Customer-Churn

Business Background

PowerCo is a major gas and electricity utility who are considering offering a 20% discount to dissuade clients from churning as it'a fair hypothesis that price changes affect customer churn. Before implementing this strategy, it's necessary to test if price sensitivity is a main driver of customer churn. If it is, then a 20% discount could be an effective customer retention strategy.

Objectives

1. Test hypothesis: Is churn driven by the customers' price sensitivity?
2. Build a predictive model for customer churning and evaluate the model
3. Examine the impact of a 20% discount on business

Data Source and Understanding

• Data are provided by PowerCo. More descriptions are included in the 'Data' folder.
• Client data includes their historical and forecast electricity and gas consumption, contract related dates, etc.
• Price data include fixed and variable pricing in off-peak/mid-peak/peak period.
• Churn indicator: whether each customer has churned or not.

Framework and Method

1. Exploratory Data Analysis:

• Apply data visulaisation to gain a holistic understanding of data set (e.g. data statistics and variable distribution).
• Test if price (generally) is correlated with churn.

2. Feature Engineering:

• Create price sensivity features:
  • Price differences across an entire year
  • Max price difference between months
  • Price differences across different time periods (off_peak, peak, mid_peak)
• Transform date features, categorical features and skewed features

3. Modelling:

• Logistic regression
• Random forest
• Predict churn and churn probability
• Evaluation: classification metrics and feature importance

4. Discount Impact:

• Revenue comparison of discount scenario and non-discount scenario
• Explore who should be given a discount to maximize profit
  • Depending on churn
  • Depending on churn and value of customer
  • Depending on churn probability

Results

• Price sensitivity is not a main driver but a weaker contributor.
• The current set of features are not discriminative enough to clearly distinguish between churners and non-churners.
• Giving discounts can be profitable to some level under assumptions.

Churning clients account for 10%:

The correlation between price and churn is low:

The feature importance of price sensitivity features are scattered around:

If we give discounts to clients with churn probability above 0.5, then profit is maximized and larger than not offering discounts:

Recommendations

The generic 20% discount might not be effective. More personalised strategies need to be applied.

• Give different discounts according to churn probability.
• Find the discount level that balances customer retention vs the cost of false positives.