FICO Decision Optimizer uses decision models to encapsulate the business impact of customer decisions. These models incorporate KPIs, business constraints and goals. They map possible actions to risks and opportunities and manage all the links between these elements. Once you have a model developed, Decision Optimizer runs data, all your customer records or a subset, through the model to determine the optimal actions to these customers – optimizing to maximize or minimize the targets you identified while meeting your constraints. Uses include champion/challenger comparison, what-if analysis, stress testing, exploring the efficient frontier and tuning already deployed rules for customer treatment. Decision Optimizer can be used in Basel Stress testing, early stage collections, credit line management, origination and more.
The process for Decision Optimizer involves going from data to developing a model, optimizing the model to drive scenario selection and then generating a decision strategy. Predictive analytics from FICO Model Builder or from SAS/SPSS/R etc can be fed into the models. Results can be as a set of actions or as a decision tree. Decision tree results can be driven out to set of business rules in FICO Blaze Advisor or exported using the well defined PMML model for Decision Trees. This allows the output to be integrated to various FICO solutions like FICO Origination Manager or FICO TRIAD Customer Manager as well as with the rest of the Decision Management Platform products like FICO Blaze Advisor.
Decision Optimizer is a web-launched client talking to a remote server for the compute power that supports multiple users on the same model. The initial interface includes a number of key elements. For scenario design you can manage:
- Data items including input data items, some of which are also potential decision keys, output metrics and reporting metrics
- The treatments or available actions to be selected from
- Facts true across all customers
- Constraints on the metrics that must or should be met
- Calculations of various types that convert defined input into outputs:
- Lookup tables/charts (values mapped to probabilities)
- PMML models
- SAS regression models
- Java code
Many scenarios can be defined using these elements and these scenarios can be grouped. The core of a scenario is a model of the components and how they interact. This connects the calculations, input data, metrics and treatments/actions. Each calculation has inputs and outputs that are connected by the equation or look up table. At one end of the model are the fixed facts and the account inputs and at the other are the metrics you want to optimize for and the constraints that must be met. One or more layers of calculations link these and the available treatments/actions into a network model: What you know, what you can do and what you care about.
Once scenarios have been executed and the results gather the overall results for different scenarios can be viewed in a grid for the scenarios in the group. All the various output metrics are shown so they can e compared.
For each scenario you can specify that the results must be created as a decision tree (that can be deployed) or as a set of optimal actions for each customer (for a batch process like a mailing campaign). Decision Optimizer can automatically simplify the generated decision trees to make them easy to read and deploy. One particularly nice feature allows the specification of business palatability constraints to make sure the tree will be acceptable e.g. specifying that, all other things being equal, higher credit scores should be more likely to be accepted than not, no matter what seems “optimal.” In addition tree templates can be defined that limit the attributes and bins that can be used also easing deployment and business believability.
Decision Optimizer allows portfolio level optimization of individual customer decisions and allows the results of this optimization to be deployed into a Decision Management System by generating a deployable decision tree that makes near-optimal assignments. This means that individual customer decisions can then be made that are very close to the optimal without the need to do optimization at run time.
Copyright © 2013 http://jtonedm.com James Taylor
(image: decision management tool / shutterstock)