I am trying to create a model for debt collections. In the past I have used logistic regression to predict pay/no-pay. This has worked well but has a few unfortunate consequences. People are more likely to pay when the outstanding balance is low. Thus, the model tends to push low balances to the good tail.
The real objective, then, is to maximize dollars collected at a particular depth of file. The default strategy is to order accounts by the debt amount which is considered a "dummy" solution. Improving on this dummy solution has been more difficult than I anticipated. The problem is that the dollars collected is contingent on the debt amount!
When I try to model dollars collected, the debt amount completely overpowers everything else in the model. Is there another way to "subtract" the influence of debt amount or otherwise control for it? The desired result is to have the model rank order dollars collected better than simply using the debt amount.
These are things I have tried:
- 2-stage model using Heckman correction predicting pay/no-pay and then dollars collected
- Built separate models predicting pay/no-pay on high, medium and low debt amounts using logistic regression. Then centered each model using their segment odds.
- Linear regression on just the payers predicting dollars paid and applied the model to the entire population.
- Straight logistic regression predicting pay/no-pay