| Over ;Using descriptive and predictive analytics, this work identifies that FICO Score, Cumulative GPA, Estimated Family Contribution, Original Balance, Credit Hours Earned, Credit Hours Failed Flag, Race / Ethnicity and Segment are the features most discriminating between borrowers who will repay and who will not. Interestingly this set of features overlaps the set of features most important for outright student loan default, but are not the same. The work also demonstrates that the distributions of these features tend to be non-stationary, as they seem to depend on the time period and collection agency from which the data was collected.;These highly discriminating features, along with the rest in the set, were used in the predictive and prescriptive portion of the work in order to generate classifiers. The best possible classifier was a boosted set of decision trees, yielding gains of 34.7% at the 10% sample level and 63.7% at the 25% sample level. The model identified ;The last portion of the work addresses the fact that the number of times a borrower is successfully contacted by a collection agency is highly related to the amount he repays. The importance of discriminating features was compared across the spectrum of the number of contacts in order to generate rules that would indicate borrowers likely to answer a collection agency's call and also repay. This effort turned out largely fruitless as rules were either not generated or insignificant in all of the instances.;Finally, various opportunities for future work are considered. Particularly, experiments into investigating the correlation between personality traits and the features examined here are posed. Personality traits are likely the driver for many of these features and their knowledge would probably strongly improve results. |
