Journal Article
Conditional trend analysis (CTA) predicts the number of purchases in a test period by all households that purchase a given number of items in a base period. The underlying model assumes that households’ purchases follow stationary Poisson processes with rate parameters that vary across the households in a market. However, stationarity is often an unrealistic assumption because of marketing variables and seasonal affects. This paper extends CTA to the non-stationary setting and compares the stationary and non-stationary models, falsely assuming stationarity systematically biases forecasts. Although modelling non-stationarity reduces bias, under-prediction, especially of the zero class persists. It is shown that this under-prediction is, in part, a mathematical artefact due to the skewness of the negative binomial distribution. The methodology is applied to scanner data.
