Monday, December 27, 2010

ABM+NKDSGE=?

Agent-based models have a track record of generating stock market bubbles when they include agents that are not optimizing and use backward-looking decision rules. But they do not seem to have convinced the profession of their relevance because of the perceived arbitrariness of model components and the fact that they basically predict that a broken clock is right twice a day. Hence, it should be quite interesting to try to embed an agent-based model into a more widely accepted model and see how far this can bring us.

Matthias Lengnik and Hans-Werner Wohltmann do this by including two type of asset traders in a Neo-Keynesian model: fundamentalists, who are forward-looking and expect that price will get closer to the fundamental equilibrium, and chartists, who are backward-looking and obey some predefined rules based on past prices. This introduces some degree of history dependence and assumes that both types of agents are fooled every time. They never learn. And asset prices are thus essentially exogenously determined. The non-financial part of the model follows some old-fashioned model where inflation linearly impacts the output gap, and inflation is determined by the output gap and the evolution of stock prices. In other words, we are back the wind-generating hand-waving of 1980's macro, and not exactly something I would call DSGE.

Anyways, let's see what comes out of this. Of course, by the very nature of the model, there can be multiple equilibria, and an unstable equilibrium is possible. So one has to be very careful with simulations as potentially a lot of scenarios are possible. Yet, Lengnik and Wohltmann base their entire analysis on a single 40 quarter run of their model. They call is "representative." In which sense? Have all runs the same statistical properties? Or did the authors mine for the most convenient one? None of the results can be believed until this is clarified.

5 comments:

Anonymous said...

Rosser will show up any minute now.5, 4, 3, 2, 1,...

Kansan said...

No, I do not think he can defend this paper. It is really bad. He will abstain, and he still has at least some common sense.

Matthias Lengnick said...

Dear Economic Logician ,
thank you for reading our paper and for writing your critique. Of course I do not agree with most of it, so please let me respond to some of your points:

You have written: "[ABC models] do not seem to have convinced the profession of their relevance because of the perceived arbitrariness of model components and the fact that they basically predict that a broken clock is right twice a day."
The behavioral assumptions seam to be quite arbitrary, however they are justified (from the micro perspective) by a number of survey studies. On the other hand, the statistical properties generated by ABC models mimic real data extremely well (macro perspective). Especially properties that are known to be very interesting for complex systems composed of interacting micro units like power law distributions of stock prices (see our paper for references). What exactly are you criticizing with your clock metaphor?

You have written: "... both types of agents are fooled every time."
I am not quite clear what you mean by this. Not everybody could expect right if different forecasting mechanisms are used. That's obvious. But is it more realistic to assume agents are so hyper-rational that everybody knows how everybody else will behave? Only in such a state of perfect ex ante coordination everybody can predict right at the same time. Empirical evidence is quite strong that financial markets do not work this way. See our paper for references.

You have written: "They never learn."
Of course our agents learn. They are continuously evaluating the different kinds of strategies according to past performance. The better a strategy has performed the more likely that it is picked by an agent. These mechanism brings about the complex dynamics that we observe in stock markets. Learning processes like this are common and the ABC financial markets literature (see for example the cited papers of Westerhoff).

You have written: "And asset prices are thus essentially exogenously determined."
Well stock prices are determined by demand/supply. Since the latter is modeled endogenously stock prices are definitely endogenous objects.

Matthias Lengnick said...

You have written: "The non-financial part of the model follows some old-fashioned model ... and not exactly something I would call DSGE."
The New Keynesian Model is one of the most important work horse of contemporary macro and not old fashioned. It is also mainly used in DSGE type models nowadays. http://en.wikipedia.org/wiki/New_Keynesian_economics#New_Keynesian_DSGE_models

You have written: "... there can be multiple equilibria, and an unstable equilibrium is possible."
We did not write anything about multiple or unstable equilibria. We identified an unstable region in the parameter space but that has nothing to do with an unstable equilibrium.

You have written: "So one has to be very careful with simulations as potentially a lot of scenarios are possible."
That's right and we emphasized this ourselves. I suggest reading the passage on page 24 following "We close this section by expressing some warnings concerning the quantitative results of our analysis."

You have written: "Yet, Lengni[c]k and Wohltmann base their entire analysis on a single 40 quarter run of their model. They call i[t] 'representative'."
Our analysis is based partially on algebraically derived results (e.g. pages 15-16) which are independent from any number of runs. Partially it is based on monte carlo experiments. On page 21 in our paper for example it is written: "We run the model for 500 quarters (32,000 days) [...] as well as 1000 different realizations of the pseudo random number generator". We proceeded in a similar way for the impulse response analysis. No analysis is based entirely on one 40 quarter simulation. We only wanted to show the results of one simulation run in order to allow the reader to get a more intuitive understanding of how the results look like.
The name '"representative" run' is of course irony. Agent-based modelers often criticize mainstream models for its flawed way of microfoundation which couples micro and macro in a too direct way. In stead of "growing" macro out of micro both are assumed to be equal by the assumption of a representative individual (or simply summing up heterogeneous individuals). The word "representative" in our paper is an ironical side blow to this line of confrontation between the competing methodological views, indicated by the quotation marks.
For an explanation of the interpretation of the word "growing" in the context of ABC models consult Epstein, J. M. (1999), `Agent-Based Computational Models and Generative Social Science', Complexity 4(5), 41-60.

Matthias Lengnick said...

You have written: "Yet, Lengni[c]k and Wohltmann base their entire analysis on a single 40 quarter run of their model. They call i[t] 'representative'."
Our analysis is based partially on algebraically derived results (e.g. pages 15-16) which are independent from any number of runs. Partially it is based on monte carlo experiments. On page 21 in our paper for example it is written: "We run the model for 500 quarters (32,000 days) [...] as well as 1000 different realizations of the pseudo random number generator". We proceeded in a similar way for the impulse response analysis. No analysis is based entirely on one 40 quarter simulation. We only wanted to show the results of one simulation run in order to allow the reader to get a more intuitive understanding of how the results look like.
The name '"representative" run' is of course irony. Agent-based modelers often criticize mainstream models for its flawed way of microfoundation which couples micro and macro in a too direct way. In stead of "growing" macro out of micro both are assumed to be equal by the assumption of a representative individual (or simply summing up heterogeneous individuals). The word "representative" in our paper is an ironical side blow to this line of confrontation between the competing methodological views, indicated by the quotation marks.
For an explanation of the interpretation of the word "growing" in the context of ABC models consult Epstein, J. M. (1999), `Agent-Based Computational Models and Generative Social Science', Complexity 4(5), 41-60.