Loss Prediction By Generalized Least Squares

The author presents analysis as to why the general belief by actuaries that development factors obtained by least squares regression are unbiased is false. He goes on to show that what actuaries should really use is a general linear model; one that uses nonstochastic regressors and an error matrix that allows for correlation. He discusses results by Stanard and Murphy regarding the bias inherent in the chain ladder methods, the Bornhuetter-Ferguson method, the Cape Cod method, and the Stanard-Buhlmann approach.