Forecasting the economic cycles based on an extension of the Holt-Winters model. A genetic algorithms approach

A key feature in fitting local polynomials and in using discounted least squares is the notion that the forecast should be "adaptive" in the sense that the low order polynomials used for extrapolations have coefficients that are modified with each observation. When the data exhibit seasonal behavior, several alternatives to ARIMA models exist. The authors focus on a direct extension of Holt's model, due to Winters and often termed as the Holt-Winters model-which is available for nonstationary time series with seasonal components. The key problems in using this model are: the optimal choice of the parameters involved and for the initial steps; the optimal choice of the number of seasonal coefficients (especially when the data are not monthly or weekly recorded). An alternative method is proposed based on a powerful searching technique, genetic algorithms, for optimizing all the start-up parameters. Numerical examples of non-stationary time series with seasonal components complete the paper.