Resumen
Decomposition regression incorporating contextual factors seems to be a natural choice for exploiting both reliability of statistical forecasting and flexibility of judgmental forecasting using contextual information. However, such a regression model suffers from collinearity due to sporadic variables or dummy variables with few variations in related observation data, leading to poor variable selection and biased parameter estimation with conventional least square estimators. In the presence of collinearity, ordinary least square (OLS) may not remain optimal and genetic algorithm can be a better alternative. In this study, we employ a log-linear regression model, incorporating promotional factors, estimated by ordinary least square and genetic algorithm as well, in which, mean absolute percentage error (MAPE), is employed instead of mean square error (MSE), in the objective function to minimize the influence of outliers and parameters to be estimated are set with practical constraints to reflect the actual world more realistically. Empirical results show that in such cases, genetic algorithm may outperform ordinary least square and ARIMA in variable selection, parameter estimation, and out of sample forecasting as well as in forecasting performance, consistently and significantly, in the forecasting of weekly unit sales of a consumer packaged product company in Taiwan. Keywords: Genetic Algorithm; Ordinary Least Square; Collinearity; Contextual Information; Sporadic Variables