Can banks learn to be rational in their lending activities? The answer depends on the institutionally bounded constraints to learning. From an evolutionary perspective the functionality (for survival) of "learning to be rational" creates strong incentives for such learning without, however, guaranteeing that each member of the particular economic species actually achieves increased fitness. I investigate this issue for a particular economic species, namely, commrercial banks.

The purpose of this paper is to illustrate the key issues related to learning in an economic model by proposing a new screening model for bank commercial loans that uses the neuro fuzzy technique. The technical modeling aspect is integrally connected in a rigorous way to the key conceptual and theoretical aspects of the capabilities for learning to be rational in a broad but precise sense. This paper also compares the relative predictability of loan default among three methods of prediction--- discriminant analysis, logit type regression, and neuro fuzzy--- based on the real data obtained from one of the banks in Taiwan.The neuro fuzzy model, in contrast with the other two, incorporates recursive learning in a real world, imprecise linguistic environment. The empirical results show that in addition to its better screening ability, the neuro fuzzy model is superior in explaining the relationship among the variables as well. With further modifications,this model could be used by bank regulatory agencies for loan examination and by bank loan officers for loan review. The main theoretical conclusion to draw from this demonstration is that non-linear learning in a vague semantic world is both possible and useful. Therefore the search for alternatives to the full neoclassical rationality and its equivalent under uncertainty---rational expectations--- is a plausible and desirable search, especially when the probability for convergence to a rational expectations equilibrium is low.