An algorithm for estimating time-varying commodity price models

Given the current financial crisis, there is renewed interest in modelling how the price of commodities change in the market. Traditionally, such models have assumed constant parameters. However, large and sudden changes in the parameters can also be anticipated due to market shocks. This paper is aimed at addressing this issue. We first describe a bias-variance trade-off in parameter estimation when sudden changes are considered. We then propose a mechanism to achieve a compromise between the observed bias and variance. A key ingredient of this mechanism is to use an estimator having a variable memory length.