Eric Siegel’s keynote focused on new and innovative uses of analytics – asking the audience to focus on their most expensive operations, their greatest operational risks, their critical operational decisions.
Predictive analytics, he says, is a business intelligence technology that produces a predictive score for each customer or prospect – something scoring their risk of fraud, propensity to buy etc. These scores are the output of a predictive model which is built from the data you have, the collective experience of your organization – I love this perspective of Eric’s, that predictive analytics let you learn from your organizations’ collective experience. Established business applications of this approach include predicting response in direct marketing, churn in customer retention, propensity to buy in product recommendations etc.
Innovative applications were the main topic of Eric’s talk. Innovation, Steve Jobs once said, differentiates leaders from followers. Eric had 7 such uses and encouraged the audience to think about them, to think beyond the standard uses of predictive analytics:
- Improved text mining
Natural language or text mining, computational linguistics is a growing area. One of the challenges is that of resolving ambiguity – time flies like an arrow, fruit flies like a banana (Groucho Marks) – and analytics can be used to improve the quality of text mining. For instance, mining the unstructured data stored in customer care systems has been used to identify the parts that a repair truck should load up with before heading out in response to the calls received.
- Online ad quality
Google is using analytics to predict the ad bounce rate of ads being bought. This uses the ad text and the landing page and predicts how likely it is that someone who clicks on the ad is likely to bounce off the landing page immediately.
- Selling without bothering
Every time a customer visited a product webpage they would follow up with an automated email to the client about the product and tracked the result. This could help but it could put people off too. In the end it turned out to be an exact tie – the same number bought in the email group as in the non email group. Uplift modeling or net lift modeling would let you differentiate between those who should, and those who should not, get the email.
- Call center and customer service
Another example used the analysis of satisfaction surveys to drive the operational decisions that were being surveyed, in this case emergency dispatch of repair trucks. The use of analytics in this case reduced both dissatisfaction and the dispatch rate. Another similar use case was the BBC’s targeting of those most vulnerable to missing the analog to digital TV switch that is coming
- Reliability modeling
Tracking when vehicles, planes etc will need repair. Using analytics to become more proactive about repair, using reliability modeling to predict when things will need repair and then scheduling proactive maintenance. This might be as immediate as scheduling vehicle maintenance or as long term as predicting solar cell failure in satellites 3 years out. The same approach has been used to also predict air traffic congestion using data from radar networks and the potential for congestion in the Visa processing system (1.6B cards, 6,800 txns/sec, millions of merchants).
- Predicting the success of startups
YouNoodle predicts startup success using data collected from financial data, founders, news feeds etc. This same data is used to generate clusters of startups.
- Anomaly detection
Detecting anomalies is used in fraud detection of all kinds and takes the kind of clustering analysis above and finds those transactions, people etc that are different from the standard clusters and therefore potentially fraudulent or abusive.
Eric closed with a call to action:
- Identify the greatest potential for operational gain in your organization
- Find the operational decisions to drive with scores
- Figure out what to predict
- Find the data you need
Begin, as I like to say, with the decision in mind.