What happens when once desirable models begin to show their age?
Not the willowy, glamorous types that prowl high-fashion catwalks. But rather the aging scoring models you use to predict risk and rank-order various consumer segments. Keeping a fresh face on these models can return big dividends, in the form of lower risk, accurate scoring and higher quality customers.
In this post, we provide an overview of custom attributes and present the benefits of overlaying current scoring models with them. We also suggest specific steps communications companies can take to improve the results of an aging or underperforming model.
The beauty of custom attributes
Attributes are highly predictive variables derived from raw data. Custom attributes, like those you’ve created in house or obtained from third parties, can provide deeper insights into specific behaviors, characteristics and trends. Overlaying your scoring model with custom attributes can further optimize its performance and improve lift. Often, the older the model, the greater the potential for improvement.
Seal it with a KS
Identifying and integrating the most predictive attributes can add power to your overlay, including the ability to accurately rank-order consumers. Overlaying also increases the separation of “goods and bads” (referred to as “KS”) for a model within a particular industry or sub-segment. Not surprisingly, the most predictive attributes vary greatly between industries and sub-segments, mainly due to behavioral differences among their populations.
Getting started
The first step in improving an underperforming model is choosing a data partner—one with proven expertise with multivariate statistical methods and models for the communications industry.
Next, you’ll compile an unbiased sample of consumers, a reject inference sample and a list of attributes derived from sources you deem most appropriate. Attributes are usually narrowed to 10 or fewer from the larger list, based on predictiveness
Predefined, custom or do-it-yourself
Your list could include attributes your company has developed over time, or those obtained from other sources, such as Experian Premier AttributesSM (more than 800 predefined consumer-related choices) or Trend ViewSM attributes. Relationship, income/capacity, loan-to-value and other external data may also be overlaid.
Attribute ToolboxTM
Should you choose to design and create your own list of custom attributes, Experian’s Attribute ToolboxTM offers a platform for development and deployment of attributes from multiple sources (customer data or third-party data identified by you).
Testing a rejuvenated model
The revised model is tested on your both your unbiased and reject inference samples to confirm and evaluate any additional lift induced by newly overlaid attributes. After completing your analysis and due diligence, attributes are installed into production.
Initial testing, in a live environment, can be performed for three to twelve months, depending on the segment (prescreen, collections, fraud, non-pay, etc), outcome or behavior your model seeks to predict. This measured, deliberate approach is considered more conservative, compared with turning new attributes on right away.
Depending on the model’s purpose, improvements can be immediate or more tempered. However, the end result of overlaying attributes is usually better accuracy and performance.
Make your model super again
If your scoring model is starting to show its age, consider overlaying it with high-quality predefined or custom attributes. Because in communications, risk prevention is always in vogue.
To learn more about improving your model, contact your Experian representative.
To read other recent posts related to scoring, click here.
