsletter — Lex, our investment column, breaks down the week’s key themes, with analysis by award-winning writers. Sign up here
#robots #replace #humans #monetary #policy