Recursive estimation as an aid to exploratory data analysis: An application to market share models

Recursive estimation methods are often adopted in order to allow for parameter changes in a model. They can also be used to explore data in the time domain before embarking upon traditional estimation methods. A case study using the data from Brodie and de Kluyver, whilst only marginally improving upon forecast accuracy, sheds some light on the problem of why explanatory models have difficulty in outperforming ‘naive’ models on some data sets. In contrast to the common assumption that model parameters are constant in time, the paper shows that there is considerable evidence for time-dependent behaviour in the Brodie and de Kluyver data. This is particularly critical for evaluation of forecasting performance, since it occurs predominantly during the holdout period of that study.