Time-series analysis to reveal fraudulent sales (anomaly detection)
Client was a newspaper distribution company, which sold subscriptions
Sales efforts were focused on kiosks – where customers park their car and get a newspaper.
Some owners of these places were cheating, storing the details of their customers directly to cut out the company.
Client had no time to review across thousands of pharmacies
Segmentation of sales points into groups by sales temporal behaviour (i.e. the time that supposed transactions took place)
The first step was to analyse annual, monthly, weekly and daily sales patterns, to establish a baseline.
Automatic statistical analysis flagged and labelled suspicious activity - per sales point - based on aberrant nature of sales time (and other factors).
“We analysed the temporal dynamics of these places (basically the peaks and troughs of sales) and we started to see clumping in the data.”