Learning, Specification Search and Market Efficiency. With an Application to the Danish Stock Market.

A new methodology is adopted for testing semistrong efficiency in financial markets where investors do not know the underlying data-generating model. Based on ideas from the literature on learning, it is shown that investors can use a dynamic significance criterion to conduct a specification search and select a model from which predictions can be computed recursively. Applied to the Danish stock market over the period 1982-91, a portfolio based on such recursive predictions is found to provide a higher mean return with a lower variance than the market index. Copyright 1993 by The editors of the Scandinavian Journal of Economics.