Reject inference with nested conditional models based on joint risk and fraud scores
Senior Research & Development Consultant, Veda
(This is a repeat of the presentation that Ross gave recently at the Credit Scoring and Credit Control conference.)
The simplest model of credit application decision-making treats the process as a single decision (accept/reject) based on a single predictor (credit risk score) of a single outcome behaviour (default). Real credit application decision-making is rather more complex, with multiple predictors (e.g. credit risk score, income, fraud risk score) of multiple behaviours and multiple decisions (e.g. accept/reject, limit setting, fraud vetting). Some decisions are made by the applicant rather than the lender (e.g. take-up offer/withdraw) and may be considered outcome behaviours in their own right.
Typically, each of the outcome behaviours (e.g. take-up, default, fraud) contributes to the profitability, so they need to be jointly considered in setting the decision strategies. Furthermore, the relationships between predictors and outcomes are rarely neatly partitioned with one outcome being related to only one predictor. Rather, outcomes may depend on the joint levels of predictors. Consequently, the predictive power of the system can be increased by explicitly modelling the outcomes as joint functions of multiple predictors.
There are also issues of censoring and outcome inference in such systems (e.g. an account can only default if the application is accepted by the lender and taken-up by the applicant). Consequently, the observed outcome behaviours are systematically biased by past decisions and decision strategies. In order to set future lender decision strategies we need to correct for these biases.
This presentation is a case study of application decision-making based on two predictors (a credit risk score and a score predicting the risk of a fraud-like behaviour) and two primary outcome behaviours (default and fraud). The joint relationship of the predictors to the outcomes (including other outcomes such as take-up) is modelled using smooth, nonparametric surfaces. The multiple decisions are treated as sequential, leading to a series of nested conditional probability models. This provides a conceptually straight forward way to generate predictions adjusted for censoring for all the relevant behavioural outcomes as a function of the two predictor scores and allows the financial consequences of proposed strategies to be calculated. The presentation traces the end-to-end development of the method and demonstrates some of the issues that arose along the way.
When: Friday, 27th September 2013, 12:00-2:00pm
Registration: 12:00 - 12:30pm (Please be early as you must be issued a visitor pass.)
Presentation & discussion: 12:30 - 1:15pm
Networking & light lunch: 1:15 - 2:00pm
Where: NAB, Level 24, 500 Bourke St, Melbourne.
Catering: There will be a light lunch after the presentation.
RSVP: Use the RSVP buttons (available until Thursday 26th) to pre-register your attendance.
• If there are no places left you will automatically be placed on a waiting list.
• Please update your registration if you subsequently find you can't attend. This will automatically allocate your place to the next person on the waiting list.
***** REGISTRATION *****
To gain access to the building you will need to be issued a visitor pass from the concierge by showing some form of ID. Please arrive early because every attendee will need to be issued a visitor pass and there is likely to be a queue.
You will be pre-registered from your RSVP information. If your Meetup profile does not contain your full name and organisation or if you are bringing any guests it is vital that you email those details to [masked]
RSVPs will close on Thursday 26th September.