State-Space versus Multiple Regression for Forecasting Urban Water Demand

Abstract

State-space and multiple regression methods were compared with each other and with simple monthly averages for the accuracy of their short-term forecasts of urban water demand. Seven sets of 24 monthly forecasts of water demand were computed. Each set is based on a different 7-year historic period, using a total of 15 years of monthly data. Based on a variety of measures of forecast error, the state-space models exhibited less bias than the other models, whereas the size of a typical forecast error was about the same for state-space and simple monthly averages. Forecast errors showed considerable variability within both state-space and multiple regression. The mean absolute forecast error ranged from 7.4 to 14.8% for multiple regression, and from 3.6 to 13.1% for state-space. For this sample data, the multiple regression model forecasts were least accurate and also had larger biases than the other methods.