Making optimal use of all the data available

Fundamental to the production of accurate and robust scorecards and analytical solutions is the existence of appropriate development and test data. The sample required must be large enough to ensure accurate statistical inference, but it must also be as relevant as possible to the situation being treated. There are no advanced statistical techniques that will produce an optimum scoring solution if the development data is too dissimilar to the current population, or is too small to exhibit accurate trends.

Experian Decision Analytics has always advocated a 'recent sample validation’ to ensure that analytical solutions are as representative of future business as possible. Essentially, the 'recent sample validation' exercise compares the profile of the latest credit applications/accounts with that from the scorecard development sample. But this can be problematic when the profiles differ significantly. In order to address these deficiencies in having the right, or sufficient development data, Experian Decision Analytics has developed a new and improved data profiling methodology, which can be used to adjust or augment existing development data.

There are a number of innovative analytical tools that underpin the new profiling process.  These tools include a data imputation technique driven by clustering algorithms that is used to estimate any missing data on the secondary sample.  The profiling algorithm itself is an iterative model-driven process.

This is a powerful analytical process that is applicable in a range of situations.  It can strengthen existing analytics, or open up entirely new development possibilities.  For example:

• A small sample can be supplemented with additional data from a secondary source, adjusted to match the profile of the original.  This can add valuable robustness to any statistical inferences. 
• A scorecard development sample can be brought into line with the current application population, ensuring that the resulting scorecards are as relevant as possible to the current business need.
• Data from an established portfolio can be profiled to match applicants for a new product, allowing ‘custom' models to be built, even in the absence of performance data.

Using the new Experian Decision Analytics profiling methodology can therefore enable more sophisticated solutions to be developed, that are appropriate to the business in situations where there is limited data available

Dr. Paul Russell - Director of Analytics, Experian Decision Analytics EMEAI

Contact us for further discussions about this article

This article orginally featured in e-news Analysis and scoring update 2007. Click here to request a copy.

Top