Specification search in nonlinear time-series models using the genetic algorithm

The Genetic Algorithm (GA) is used to estimate dynamic nonlinear time-series models from nonstationary data. Specification search takes place at three different levels: between competing covariates, between different dynamic specifications, and across functional forms. A variation of GA is developed that operates on strings representing functional forms. Although the dimensionality of the specification space is very large, we show that GA succeeds in estimating strings that have straightforward economic interpretations. The nonstationarity of the data gives rise to the problem of spurious fitness in strings obtained by GA. We suggest the use of stationarity tests on the residuals obtained from static versions of dynamic strings to determine whether the underlying relationship is cointegrated. We use data on "Lotto" sales in Israel to illustrate the application of GA. Finally, we compare models estimated by artificial intelligence (GA) with models estimated by conventional specification search.