Optimize the Decision-Making Process
One of the grander visions belongs to Fair, Isaac (www.fairisaac.com). From its base in credit scoring, Fair, Isaac has developed what it calls business science: automating, optimizing, coordinating and executing all of an enterprise's business decisions. This is a grand vision indeed.
Fair, Isaac has moved aggressively to support its goal. One major move was to buy fraud detection specialist HNC Software. HNC brought its own predictive modeling technology plus two recent acquisitions: the Blaze rule-based decision engine (reviewed here in July) and the Trajecta OSE optimization system. These products are now offered with Fair, Isaac's own software under the label "Business Science Suite." The components remain largely separate, though the vendor plans to integrate them over time.
OSE, now enhanced and renamed Decision Optimizer, may be the most visionary Business Science component. The word "optimization" is often used loosely to describe any technology that improves business results, but Decision Optimizer meets the formal definition of a system that automatically finds the best combination of decision rules to reach a specified outcome within a defined set of constraints.
Though constrained optimization software is widely available, only a few other vendors -- most notably MarketSwitch -- have adapted it specifically to marketing. Placing optimization at the center of marketing decision making is hugely valuable because it relates decisions that otherwise would be made independently for each promotion or customer. Coordinating these decisions lets an enterprise manage long-term, aggregate objectives such as total risk and return on investment.
Decision Optimizer implementation begins with construction of a mathematical model describing the operations of the business being optimized. Users specify inputs, constraints, constants, available actions and expected action results. Action results often are based on predictive models -- for example, probability of response to a specific offer. These models are built outside of Decision Optimizer, and scores can be imported with customer data or calculated as needed during the optimization process. Users can define basic scoring formulas within the system or call executable models developed in SAS or other Fair, Isaac tools. The need to build a large number of predictive models, along with the complexity of defining the business model itself, typically leads most organizations to rely on outside consultants for help.
Inputs are read from a flat file or Oracle table, which is prepared outside of the system and has one row per customer. Constraints can be global (e.g., total marketing budget) or local (e.g., maximum promotions per customer). Constraints also can have weights that determine how closely the system adheres to them. Similarly, the objective value - the value the system seeks to optimize -- can be the weighted sum of several components. This gives a limited ability to pursue multiple objectives simultaneously.
Users also can specify combinations of actions that are permitted or forbidden, either for all customers or particular segments. This is a powerful feature that can capture real-world requirements.
In another nod to real-world complexity, the system lets users specify a range of variance between the expected and actual values of a data element. It applies this range to create values used during execution. This simulates uncertain events such as changes in interest rates or customer behavior. It is important because it lets the system estimate the range of risks, as well as the expected value, of a given set of decision rules.
The business model is displayed as a diagram showing relations among inputs, intermediate calculations, and results. Users can select an object on the diagram and explore its properties, or manipulate the diagram to change the underlying model.
Once the base model is complete, users can test alternative strategies by varying input data, constraints, objective functions, uncertainty ranges and other elements. The optimization process itself breaks the customer base into clusters of similar records - typically several thousand - and uses conventional linear programming methods to find the optimal solution within each cluster.
Users can set some parameters to determine the amount of processing used to complete an optimization run. Though performance depends on several factors, a single scenario generally runs in 10 to 20 minutes.
The system automatically transfers scenario results to the Essbase multidimensional analysis system. Standard reports show outcome values, results by time period and how many times each action is executed. Users also can compare results of different scenarios. Once the user selects a preferred solution, a production server assigns actions to the actual customer population.
Decision Optimizer is a work in progress. The 4.0 version, released in September, is basically the software that was purchased with HNC, though Fair, Isaac has upgraded the interface and added support for Unix and Sun Solaris servers. The 5.0 release, set for March, is to incorporate features from Fair, Isaac's internally developed optimization tool, now called Strategy Optimizer. These include nonlinear optimization and the ability to display and export assignment rules in a decision tree - a more convenient approach than Decision Optimizer's method and one that integrates easier with other Fair, Isaac products.
Though Trajecta OSE was introduced in 1999, the system had only two production installations at the time of the Fair, Isaac purchase. Fair, Isaac expects to sell Decision Optimizer mainly with custom consulting and prebuilt solutions to specific business problems. Pricing for the software itself is expected to range from $400,000 to $1 million a year for use in a specified decision area.