Gunjan Thakuria, one of our finest consultants, educates us on Churn Analytics

The landscape and dynamics of the telecommunication industry has changed drastically with so many service providers entering the market. The Indian telecommunications industry is one of the fastest growing in the world and India is projected to become the second largest telecom market globally. According to TRAI, the number of telecom subscribers in the country increased to 562.21 million in December 2009, an increase of 3.5 per cent from 543.20 million in November 2009.With so many service providers fighting it out for the same customer base, there is  lot of focus and attention given to churn reduction and customer retention. The fact that customer acquisition is a very expensive exercise has led to more emphasis given to customer retention strategies.

To determine customer retention strategies, it is very important to determine which are the customers who are most likely to churn, and then device strategies based on that. Logistic regression methodology is extensively used to predict churn. Using logistic regression it can be determined not only who is going to churn, but also what the drivers of churn are. It tries to model the log of odds of churning taking into consideration the various characteristics of the customers. The equation in a logistic regression is as follows:

Log (p/1-p)= B0 +B1 X1 + B2X2 +B3X3

Where,

P is probability of churning

X1, X2 and X3 are the covariates effecting churn

Bo is the intercept and B1, B2 and B3 are the coefficients

The intercept and coefficient values are determined using the maximum likelihood estimation. Before performing a logistic regression, the data set is divided into two parts, training and validation. The model is developed on the training set and the probability model is validated by using the equation on the validation set. For the model to be validated decile analysis, lift chart and confusion matrix has to be checked.

The customers are then grouped according to their propensity of attrition. Once this is done it becomes easier for the CRM team to device strategies for customer retention. Usually different types of customer retention campaigns are done on the different groups of customers.

This methodology of churn reduction has proved to be highly profitable and productive for telecommunication companies in reducing churn and hence increasing profitability.

Advertisements

How TV Ratings Work

January 15, 2010

In light of all the brouhaha about the late night shows in the US, it would be interesting to understand how television ratings really work.

The company Nielsen that conducts statistical research began operations in 1923 to sell engineering performance surveys.  The company soon moved on to market research eventually expanding its repertoire to include a national radio rating survey (yes, those were the days). The company eventually moved on to television rating in the 1950’s and has formulated the analysis pattern used by TV rating agencies across the world now.

The key component of the Nielsen ratings is, of course, statistical sampling. Most of what we refer to as television rating comes from a little box called the daily meter which captures what channel the household logs in to and the  People Meter which captures which members of the household watch the show. The all-inclusive number of households  participating in Nielsen daily meter system each year  25,000 out of a grand total of more than 110 million American households that own a Television Set. In addition, Nielsen also sends out millions of paper diaries during November, February, May and July (called “sweeps” in the industry parlance) for households to enter all viewing decisions during these periods.

A combination of these two data sets is used to create TV ratings. A TV rating of 8.5/12 means that 8.5% of the total households watched that show and that 12% of the households who were watching TV at that time watched that show. The numbers that most networks look at though, are the rating in the 18-49 age group as (presumably) that’s where the big spenders lie.

And how do we measure television ratings in India? The process is very simple. Sample households across India’s 75 largest towns are given the  People Meter to capture their viewing decisions. This sample, of course, comes from the estimated 130  million Indian households that have television. The company that conducts this research in India is an industry body called TAM (Television Audience Measurement)

The key issue in case of both of these ratings is obviously the sampling decision. Are these sample sizes enough? Do they accurately cover all ethnicities and economic sections? In countries with vast economic and cultural diversities like India and US can any sample size except for the absolute population be representative? A key part of both TAM and Nielsen’s strategy, therefore is sample design. And while these ratings may not be perfect, they ARE representative. If there’s one thing our many years in the Analytics Industry has taught us, is that some information-however little-is ALWAYS better than no information.

The problem however is the lack of options. For better or for worse, with minor changes, the Nielsen way of doing things has been the established industry norm largely because the huge size and cost of this endeavour. With media spends on television advertising increasing by the day, and television viewing patterns altering dramatically with the introduction of videos on the internet and the DVR, it is only time before someone comes up with a better algorithm to tell us what programs on television are really being watched by most people.