| Abstract: | To accommodate possible parameter changes in time series at times which are not specified in advance, we propose an adaptive procedure for estimating parameters and for forecasting. The mechanism for activating the adaptive procedure is a successively updated change-detection statistic. The statistic has small expected value when no change is present and has large value when change takes place - the larger the change, the larger the statistic. The statistic defines discounting factors which determine how much of the past will be used both for estimating parameters and for forecasting. The change-detection statistic is designed to effect major changes to parameter estimates and to forecasts in a discrete fashion only, as opposed to certain other adaptive procedures that react continuously to perceived fluctuations in data, and so indicate change even when parameters remain fixed. The procedure is illustrated using exponential smoothing and Holt's linear exponential smoothing and is applied to a hydrological series. |
