Raman Mandapaka, Ph.D., is managing director for the Quantitative Analytics and Risk Management practice at Navigant Consulting, Inc. He has extensive experience in mortgage finance, capital markets, data quality and integrity, data analytics, and econometric and financial modeling. His experience includes regulatory compliance, risk management and model validation.
Mandapaka, when speaking at the recent Consumer Credit Summit at Card Forum, drew on his expertise in automated solutions development to help predict the future of credit-scoring model design.
Decisioning models have become highly advanced in the past few years. How much further will the industry go? In other words, what is the future of model design?
Model design will always be evolving. This evolution is a function of data availability, business need (new products/markets) and processing time. Techniques such as machine learning, although they existed before, are now in greater use because of the computing power we can readily access.
As technology advances, what constraints must be considered by model developers and lenders who use models?
The availability of data is a constraint for modelers — the paucity of data depends on the area of specialization. It will take time to analyze and internalize new and unstructured data that is available and being used. Modeling is not done in abstract; in the context under discussion, it serves a business purpose. Questions modelers may be grappling with include: Does the use of alternative data necessarily improve prediction, and what additional insights might alternative data provide us on the creditworthiness of borrowers? New models will present a challenge for the model developers from a model risk/model validation standpoint. Model developers will also be challenged to document new models and design them in a way that is transparent to independent validators.
Where do you see the more advanced models used? Do you think these will be tools that aid in decision-making, marketing or reporting…or some combination thereof?
Advanced models will be used in many areas, including decision-making, marketing and reporting. Their use in regulatory applications (such as credit stress testing or capital requirements) may be limited in the short run as a result of constraints in auditability, independent validation or replication aspects.
How should consumers view this trend? What’s the potential benefit to them?
We are already seeing some of the positive effects caused by the new model developments. For example, new developments permit the scoring of more borrowers. This will, in turn, improve access to credit and spur growth in the economy as a result of higher consumer spending or increased home ownership. This will also allow more targeted marketing.
At the Consumer Credit Summit @ Card Forum, your panel discussed machine learning. Briefly, what is machine learning and why does it matter?
Modeling is essentially the identification and quantification of the relationships between inputs and output in order to either predict or forecast certain results. These relationships can be identified or pre-specified based on an understanding of the phenomena we are trying to model (for example, borrowers’ default probability is understood to be predicted by their credit scores). These postulated relationships can then be modeled and tested using a methodology such as regression analysis. Machine learning is another way of identifying relationships in the data, — using a series of search algorithms. These searches can be conducted in many dimensions. We can also unearth a finer set of relationships that can be captured by business intuition or exploratory analytics. Machine learning models have the potential to provide greater understanding of the relationships within the data and they can also be used with a large volume of data. These algorithmic searches can be computationally intense and time consuming. The industry is nonetheless returning to experimentation with machine learning models as computing power increases exponentially.