Hey everyone,
We are in urgent need of a fraud detection expert.
We're trying to predict the fraud risk over time periods (3 month risk, 6 month risk etc.) of each observation. The dataset is very imbalanced with only -2% of observations being fraudulent.
Here are the things we've tried, and the problems we've encountered:
* Logistic regression with over sampling methods (problem: too many false positives)
* Outlier detection with SVM (problem: same as above)
* Survival analysis (problem: all our models has low concordance and R-squared)
* AdaBoost with Decision Trees (problem: low detection rates)
* Principal Component Analysis with all the above methods
We're thinking that a Bayesian or probabilistic methodology could work, or random simulation. Of course, we're open to other ideas.
The first deliverable would be a paid advice session, followed by a completed model w/source code.
Please include some questions about the things we've tried so that we can get an understanding of how you approach problems.
Thanks!
About the recuiterMember since May 20, 2018 Chandradeep G
from Tuamotu, French Polynesia