Aligning marketing and risk
09/06
Peter Constance outlines one approach which illustrates some of the challenges
As the use of analytic tools has increased and become more refined, attention has moved away from the more traditional method of treating risk assessment and marketing as discrete operations, to a more holistic approach where our assessment of risk factors informs our marketing decisions. Indeed, it is true to say that avoidance of risk per se, is not the key issue. What we are seeking to achieve is an alignment of our marketing with one or more of our corporate objectives. This is usually best served by targeting the profitability of potential new business. In this situation, we are not interested in risk in the way that we would be if designing an underwriting scorecard, we are interested in the predicted dollar value of loss as a modifier of income.
To illustrate some of the issues surrounding the balancing of marketing and risk, we will take as a scenario the planning of a direct marketing campaign by a consumer lending company based on a bought in list. This involves the development of a series of models to arrive at a target list, based on an assessment of the likelihood of response, the likelihood of acceptance and a projection of the likely profitability.
The first stage is to develop a propensity to respond model in the usual way from a test marketing exercise, enabling us to give each record in the external data set a p(response).
Profitability is a good overall indicator of the quality of a loan but can be difficult to model accurately. Taking a data set from our existing customers, we can start with the gross interest income plus fees less cost of funds to give a baseline net interest income, and we can attribute administration costs to give us a margin. The first difficulty comes when considering modifiers for early settlement or loss. As the lost income will be greater the earlier in the term that the customer settles (or defaults), we must develop a series of models for each of these factors. A balance therefore needs to be achieved between the level of granularity in time that we want and the practicalities of having too many models. If loan terms extend to say 5 years, then to model at the monthly level would require 60 models for each of early settlement and default. The most appropriate level can only really be found through experiment and measurement, but let’s take quarterly as our start point. Our profitability models will result in us being able to attribute a predicted profitability value to each loan in our internal data set that takes account of the above factors.
As we know, the quality and completeness of data is essential to the development of effective models, and as the external data set is likely to have a limited set of demographic characteristics, our profitability model must be limited to identifying significance among this shared set in order that we can extend the profit projection from out internal to our external data set.
We may chose to pre-qualify the list for acceptance to avoid the unnecessary cost (and bad customer experience) of declining responses from the campaign. Again, however, we will need to limit our model to shared characteristics between the data sets and this will involve developing a specific model rather than using our normal credit risk models (see diagram). One possibility is to use credit agency data to extend the characteristics available to the accept/decline model, but depending on the numbers involved, this can be a costly exercise.
Having now screened our list for the likelihood of acceptance, we are in a position to attribute a predicted profitability to the remaining records in the external data set. We can then use this in one of two ways. We can set a threshold target profit and market all prospects where:
(p(response) * profitability of responder) – cost > target)
where cost = the unit cost of marketing.
Alternatively, we may prioritise the list according to profitability and set a cut-off based on the number of customers we want to market.
Finally, we can use the models to feed our predictions back into our database. This will allow us to make and publish direct comparisons between expected and actual performance. These predictions can be used to monitor how effective our campaign has been as customer responses are gathered, applications are underwritten and through the performance of the customers throughout the loan lifecycle. By providing feedback both on the marketing strategy and the implications to profit and risk we are able to build stronger cooperation and collaboration between risk and marketing departments, increasing the chance that future strategies will be aligned.
As always with modelling, there is no substitute for experiment and measurement to refine our models but the approach described above offers a reasonable starting point.
