Introduction

Dataset: https://www.kaggle.com/blastchar/telco-customer-churn

Software: Python,SPSS

Dataset Description: 7043 rows, 21 columns, 3 of them are numeric variables, rest of them are categorical varibles

Goal

• Identify the users with high probabilities to churn

• Uncover the factors may lead to the customer churn.

• Provide managerial recommendations to the company to improve performances

Background

Macro Environment Background(PEST)

Economic : Economy increase support development of telecom industry

Social Culture: Social Culture needs telecom companies to grow

Telecom Industry in the US

Providers: Main players in the telecom industry in the US are Verizon, AT&T, T-mobile and Sprint

Subscriber : Postpaid segment accounted for more than 60% of all connection, while prepaid accounted for less than 20% of all connection

Churn Customer: Annual Industry churn rate was 15.9%

Exploration Data Analysis

Categorical Variable

After doing a chi-square test, Phone Service and Gender doesn't have much impact on target.

Customer who churn are not senior citizen, Using check more and using month to month contract.

For Service, customer who are churn, using Fiber optic Internet Service and single lines phone services.

They tend less use other internet services like online backup etc m but they do using streaming TV and streaming movies.

Numerical Variables:

• From boxplot,tenure, monthly charges and total charges do have impact on churn. But based on domian knowleges, total charges equals to monthly charges multiply tenure. So total charges should be exclude in the analysis.

Most people churn with shorter tenure and more monthly charges.

RFM Analysis

Using tenure as frequencies and monthly charges represents monetary.

Since dataset doesn't have information about recency, only use frequency and monetary to segment customer initially

From barplot, tenure is negatively associated with churn. Less tenure with high probabilities in churn. And more monthly charge is associated with high probabilities to churn.

Decision Tree

Result:

Mainly two kind of people are more likely to churn based on decision tree result:

• Month to Month contract using Fiber optic with tenure less than 17.5 month

• Month to Month contract using DSL or no internet service with tenure less than 3.5 and don’t use tech support

Accuracy and Utility:

Accuracy is about 80%, and our model is more accurate in predicting in No churn customer due to our target distribution.

Overfitting:

The lift chart showing that the trend is nearly same between training sample and test sample. No evidence of overfitting in decision tree.

Logistic Regression

Logistic Regression Model Result:

• From Logistic Regression, the relationship between monthly charges and churn are negative, which doesn't fits the result from RFM and Exploration Data Analysis.

• But the extreme relationship between internet service and churn is strange. From boxplot, Internet Service between monthly charge are highly correlated.

• So the relationship between internet service and churn can somehow represent information between monthly charges and churn.

• Still think monthly charges are positively correlated to the churn. The more customer spend in month, the higher probabilities the customer will churn.

Banana Graph and Cum Lift:

• Compared to benchmark, most customers are just slightly better than random guess based on the banana graph.

• Cum Lift graph shows that less churn probabilities customer just slightly better than those who are more likely to churn.

• It depends on telecommunity industry’s customer characteristics.

What customer's demographics of those who are more likely to churn?

Conclusions,Limitations and Future Work

Conclusion:

• People churn with shorter tenure and higher monthly charges.

• Tenure, Monthly charges and Tech Support are important in predicting Churn.

• Churn customer paid more on Fiber Optic and using less services.

• Customer are not sensitive to service, so the companies can focus on the existing customers.

Limitations:

• Imbalanced Target

• Different accuracy in predicting churn customer and non-churn customer

• Time - scale effect, such as seasonality

Future Work:

• Resampling or collecting more data

• Using different cutoff

• Include cost of acquiring a new customer and remaining an existing customer