An Application of Regression Smoothing for Forecasting Customer Utility Bills

This paper develops the methodology for regression smoothing models and then applies it to the problem of forecasting utility usage rates for individual customers in a city in Northern California. A class of models which use regression as a framework for making forecasts repeatedly over time is developed. The models permit the basic underlying relationship between the dependent variable and the independent variables to change over time because they exponentially weight data through time. They are appealing because of their computational simplicity, low storage requirements and adaptive characteristics. Using a sample of 53 customers forecasts of water, gas and electricity usage and the monthly utility bill in dollars are prepared using regression smoothing models, moving average regression models and models which average previous usages. Overall, the smoothing models significantly outperform the other models. In addition the smoothing models can be improved by using adaptive procedures which weight recent observations more heavily when there is a change in customer or an observed change in usage pattern.