Money, policy errors, and the Fed
All roads lead to the Sumner singularity
Bill Nelson notes via his regular Bank Policy Institute email (also on LinkedIn) that the Fed’s latest Monetary Policy Report to Congress included a short discussion of the monetary aggregate M2. According to Bill, this is the first time in 10 years that the biannual report has mentioned M2 or any other monetary aggregate (he also notes that the Fed and Congress skipped the report and testimony in January-February this year).
As noted here last week in our discussion of Ireland, Miran and Roubini (whom Bill links to in passing), Fed Chair Warsh has shown some interest in the role of monetary aggregates as an information variable the Fed should pay attention too. In his accompanying Senate testimony to the Monetary Policy Report, Warsh suggested that if the Fed had paid more attention to monetary aggregates in 2021-22, it might have better anticipated the subsequent pickup in inflation.
Back in 2023, I wrote about how the global acceleration in inflation in 2021-22 was indeed forecast by many analysts in the monetarist tradition, including Tim Congdon and Scott Grannis. Both Tim and Scott directly referenced M2 growth. Although Tim was wrong on the magnitude, he was directionally correct. Josh Hendrickson subsequently made the argument that the magnitude of the acceleration (which his illustrative model under- rather than over-predicted) was a second-order issue relative to the more important demonstration that inflation was indeed forecastable from monetary aggregates, disproving the frequently heard claim that we cannot reliably forecast inflation.
Josh’s broader argument does not depend on a view that changes in monetary aggregates are causal or exogenous or that monetary aggregates (however measured) are even a good model of inflation. If monetary aggregates contain information that predicts the main macro variables of interest to the Fed, that is sufficient reason to pay attention to them, as Warsh suggests. During the pandemic, I referenced the money demand shock as part of my argument that monetary conditions in Australia were too tight, noting that $100 billion in QE looked less accommodative alongside a money demand shock equal to 17% of GDP at the time of writing.
As Scott Sumner would argue, monetary aggregates may or may not be a good indicator of the stance of monetary policy. Scott called out the overly accommodative stance of US monetary policy in September 2021 based on his view of nominal GDP as a sufficient statistic for the effective stance of policy, well before most central banks started their post-pandemic tightening cycles.
The Fed and M2
Bill Nelson reviews the history of the Fed’s interest in monetary aggregates and argues that there are good reasons why the Fed stopped paying attention to M2 in particular. The Humphrey-Hawkins Act of 1978 required the Fed to include the ranges of growth for the monetary and credit aggregates that it considered appropriate for the year ahead, a requirement that was subsequently dropped in 2000.
Nelson quotes Greenspan’s July 1993 Humphrey-Hawkins testimony to the effect that:
The historical relationships between money and income, and between money and the price level have largely broken down, depriving the aggregates of much of their usefulness as guides to policy. At least for the time being, M2 has been downgraded as a reliable indicator of financial conditions in the economy, and no single variable has yet been identified to take its place.
Nelson maintains that there is no reliable relationship between money and economic activity or the Federal Reserve’s balance sheet, although that is partly a function of the mismeasurement of monetary demonstrated by Peter Ireland. More specifically, he suggests that the link between M2 growth and inflation in the pandemic episode was through the traditional Keynesian impact of savings on future consumption and aggregate demand, not through money. This is actually consistent with the money demand story that Scott Grannis and I suggested in relation to the US and Australia, respectively, in real-time during that episode.
Nelson is critical of what he sees as a failure of monetarists to specify the transmission mechanism from money to activity and inflation in a way that is consistent with the current institutional arrangements for monetary policy, which should render money endogenous. As Marcus Nunes argues, Nelson is actually invoking a monetarist view of the transmission mechanism without realising it (I wrote most of this post before I saw Marcus had also responded to Nelson).
Nick Rowe’s bullets
Nelson’s argument in relation to monetary aggregates should actually apply more broadly, at least under a well-functioning and credible inflation targeting regime. The only thing that should forecast inflation in that setting is changes in inflation expectations. The flattening of the Phillips curve since the introduction of inflation targeting highlights this phenomenon. Any residual explanatory power of the Phillips curve for inflation suggests a failure of monetary policy to fully endogenise economic conditions.
But even then, we can still identify a potential role for monetary aggregates and other variables in forecasting monetary policy errors. Nick Rowe and James Yetman had a nice Bank of Canada working paper back in 2000, in which they turn the standard approach to testing the central bank’s policy reaction function on its head. In classic Nick Rowe style, they argue:
The standard approach looks for where the monetary authority’s gun is pointing; our approach looks for the impact of the bullets...deviations of the bullets from the target should be random errors that are unforecastable from the information available to the monetary authority when it pulled the trigger.
Their model seeks to explain the deviation in inflation from target based on a lagged information set which, in their case, included a monetary aggregate. Under a credible inflation targeting regime, the information set should not have any explanatory power for those deviations. Technically, we want to accept the null hypothesis that the parameters of the model are jointly zero. Their approach actually anticipated some of the more recent work on a sufficient statistics approach to macro policy, an approach broadly consistent with the new market monetarism.
Strictly speaking, the Rowe-Yetman model is a joint test of rational expectations; the lag length with which the policy instrument influences the target variable; and the policymaker’s target. Their approach trades off structural identification for robustness to multicollinearity in the information set. But putting aside the issue of lags and transmission length, and assuming the policymaker is not ignoring information, we can still think of this as a test of the inflation target. If we reject the null, then the central bank is not targeting inflation, possibly because it is distracted by some other objective (such as leaning against the wind on asset prices (to take a not entirely random example).
I gave Claude the task of replicating Rowe and Yetman using Australian data to see if deviations in inflation from the RBA’s target are forecastable. To reduce the dangers of p-hacking, I use the same information set as Rowe and Yetman based on equivalent Australian data, including broad money. I use the trimmed mean as the measure of inflation, which the RBA only started to reference in the mid-2000s, but previous approaches to measuring underlying inflation would have arguably yielded similar results. I also replicate the Rowe and Yetman parameter choices.
Given the original paper and the Australian data, Claude wrote the necessary R code to replicate the paper from a single prompt. I got Claude to write the code so I could be sure it was not making up the results or making obvious errors. I’m actually pretty confident it could have done the full replication, including accessing the data, from a single prompt. Partly, I think this is due to Rowe and Yetman’s very elegant and intuitive explanation in the original paper.
