Building Blocks of Decision Management

Michele Edelman of Discover presented on Building Blocks of Decision Management: “Tools to Rule”. Michele spends a lot of time educating people inside Discover and her team use sources like McKinsey to show executives why EDM matters. For instance, a report on top 10 macro-economic trends:

• Centers of economic activity will shift profoundly not just globally but regionally
  So they should find competitive strategies that cannot be duplicated without strong analytic capabilities and focus on customer analysis as a core competency
• Technology connectivity will transform the way we live and interact
  So they must build insight into how customers will use various channels – track and understand them
• Management will go from art to science
  So interesting analysis is useless unless you can deploy that insight to drive business change

All this leads to business reasons for EDM. Discover’s definition of EDM is

“an approach to automating and connecting decisions across the enterprise that is Precise, Consistent, Agile, Performs” (my emphasis)

EDM to Discover is about automating decisions to enhance business performance – approve/decline decisions (their most important decision), marketing decisions, credit limit assignment (second most important) and underwriting decisions and many more. Interestingly Discover focuses on the really tough decisions first not a low-risk, start small approach.   This has worked for them although it is a higher-risk approach. To give a sense of the scale of their work, one decision has 1,200 decision points and hundreds of rules and model attributes are used in the credit limit decision!

The business objectives for Discover’s use of EDM are to:

• enable analytically driven business strategies
• create organizational agility
• ensure quality
• drive decision management across the enterprise.

The decision management group is an “enterprise utility” separate from but works closely with the IT group and the drivers for their business architecture are precision (increasing intricacy of models, delivering robust and accurate data), consistency (processing controls), agility (event-driven processing, adaptive v predictive decisioning, reusability/extensibility) and performance (millions of daily decisions, scalability, business forecasting).

Michele made the point that “coding it right” is not the issue with decisioning – business impact is. Making sure the business can simulate and forecast the business impact of a decision change is critical.

Their data strategy involves pulling a lot of different sources of data together into a robust, reliable platform for delivering data. Initially this was focused on the modeling analysts but the availability of the same data in production as was available to analysts for modeling was critical for ensuring that models could be developed and then pushed into production. As a note, a small data project for Discover is to add a few hundred attributes and a large one is 1,000! The analysts and reporting both have the same timeliness also – overnight updates. For performance reasons they use Teradata on the data side.

Authorization has a separate and more limited view of data for both performance and reliability reasons – authorization cannot be allowed to go down or take too long. Some of the data used for authorization is generated by models that run overnight.

From a tools perspective they try to avoid focusing on specific products – they use two decision engine platforms (Blaze Advisor and Strata), several modeling tools (SAS, Model Builder). They are focused on parallelism and avoiding duplication. A business user environment allows them to manage the rules and models in the engines and they have a single administrative interface for multiple Unix/Mainframe instances. Deployment is handed off to the technical folks but all the rule and model management is handled by the business.

From a results perspective:

• More than 10 decision management applications (3 online, 8 batch)
• More than 100 model scores introduced through these applications
• From portfolio analytics to customer level analytics with much more precise segmentation
• Rapid response and agility, release management to keep data/policy changes synchronized
• Scalability for Blaze Advisor got to 600 TPS for online and 1,000s for batch
• No downtime for the decision engines
• Successfully passed all the audit and compliance tests

Lessons Learned

• Data, data, data – infrastructure and integration are big challenge
• Simulation environment prior to production
• Plan for ongoing expansion
• Properly trained resources are key
• Senior management commitment

Michele made the point that they have not done a formal ROI study, in part because the costs of the whole program pale into insignificance relative to the value created. She had a great phrase “execution latency” and emphasized that in the current climate how valuable the agility and change management that EDM provides. As she said “there is no data on the future”!

Discover is one of those companies that are truly adopting EDM. They have gone from rules-based decisions to statistical models and data mining to account level economic models to customer level economic models but see that there is more to do, more analytics to use and more data.