Identify Potentially fraudulent claims through statistical and data mining technique with a view to
Replace existing manual process by reducing man hours of effort and
Focus on investigation of fewer loans flagged as fraudulent from the model
Machine learning technique Boosting (Adaboost algorithm) was employed for supervised learning primarily to account for the < 1% incidence of scam.
Scoring model was highly discriminatory with a KS of 0.84. Fraudulent claims were assigned high scores in the model.
Scrutiny was limited to the high scorers – suspicious cases. Investigation of only 5% claims succeeded in detecting 90% sham thus reducing human effort by a great extent.
Recommendation on better acquisition and organization of data to save time
Data checks and business rules were also laid out by Smart
Data anomaly included duplicate cases, mismatch between claim type, cause type and loss type
There were < 100 fraudulent claims out of 30K claims – a hurdle for fraudulent pattern recognition and model building
For a client dataset we could identify 92% of potentially fraudulent cases by examining only the 8% case with highest fraud score a lift factor of more than 11. It all cases it suffices to examine less than 20% of all cases (and usually less than 10%) to identify all fraudulent cases, cutting down manual work by a factor of 5 -10