This piece is written by someone who has built DSGE models at various central banks. They prefer to write anonymously:
Having read the recently published Macro Hive Deep Dive, Demystifying the DSGE Model, there are a few things to address.
First Things First
The models are only stable when the Taylor rule features a sufficiently strong long-run reaction to deviations of inflation from target. The nominal policy rate must move more than one-for-one with deviations of inflation from target. Here’s another way to think about it: the real interest rate needs to rise when inflation rises and vice versa.
This article is only available to Macro Hive subscribers. Sign-up to receive world-class macro analysis with a daily curated newsletter, podcast, original content from award-winning researchers, cross market strategy, equity insights, trade ideas, crypto flow frameworks, academic paper summaries, explanation and analysis of market-moving events, community investor chat room, and more.
This piece is written by someone who has built DSGE models at various central banks. They prefer to write anonymously:
Having read the recently published Macro Hive Deep Dive, Demystifying the DSGE Model, there are a few things to address.
First Things First
The models are only stable when the Taylor rule features a sufficiently strong long-run reaction to deviations of inflation from target. The nominal policy rate must move more than one-for-one with deviations of inflation from target. Here’s another way to think about it: the real interest rate needs to rise when inflation rises and vice versa.
Habit Formation
Although the article is correct that habit formation is frequently used in the models, this is not to make the policy response smoother. The policy response is made smoother by simply making the Taylor rule a linear combination of the desired policy rate in the absence of smoothing and the lagged policy rate. Typically, the weight on the lagged policy rate is high (>.9). The central bank can easily construct deviations from this in its actual policy scenarios by using anticipated shocks (anticipated by the rest of the agents in the economy, they provide a different result than if the central bank policy path is not anticipated). Anyway, habit formation is included because otherwise you can’t get the volatility of investment to be three times as high as the volatility of consumption, or whatever the stylised fact is in your particular country (it’s about that for the US – or it used to be). Habit formation makes consumers care a lot about fluctuations in consumption and pushes a lot of the shocks to investment.
Shocks
The article is correct about data limitations and lags in data availability. However, many central banks are going to have auxiliary models, nowcasting models, and sectoral experts, which mitigate a lot of these data lags. These outcomes can be pre-input into model scenarios using anticipated or unanticipated shocks as the analysts deem suitable for the situation. Using anticipated shocks does require the economist to take a position not just on what is happening (the likely data outcome) but on why (everything is a function of everything else in these models, and so multiple causal explanations or shocks can be chosen to impose a given outcome).
The article mentions this in other language (“shocks large enough to lead to structural damage”), but specialists would put it in slightly different terms with also slightly broader implications. One of the greatest weaknesses of DSGE models is that they are solved relative to some balanced growth path. Typically, we refer to that balanced growth path as steady state, although technically it is a slight adjustment from a pure steady state to one that exhibits trend exponential growth. The implications of this adjustment are fairly minimal, at least as far as business cycle dynamics are concerned. At any rate, the assumption that the shocks which push the model away from this long-run equilibrium will eventually vanish and that the economy will return to that trend growth path is an extremely strong assumption.
Data and Trends
Related to the last point, the data on which these models are estimated are typically detrended. The model itself typically has something to say about trends, like the Smets-Wouters mentioned in the article assumes all real variables grow at the same deterministic rate, linearly in logs (or exponentially in levels). But then typically in the data, this doesn’t hold. Real consumption divided by real output typically isn’t a constant trend. If you smooth it, you get growth, not a flat line.
DSGE and COVID-19
The biggest issue when considering what a model like this has to say about something like coronavirus is that it wasn’t designed to answer that question at all. It doesn’t really have any ability to speak to issues like the economic impact of people being immobile. There’s really only one consumer, or anyway a representative consumer, in the model. The analyst is then left trying to use a combination of the existing demand and supply shocks to try to predict what will happen. It’s a complicated tool, and very good for assessing things like how to move the nominal policy rate to stabilize output and inflation in response to typical business cycle shocks, exchange rate movements, and so on. But used to answer questions like the current crisis, despite its complicated math and the calculation, it is still effectively back of the envelope.
Complications
Another issue is that the models are fairly complicated because you have everything as a function of everything else (that’s the general equilibrium part). Add to that dynamics and it’s already fairly difficult to solve. And you can’t really solve this model properly because it’s too large. While it is possible to do approximate non-linear stochastic solutions to DSGE models, the models for which you can use those solution methods are very simple. For models typically used in policy analysis at major central banks, there is way too much complication (technically speaking: too many state variables) to enable the use of these non-linear methods. Therefore, the model equations are linearized, and that approximate linear dynamic system is solved. While useful again for answering many questions, this has the side effect of removing risk and uncertainty from the model. Increase uncertainty, no effect in the model. There is no uncertainty in the linear model.
The Absence of Financial Markets is a Major Flaw
The latter makes accounting for even basic financial markets difficult. In the wake of the GFC, central banks scrambled to get some type of financial market mechanism in these models. Typically, however, the models lacked even so much as a term structure of interest rates. And without risk and uncertainty and true nonlinearity, there is no motive for it. There is no need for anything except a one period bond unless you really cram it in there by just saying some consumers had utility for holding long term assets. We try to be sort of reasonable about it and do it in a way that we could say this accounts for pension funds, etc., but nonetheless we’re aware that this isn’t really consistent with the whole idea of DSGE models, which is a macroeconomic model with microeconomic foundations. There isn’t actually a microeconomic explanation for why the rate on a 10-year bond in the model is anything other than the product of expected future short rates (the expectations hypothesis).
(The commentary contained in the above article does not constitute an offer or a solicitation, or a recommendation to implement or liquidate an investment or to carry out any other transaction. It should not be used as a basis for any investment decision or other decision. Any investment decision should be based on appropriate professional advice specific to your needs.)