By Larry Marsh, Kansas City Star Midwest Voices columnist 2009
Is the worst over? Is now the right time to jump back into the stock market? Economic models will fail to tell us where our economy goes from here unless they incorporate system dynamics.
Traditionally economic models have consisted of a number of equations that logically explain how the various sectors of the economy interact with one another. The money supply affects interest rates which are inversely related to bond prices. Wages and salaries affect consumer demand which affects output and, ultimately, employment and inflation.
To use these models for forecasting, they must allow past values to predict future values. This means introducing lagged variables into the models. This works fine for describing the past where the lagged values are all known such as when you want to predict just one period (say, a month) ahead. But what if you want to forecast the state of the economy many periods (months) in the future?
The problem is that the future values of the lagged variables are generally unknown. The "solution" that economists have come up with is to use so-called "time series" models which express the value of a variable such as stock prices as a function of its past values. In other words, you get next period's predicted values on the basis of last period's predicted values. This quickly compounds prediction errors. More importantly, the time series models lack much of the intuitive, logical relationships of the traditional economic models.
The real solution to this problem has been largely ignored by economists. Engineers at MIT and elsewhere have developed models that are much better equipped to deal with the interrelationships involved in forecasting. Their "system dynamics" models are basically systems of differential equations with positive and negative feedback loops and time delays. These models allow for both the traditional logical and intuitive economic relationships and an effective way of forecasting their values.
The reason that economists have eschewed system dynamics models is that their statistical properties have been difficult to determine. They could produce biased and inconsistent estimates. They also quickly turn into a rather complex system of equations that become impossible to solve analytically. Fortunately, there are now many efficient methods that can solve these equations numerically instead.
The reason that system dynamics models work well in forecasting is that they extend the traditional economics models to explain all values of all variables within the model except for the value of time itself, which, of course, is already known infinitely in to the future. In statistical lingo, this is to say that all values in a system dynamics model are "endogenous" (determined within the system) except for time which is "exogenous" (determined outside of the system).
Until rather recently, economists also rejected so-called "nonparametric" models that build themselves, such as the old "stepwise regression" models that allow a variable to be tried on for fit before deciding whether to include it or not. These models had also been rejected because they could not be shown to produce unbiased and consistent estimates.
Advances in asymptotic analysis have allowed the determination of statistical properties for nonparametric models when large data samples are available. Consequently, nonparametric methods are becoming popular in economics.
Advanced asymptotic analysis should now be used to validate system dynamics models for forecasting in economics. This has already been demonstrated for some system dynamics models. If there were ever a time when we needed to be able to peer into the crystal ball and foretell our economic future, now would be it. System dynamics models can show the way.
Free system dynamics software is available at Vensim.com. MATLAB has an extension called "Simulink" for creating and running system dynamics models.
. . .
Also see:
Bernanke's KC talk points to deeper problems in business and economic analysis.
Research needed to end economy's boom-bust cycle.
Enhance financial security, cut income tax with tax-deferred savings plan.
Financial crisis exposes deficiency in economic theory.
Without taxes, money would have no value.
Should California be allowed to create its own money?.
Trashing diet for cake and ice cream exposes flaw in economic theory.
Ostrom and Williamson win Nobel Prize in Economics.
Some Wall Street greed can never be satisfied.
. . .
Follow Larry on Twitter.
. . .
Delicious
Digg
Thank you for responding
All your points on robust model development and relevance to forecasting sound like good advice. I think your ‘system dynamics’ insight here is that robust system dynamics models make explicit the balancing loop forces that make economic behavior and models of it well-behaved over time in spite of periodic bursts of vicious positive feedback loops that rear up and dominate the economy during bubbles from time to time.
The challenge with respect to using asymptotic properties is that in economic forecasting models the ‘true population statistic’ cannot ever be known and if it is assumed from the data then one must be careful of data mining. I think your views on robust model design are also crucially relevant here as well.
Regards,
Michael
Questions about System Dynamics efficacy in forecasting
As an avid hobbyist of System Dynamics modeling ( as opposed to an expert ) I
have to ask; what proof do you have that System Dynamics Models “work well in forecasting”.
I ask because I do not think endogenous model design is, in and of itself, a valid basis for saying System Dynamics models “work well in forecasting” . In addition, any model used to forecast the economy will likely have ‘biased and inconsistent’ results so I question whether that is a valid criticism for those that choose to seek alternatives to System Dynamics models for such purposes. Finally, if any and all models used for forecasting the economy are inherently biased and inconsistent, in the sense you use, how then are “advances in asymptotic analysis’ relevant to assessing any model for the purposes you discuss ? I would like to hear your thoughts on these issues.
Regards,
C. Michael Reilly, CFA
Reply to Michael Reilly, CFA
Thank you for your comments. The key to successful forecasting with System Dynamics is to develop a robust model that captures the essential aspects of the interrelationships between the real world variables. If your model is well-developed it will become relatively robust to modest changes on the periphery. Poorly developed models tend to be sensitive to small changes in structure. The reason for this appears to be that in the real world various forces and systems learn to live with one another and to accommodate themselves to one another. If one factor gets out of line the others still hold the group of factors together for their common purpose. This is why a well-developed System Dynamics model will provide good forecasts. If you are getting poor forecasts, then your model might not be sufficiently developed or the system has changed in some fundamental way. It may be missing key variables or key relationships between the variables. The asymptotic properties are in reference to the parameters of the model which are estimated. Does the sample estimate of a particular population statistic converge onto that true population statistic as the sample size increases? This is a different issue from the one about forecasting. It is saying as you get more and more data of the same type does the sample estimate converge onto the population statistic it is trying to estimate. The mean of Y = X + 1 will not converge to the mean of X asymptotically (as n increases), but the mean of W = X + 1/n will converge to the mean of X as the sample size n goes to infinity.