Next up was a group of Fair Isaac folks introducing and discussing decision optimization using Fair Isaac's Strategy Science approach. In terms of Strategy Science the focus is on choosing an optimal set of decisions across a portfolio when there are conflicting goals and constraints. It is not about optimizing a single decision, but a set (often a very large set). For instance, how do you collect the maximum amount from your bad debts given that you can only call a certain number of people, write up a certain amount and so on.

• Decision Modeling - establish a mathematical model of the elements of the decision
• Optimization and Simulation - multiple goal/constraint optimization
• Interpretation - typically turning the optimal model into a set of rules
• Accelerated Learning - efficient test design and learning

One of the most unique features is "action-effect" modeling which takes account of the reactions to actions you take. This is modeled using a diagram like that shown here where things you know or can predict are input to actions that cause reactions. Having built the model you must establish constraints (like amount of bad debt allowed, risk tolerance etc) and objectives (typically several being balanced). Within this space the simulation engine can try many scenarios to see what works best so that you can manage the trade-offs. One of the best final outcomes is to generate a decision tree that applies the optimal action to each transaction.

Decision modeling is about breaking down a problem into its parts, building the relationships between these parts so that we can verify it matches history and then using those relationships to simulate the future. Key pieces:

1. Inputs. Known information, or the results of known predictive models, that is input to the model
2. Decisions. The possible actions you might take
3. Intermediate Outcomes. Unknown but measurable customer reactions that drive objective like use of a product, payment of bills etc.
4. Objective. What is it that we are trying to optimize

These then go into an influence or action-effect model. How do the inputs affect the decisions, how do the decisions drive the intermediate outcomes and how do these drive our objectives. It is important to focus on those elements that are really going to matter in terms of influence and to measure all the things you might change as a result.

Examples of problems suitable for this include authorization/pay-no pay or marketing/cross-sell. In authorization, for instance, there is income related to rejecting over-limit charges but also attrition risk if customers are rejected too often. This kind of conflict is typical in these models. The team then built a decision model for authorization live in the session! Sadly I cannot reproduce them but it was interesting watching them take input from the floor and build a decision model from it. A "pre-baked" one is in the slides, however. Each link on the diagram requires a model showing how the linked elements influence each other - this means taking historical data to see what relationship these things have had in the past.

Some summary slides are here, including:

• Four steps to Strategy Science
• Anatomy of a Strategy Science project
• How we model decision problems
• Influence Diagram