1 of 14

Cutting the Cord: Predicting Customer Churn for a Telecom Company

Brenner Heintz

2 of 14

Why Study Customer Churn?

  • It costs 5x - 25x as much to find a new customer than it does to retain an existing customer
    • Source: Harvard Business Review (link)
  • We can spend $ to save $$$$

3 of 14

27%

of customers left our company

within the month

4 of 14

Data & Methodology

  • We want to predict: Monthly customer “churn”
  • Kaggle dataset, n = 7,033
  • Independent variables:
    • Which services customers had
    • How much they paid per month
    • Tenure
    • Basic demographic info
  • Assumptions:
    • It costs the company $500 to find a new customer
    • It costs the company $100 to retain an existing customer

5 of 14

Setting the Bar

  • Build a model that’s better than doing nothing, or paying to retain everyone

?

-$654,000

$162,000

6 of 14

Our Model

  • Recall
  • Top 3 models moved forward
  • Used GridSearchCV to cross-validate our models and tune our hyperparameters

7 of 14

Results

  • “Money Saved” per model
  • Low probability thresholds were ideal

8 of 14

Results

  • “Money Saved” per model
  • Low probability thresholds were ideal
  • Logistic Regression won out

$272,200

9 of 14

Results

-$654,000

$162,000

$-654,000

$272,200

10 of 14

Future Work

  • Focus efforts on retaining high-revenue customers
  • Create bundling options for customers that cost less, but lock customers in to longer term contracts

11 of 14

Thank You!

​

github.com/athena15

brenner.heintz@gmail.com

12 of 14

Results

Model

Average Recall

Net Gain

Logistic Regression

81%

$272,200

Random Forest

72%

$262,800

Gradient Boosting

68%

$263,600

AdaBoost

74%

$257,200

13 of 14

14 of 14