Recent Posts

Wednesday, January 20, 2016

DSGE Macro As An "All You Can Eat" Buffet (Part 1)

The ongoing online debate regarding mainstream Dynamic Stochastic General Equilibrium (DSGE) models has again heated up. Professor Simon Wren-Lewis of Oxford discussed whether mainstream macro was eclectic; provoking a response from Professor Lars P. Syll of Malmö University, the latest of which is "'Deep Parameters' And Microfoundations". I am not interested in the entire spectrum of their debate, rather on the question of eclectic models. (This is the first part of a two-part series.)

Axiomatic Nature Of DSGE Macro

In an earlier article, Professor Syll complains about the "axiomatic" nature of mainstream macro; although I am not a fan of DSGE macro, I do not see this as being a problem. The mathematical affectation of DSGE macro is somewhat reasonable given the academic context the theorists work in.

Modern formal mathematics consists of moving from an initial set of assumptions about a mathematical "system," and then seeing what new properties of that system can be derived using the rules of mathematical logic. Therefore, it is not unexpected that papers would be written with a formalism of starting from a set of assumptions, and then moving towards conclusions. You can skip over specifying your starting assumptions -- which is commonly done in math courses for non-mathematicians -- but this is sloppy, and can lead to errors in comprehension.

Instead, the only problem that I see is that much of the mathematics within DSGE macro is there to make the paper "mathy," and not to inform. Unexplained leaps of logic are unfortunately common.

I would note that not all aspects of economics can be treated with mathematical models. For example, economic history should follow the academic standards for historical analysis. However, I am mainly interested in questions that are quantitative in nature, and so my discussion here is only focussed on those parts of the subject.

Types Of Mathematical Economic Models

Returning to Professor Wren-Lewis' discussion of eclectic models, I should note that I roughly in agreement with him. I also think that we need to work with a variety of simple models, none of which are full simulation of an economy. (This is in distinction to the large models that reached their peak of popularity during the 1960s/1970s.) That said, we distinctly disagree about what set of economic models should look at; he generally favours DSGE models, while I think that entire modelling approach should be abandoned. I will turn to the specific problems of DSGE models in the second part of this article.

I see three potential modelling strategies in macro.
  1. Big Models. A large model that attempts to describe the entire economy, with specific sectors modelled (and presumably sectors in foreign economies as well).
  2. Aggregated Models. Smaller models that attempt to capture the broad trends within the entire economy; the business sector is either treated as a single sector, or possibly a small number of sectors (e.g., consumer goods and capital goods). 
  3. Partial Models. Models that attempt to capture dynamics of a small aspect of the economy. Other parts of the economy would not be specified; analysis would be conditional on a forecast for broad economic trends. (For example, we could model the housing market, and have the unemployment rate, wage growth, and interest rates as external ("exogenous") variables.) If you use this model to generate forecasts, those forecasts are conditional upon your forecast for those input variables.
The first strategy (big models) would be very attractive -- if it worked. These models were the focus of research in the 1960s, but they have largely been abandoned. I have not directly worked with such models myself, but I have heard enough horror stories from people who had to work on the plumbing of those models to say that they did not work in practice. I will give a slightly more formal reason for my rejection below.

The second strategy (aggregated models) might be similar to the first type, but the model is working with aggregated behaviour. This makes it harder to align the model data to measured economic data; for example, do we use CPI or PCE to stand in for the price index within a model? The DSGE models at central banks are generally aggregated in this fashion. My preferred alternative is the use of Stock-Flow Consistent (SFC) models (my page on SFC models).

The use of this second class of models is why I doubt the viability of the first class. If we have a model that properly incorporates business sector planning and investment, the model will end up embedding the Kalecki profit equation. What that relationship tells us is that increasing dividends and investment  raises profits. Since the objective of investment is to increase profits, the model should end up with a "positive feedback loop"; pro-cyclical investment drives the business cycle. Profit growth continues until something impairs fixed investment, at which point the cycle goes into reverse.

The propensity of businesses to invest will drive activity within the model (unless government activity is even more volatile); we could call this tendency "animal spirits." Forecasting the path of the model economy entirely turns into forecasting animal spirits. I will assert that forecasting animal spirits is difficult; for a proof, examine business stories at the time of writing (January 2015). (Standard DSGE models avoid the problem of "animal spirits" by almost completely ignoring business investment decisions. This has the side effect of business cycles disappearing from the models; the models can only induce recessions via dubious concepts like "productivity shocks.")

The hard to measure concept of animal spirits is what dooms the big macro models. The uncertainty around animal spirits dwarfs any additional information that we can glean from dis-aggregating sectors.

The final strategy, partial models, are what are mainly looked at in financial market research. These models are easier to estimate, but hugely depend upon the quality of the inputs you supply for the broad macro trends. Moreover, they are unsatisfying to academia, since they can be incoherent -- a set of models for various sectors may generate inconsistent forecasts if you aggregate them.

Eclectic By Necessity

We end up with a need for eclectic models by necessity. If "big models" worked well, they would be all that you need. Unfortunately, they do not, so we are stuck with partial models. They are either aggregated models which are possibly too abstract to fit to real world data (or entirely depend upon non-measurable variables), or partial models that rely on forecasts for broad economic trends. Although the details of my argument may differ from that of Professor Wren-Lewis, we end up with a similar view on the need for many models.

Where we differ is on the acceptability of DSGE models within that mix of models. I will turn to this in the second part. 

(c) Brian Romanchuk 2015


  1. "The hard to measure concept of animal spirits is what dooms the big macro models"

    The way to make your models mimic human behaviour is to invite humans into your models.

    1. Fun fact: aeronautics engineers have linear models describing human behaviour. They need them in order to see if their designs might trigger PIO - pilot-induced oscillation.

      Although I recognize that we can have non-mathematical models, it's hard to get people inside mathematical ones. People refuse to become sets for some reason or another.

    2. There's your problem. Mathematical models. Or spreadsheets as they are otherwise known.

      I can get people inside my models without any trouble. In fact none of them would earn any money without people using them.

    3. I'm an ex-applied mathematician, so I am fond of those models. I realize that there's things that you cannot do with them. But I find it hard to see how to answer quantitative questions in economics - what happens to the unemployment rate if there is fiscal stimulus which is 2% of GDP? - without them. They do not have to be fancy-looking calculations; even a rule-of-thumb has an embedded model behind it.

    4. Brian-

      Maybe I've become more of an economics nihilist over time but my current position is that economists cant even answer this question:

      "what happens to the unemployment rate if there is fiscal stimulus which is 2% of GDP?"

      any more definitively than...."will probably go down by some amount".

      If you think climate modeling is difficult with supercomputers and 100's of equations, thats nothing to the chaos that is the global economy. THere are simply way too many assumptions in these models to be of any predictive use. Everything is contigent on ceterus paribus assumptions which never hold.

      IOW what happens to the UE rate w\ a 2% of GDP deficit increase?

      what happens to bank lending?
      Are there changes to regulation that makes lending and speculating easier or harder?
      Whats the distribution of the deficit increase?
      What will happen to the savings rate?
      Is it on spending or tax side?
      Whats the impact on the trade balance?
      Whats going on in EU, CHina, Canada?
      Whats the UE trend and position prior to the fiscal adjustment? 2% deficit will probably be more impactful on UE if its at 10% than 4%
      What time scale are we talking about? is the fiscal adjustment ongoing or just for 1 year? If its an infrastructure deficit it might take a year just to get the projects rolling.

      Etc. etc.

      Now sure you can make assumptions for every single one of these questions and run the model through a computer to get a projection. but to assume that even one of your assumptions let alone all or most of them will hold is completely misguided and naive IMHO. The other problem is that alomst every variable is a dependent variable.

      Bank Lending rates cant be judged without knowing the future regulatory framework and "animal spirits" culture. the regulatory framework cant be guessed without knowing what the political culture and the relevant elected legislator breakdown (Ds vs Rs). Are there elections coming and when? So you need to try and predict that too. Is a new fed chair going to be appointed? will they be more hawkish? Will they raise rates higher than you expect? Lower? At what pace?

      Sorry, started to go down the rabbit hole again there. Every question you ask needs to have some sort of assumption made for it either explictly or implictly. So what good are these models for answering a question like the one you asked, which is a rather simple question on the surface? Almost zero given the current state of economics knowledge, understanding, and modeling.

    5. I am not hugely optimistic either; you may note the general lack of forecasts in my writing here. I do not particularly mind have relatively huge error bars on forecasts, so long as they are directionally correct. And to be honest, the huge forecast errors happen around recessions. The behaviour of the economy is in a different mode during a recession, so I think any quantitative forecasts are useless.

      I am finishing my latest round of complaints about DSGE models at the moment. I will eventually turn to the constructive question of what we can hope to do. I introduced that in this part, but I could give more details later.


Note: Posts are manually moderated, with a varying delay. Some disappear.

The comment section here is largely dead. My Substack or Twitter are better places to have a conversation.

Given that this is largely a backup way to reach me, I am going to reject posts that annoy me. Please post lengthy essays elsewhere.