Mainstream Economic Forecasting - Too Complex To Be Predictive

The Conference Board of Canada
model includes some 1,250 equations.

Forecasting is a tough business. It is particularly true for economic forecasting. Nowadays, economists constantly miss targets by a mile.

In the previous posting I discussed problems associated with economic forecasting in general. In this posting we will talk about the major intrinsic problem of mainstream economic models - their complexity.

There are a couple of major complications associated with complex mathematical models. The first major problem are the parameters and variables themselves. With thousands of parameters and variables it is almost impossible to monitor or assess them correctly. Typically, economists try to minimize the error on key variables only. The major consequence of this problem - inability to see the whole picture. Every physicist or mathematician knows that it is easy to get lost even in the system of 3-4 equations.

However, the biggest problem is the derivation of these models. In natural sciences many of these equations are derived from the first principles (e.g. Maxwell’s equations, gravitation law) which were tested in countless experiments. Thus, these equations allow creating extremely accurate models that guide ballistic missiles and satellites.

Predictability of other natural science models which are not based on first principles (e.g. forecasting of earthquakes and weather) are much less accurate; however, they typically have enormous databanks for testing and improving their models.

Social science models, however, lack both rigorous first principles and extensive data (it looks like we have tones of economic information, however, i) its amount is not adequate to the complexity of the system, ii) this information is much less accurate and consistent than one typically used in natural science models).

In these circumstances it is possible (and what economists actually do) to develop a model by linearizing a system in some vicinity of an equilibrium state. Thus, such model may have predictive power in some narrow range of parameters. It may run pretty smoothly if the most important parameters change insignificantly and smoothly (in this case you can adjust the model to a new state using the same set of data)- e.g. economy between two recessions.

However, with a system of 1250 equations you have thousands of parameters which almost impossible to monitor or assess correctly. And due to construction of the model, if even one important parameter or variable is changed significantly you have to re-adjust the whole model, because the system can leave the previous equilibrium state and move to a new one. However, the problem is that we don’t have factual data to feed this model at a new state. Particularly, referring to the current situation we never had a credit slump, housing bubble and $147-a-barrel oil simultaneously.

A good example of economic forecasting models is the so-called Fairmodel, a publicly available (http://fairmodel.econ.yale.edu) econometric model created and maintained by Professor Ray C. Fair of Yale University. The Fairmodel is a compilation of multiple regression equations based on historical data collected from 1952 and established relationships among these variables. This model produced accurate predictions for many years, however, all recent forecasts were overly optimistic.

In summary, we can conclude that most of mainstream economic models do not have predictive power in forecasting turning points or unusual market events. What is the reason why they run smoothly under the normal circumstances but failed to predict the current recession and constantly fail to predict the correct dynamic of the recession? They were developed and trained on a completely different (and irrelevant to the current situation) dataset.