"The role of design in the development of an applied statistical model:
A case study in fraud detection"
Ross Galyer (Senior Research & Development Consultant Baycorp Advantage)
12th September 2006, 2pm
Room 1.01, Institute for Mathematical Sciences,
South Kensington campus,
Imperial College London,
The development of predictive statistical models appears to be a
science, as it is based on standard statistical techniques. However,
there is a great deal of art in selecting and modifying the techniques
to suit the specific context. This is particularly true of fraud
detection models because of the volatility of fraud characteristics
and the low number and proportion of frauds in typical datasets.
This case study concerns the development of a scorecard for predicting
application fraud on a retail lending portfolio. The model was
developed over a short period and successfully implemented into a
high-volume consumer lending environment. This model was designed
paying particular attention to the intended purpose. Another model was
also built at the same time using standard scorecard development
techniques. This comparison model was not implemented. At the time of
implementation both models showed equivalent predictive
power. However, six months later the comparison model was
significantly less predictive than the implemented model. The design
of the implemented model is analysed and the reasoning behind the
design choices is explained.
|