BIVARIATE NON-PARAMETRIC REGRESSION AND VARYING PARAMETER REGRESSION AS APPLIED TO WEATHER NORMALIZATION (ECONOMICS, ECONOMETRICS, TIME SERIES)

摘要

In this dissertation, two estimation techniques are studied, and applied to the problem of weather normalization. Weather normalization is the name given to the estimation of what electricity sales would have been had weather been "normal" instead of the weather that actually occurred.;In the first chapter an extension of the non-parametric regression technique is made to the bivariate case. Estimation is carried out by dividing the data of the bivariate relationship into a two dimensional grid. One variable is estimated for each grid element with a penalty imposed for lack of smoothness over the surface being examined. The degree to which this penalty is imposed, in relationship to the usual lack of goodness of fit penalty, is determined by the data through the use of a generalized cross-validation procedure. Linearly related variables as well as an autoregressive error are added to create a complete model. An application to the relationship between temperature, humidity, and electricity sales is carried out.;In the second chapter an application of varying parameter regression is done. Estimation is completed with the use of Kalman filtering and the EM algorithm. Parameter variation on weather sensitivity is found to exist in certain samples. Estimation shows gradual movement in the parameter on weather sensitivity during the sample period.;In the third chapter an illustration of the ability of the EM algorithm to fit various simulated models is carried out. Accurate results are found to exist, whereas mispecification by using OLS leads to very poor results.;In the final chapter an illustration of the bivariate nonparametric regression is done to get some idea of how well this technique can identify an unknown non-linear surface.;The first chapter represents an extension of an existing technique to the development of a new one and its application. The second chapter is an interesting application to a rarely applied, but much studied technique. The last two chapters are studies of the reliability of these two techniques.