How does predictive analytics work?

Different types of predictive analytics are applied distinctively in business. Here’s how some frequently used models work. Scorecards-predict customer behavior Scorecards are a…

Different types of predictive analytics are applied distinctively in business. Here’s how some frequently used models work.

Scorecards-predict customer behavior

Scorecards are a popular form of predictive model, used in risk assessment and other areas. A scorecard produces scores that "rank-order" customers according to their likelihood to exhibit a specific behavior, from low likelihood to high. Numerical scores make it easy to set "cutoffs," above and below which you take different actions. To build a scorecard, the analyst compares data snapshots for thousands of customers from "before" and "after" a particular decision—for example, before and after an offer was extended, or a loan was made. The analyst assigns values to relevant pieces of data according to which are most predictive, and this information is used to build the scorecard.

Say you run a collections department, and you want to make sure your collection officers collect the maximum amount per hour. For each account, a collections model could deliver a score that would indicate the relative expected collection amount. Your system could send high-scoring (high expected collection amount) accounts to your best collectors, and sell low-scoring (low expected collection amount) accounts to an outside agency.

Scorecards are often good options for industries that must explain decisions to customers or regulators. Since there’s a clear relationship between the "points" assigned to each piece of data (see graphic), scorecards make it easier for businesses to understand why customers score the way they do. In addition, regulatory and business considerations can be built into the scorecard—controversial or prohibited data can be excluded, for example.

Neural Networks-spot abnormal data patterns fast

What if data relationships are highly complex, but your profitability depends on identifying them quickly and precisely? In environments with heavy transaction volume and abnormal data patterns, like fraud, neural networks may be the answer. When your credit card is used, for instance, the transaction is probably examined by a neural network model for signs of fraud. Neural networks are trained in development on a data set where they "learn" by example. The models initially make a number of random predictions. By assigning "penalties" to incorrect predictions, the neural net is systematically updated so that the models eventually learn to make correct predictions. Neural nets are considered artificial intelligence because their structure is based on the way the human brain processes information.

Neural nets are good at processing large amounts of transaction data at incredible speeds. They are also "trained" to identify what’s known as "non-linear" relationships in the data. In a "linear" data relationship, fraud risk would increase as the transaction amount increases or as the distance of the purchase from home increases. But fraud patterns are not so linear—what if a customer makes a large purchase while on vacation? Neural networks can more easily identify this non-linear situation as non-fraudulent versus models built to predict linear relationships.

Clustering-segment customers into groups with similar behavior

"Show me how to split my customer base using demographic and customer data to create groups with different price sensitivities to our most profitable products." This is one kind of problem tackled by clustering, a type of descriptive model that associates customers with each other relative to a certain dimension.

If you needed to execute a marketing campaign against a universe of customers, for example, one step might be to vary the creative and the offer according to the customers’ needs and desires. Using clustering, you could assign each customer to a particular life stage and buying profile—for instance, young urban parents or cautious retirees. You could then build your messaging and product or service terms to appeal to each of these groups.

As a result, clustering can be used as a precursor to predictive modeling. Once clustering identifies key customer segments, predictive models can help tailor action for each segment.

chevron_leftBlog home

Related posts

Take the next step

Connect with FICO for answers to all your product and solution questions. Interested in becoming a business partner? Contact us to learn more. We look forward to hearing from you.