Modelling Rule- and Experience-Based Expectations Using Neuro-Fuzzy-Systems

Expectations modelling in macroeconomic theory is often done under restrictive assumptions regarding people's ability to learn and the level of their knowledge. Either it is assumed that people do not learn at all, which justifies the use of simple autoregressive forecasting methods, or the model makers believe that the relevant agents know everything about the (long-term) behaviour of the economic system (rational expectations). Neither of these seems realistically to describe what people really do in anticipating future developments when making current decisions. The lack of an adequate expectations model is especially problematic in business cycle theory where expectations play a dominant role in the cyclical behaviour of main macroeconomic indicators. This paper provides a more realistic description of human forecasting behaviour by using neuro-fuzzy-systems to model economic expectations in a simulation environment. Fuzzy-rules allow the expression of vague knowledge, e.g. "IF the money supply is fairly high and the unemployment rate is rather low THEN inflation tends to rise considerably." This approach, then, assumes that people know something about economic dependencies but that they are not informed of the exact formulas. Neuro-methods are able to train on what people mean when they qualify a certain growth rate of money supply in terms such as "fairly high" or "very low." These two techniques are hybridized as a neuro-fuzzy-system called the "Neuro-Fuzzy Expectation Generator (NFEG)." This module is connected to a business cycle simulation model using MAKROMAT-nfx (designed for WinNT 4.0 and Win98). This software allows us to analyze how the NFEG interacts with the economic system when the later is exposed to exogenous shocks. Since traditional forms of expectations modelling are also implemented in the software, interesting comparisons between rule- and experience-based expectations and autoregressive or rational expectations are possible as well.