Likelihood-ratio changepoint features for consumer-behaviour models

Some predictive models for customer value management might benefit from information about certain changes in individual-consumer behaviour. We take changepoint methods as the first step in producing a model-input feature for this purpose. An unusual feature in the application of changepoint methods to consumer data is there are as many streams of data as there are customers. This property is used to help decide whether an individual has changed their behaviour by ordering likelihood-ratio statistics from the changepoint models. Following a review of changepoint methods, the approach is demonstrated on cash machine transactions. Models for the amount, location and time of transaction are used and accounts exhibiting large evidence of change are examined in detail. For the data set used the approach performs sensibly. The worth of likelihood-ratio statistics to rank evidence for change is considered more generally through some of the literature.