I was at the IQPC Mobile Prepaid Summit this week in Kuala Lumpur Malaysia. Many of the sessions were focused on different prepaid strategies and two sessions focused on some form of analytics.
Vishal Dubey of Idea Cellular presented an interesting journey of trying to understand why ARPUs and MOUs were declining despite increasing number of subscribers. He told the crowd that one key predictive variable for a person who would churn from the network was his/her time on the VLR. For the un-initiated like me, VLR stands for Visitor Location Register (see wiki link) and time on VLR basically means how often the user has his SIM card connected to the GSM network. In layman’s terms, it basically indicates how often a user has plugged his SIM into the phone and turned on the phone. This type of insight could be specific to one market but is clearly one that Idea Cellular’s data miners arrived at.
Mikko Rontynen from Xtract then presented social network analysis. One interesting case study that he presented was the Xtract system used to predict churn on an Asian carrier with 10-20M subscribers. Specifically he showed a chart with lift on the y-axis and different groups of subscribers on the x-axis. For the top 100,000 subscribers predicted to churn, the Xtract system was able to accurate predict the churners 6x better than the base rate in the population.
After watching these two presentations, I was wondering which model is better: Xtract’s churn prediction model or the churn prediction models developed by the operators? How does the 6x “lift” over base rate rate compare to the churn prediction models created by tools such as SPSS Clementine and SAS Enterprise Miner? It would have been useful if Xtract had presented what the lifts were for the existing churn prediction models at the Asian carrier. Secondly, if you identify a predictive variable such as “time on VLR” how would you integrate this finding with social network models results?
Tags: churn, churn prediction, predictive analysis, xtract
