Table of contents for Live from EDM Summit 2008
- Live from the EDM Summit – From Here to Agility
- Hotwire.com Revenue Management
- Rules and Process Management for Insurers
- EDM Summit – Day 2 Begins
- Risk Management at Sun
- Rules in tables, spreadsheets and diagrams
- Predictive Analytics Produces Business Rules That Deliver
- EDM Summit – Day 3 Begins
- Building Blocks of Decision Management
- Integrating Predictive Analytics and BRM to Improve Health Plan Member Experience
- New Approaches to Creating, Simplifying and Visualizing Rules
- Optimizing Customer Lifecycle Management
- EDM Summit – Emerging Trends Panel
Eric Siegel, who is chairing the new Predictive Analytics World show, presented on predictive analytics and business rules. Predictive analytics, says Eric, is a business intelligence technology that products a predictive score for each customer or prospect … and explanations thereof. These scores come from predictive models that are developed across your historical data. This historical data is, at some level, a collective memory for the company and is a core strategic asset. You must learn from this data.
Predictive analytics allows your organization to learn from its collective experience and puts this knowledge to action. Let’s say, for instance, that people who buy life insurance are likely to buy a luxury sedan. This is knowledge discovery but you must decide what to do – you might use this knowledge to drive offers to this group (do cross-sell) but not others (don’t discount).
- In development, modeling takes historical data and produces a model.
- In production this predictive model takes the customer profile and customer behavior and generates a predictive response.
- Deployment takes the predictive response and applies business logic to take a business action.
Predictive analytics is particular effective in the low value, high volume operational decisions (Micro decisions, for instance). Because the actions taken in these decisions are individual and per-customer, predictive models work perfectly to improve them. Some key concepts:
- Predictors are the key building block for models – what characteristics of the customer predict the desired outcome. Generally predictors should be combined to improve the quality of the model, based on the analysis and objectives. The actual weights of different predictors will come out of the analysis.
- Training data is a flat table – extracted from the operational system with one row per customer. This is historical data of what happened in the past. Generally many examples (100s of 1000s) and many potential predictors
- Can’t just memorize training examples and then look up records – too many combinations because customers are unique and because would be generalizing on a set of 1 which is a bad idea.
- Need both positive and negative examples
- Models can be represented as a decision tree and the various nodes in the tree are each business rules. The training data that built the tree can also be used to see how likely different outcomes are.
- Predictive models can augment the business rules being used in a system, especially if the effectiveness of the decision making can be captured and subsequently analyzed.
Key online applications – content selection, retention and product recommendations. Retention, for instance, is much cheaper than acquisition. Improved retention often has a high ROI (NPV of 75%, growth of 12%) and predicting who is at risk of churn and what might prevent them is very effective.
Eric walked through an example of an online dating site where they targeted dating defections. Training set of 57K subscribers with about 20% likely to leave. Predictors included length of membership, how they were acquired, residental location and others. For instance, twice the churn rate seen in people who were trying to chat AND were less than 237 days AND less than 1.85 days since their last failed login. Lots of small discoveries like this one.
Some derived rules are obvious but:
- Just because a rule is obvious does not mean that being able to prove it is not helpful
- Exact thresholds come from models where the obvious rule might have been approximate
- Some “obvious” rules are not true