Publications

The 11th Asia-Pacific Knowledge Discovery and Data Mining Conference held a competition where the task was to rank order bank credit card applicants according to their likelihood of subsequently opening a home loan. A financial institution, with the aim of finding better methods for developing cross-sell models, provided the data.

The main reason the particular data set was chosen was the relatively low incidence of the target group in the overall population, or in other words, it was similar to trying to find a needle in a haystack.

47 entries were received from various sources including academia, software vendors, consultancies and in house modelling departments. Each submission included a short description of the methods used.

The reason for this document was my intial query to the competition organiser, Nathaniel Noriel, if a study of the final results would be made, as I would be interested in knowing how a "committee of experts" model would have performed. After the conclusion of the competition, Nathaniel kindly provided me with the submissions for further analysis.

The quick answer to my query was that a straight average of the top 27 entries would have won the competition and a straight average of the top 40 submissions would have come first runner up. The best solution found was an average of 5 submissions that individually used very different methods. This solution would have won the competition by a significant amount.

This document summarises our findings and will hopefully be of interest to others and provide ideas and stimulus to better improve predictive modelling.

Read more.

What's this?