How do we stop the Churn Burn?

Goal:

• Discover driving features affecting churn
• Use drivers to develop a machine learning model to predict churn
• Use these predictions to inform preemptive decisions aimed at alleviating future churn

Acquire

• telco_churn data from Codeup SQL database was used for this project.
• The data was initially pulled on 26-OCT.
• The initial DataFrame contained 7043 records with 44 features
  (44 columns and 7043 rows) before cleaning & preparation.
• Each row represents a customer record both current & historical.
• Each column represents a feature provided by telco or an informational element about the customer.

Prepare

Prepare Actions

• DROP: Removed 4 index_id, 18 duplicate, and 1 corrupted data column
• RENAME: Initially did not need to Rename any original columns
• REFORMAT: 2 columns contained inappropriate data types that needed to be reformatted
• REPLACE: 7 columns had a third value that could be determined by another feature, replaced third value in each column with appropriate yes/no value. 1 column had empty non-null values that were replaced with 0
• ENCODED: 14 categorical columns from variables to boolean numeric values
• MELT: No melts needed
• PIVOT: 3 columns with more than two variables were pivotted
• FEATURE ENGINEER:: No new features were added
• DROP2: 16 Columns duplicated by Encoded and Pivot Columns were dropped
• RENAME2: 13 encoded columns were renamed after original columns were dropped

NaN/Null: Only one column contained NaN/nulls in the data (it was in the corrupted field that was removed). OUTLIERS: No outliers have been removed or altered IMPUTE: No data was imputed

Split

• SPLIT: train, validate and test (approx. 60/20/20), stratifying on target of 'churn'
• SCALED: no scaling was conducted

A Summary of the data

Most features with 0min and 1max, the mean will represent the percentage of True values

Print nunique of all Columns shows a count of True and False for each feature, giving a quick glance at variance between feature values and allowing a quick infference into approximate percentages.

Explore

• Each of the three features were tested for relationship or difference against Churn.

  1. Tenure
  2. Monthly Charges
  3. Tech Support

• All three comparison features showed a significant relationship with the target feature Churn.

• There were four feature specific questions asked across three features all compared against our Target Feature of Churn.

  • 1.1 Is the average Tenure of Active customers greater than the average Tenure of Churned customers?
  • 2.1 Are the average monthly charges of customers that Churn higher than the average monthly charges of Active customers?
  • 3.1 Is the average of customer Churn without Tech Support greater than the average of customer Churn with Tech Support?
  • 3.2 Is the average of customer Churn without Tech Support greater than the average of Active customers without tech support?

• Three statistical tests were used to test these questions.

  1. T-Test
  2. Pearson's R
  3. $Chi^2$

• The first two questions 1.1 and 2.1 did not test positively against our stated question.

• The remaining two questions 3.1 and 3.2 involving Tech Support both tested positively against our stated question.

Exploration Summary

30% of all customers without tech support churn 82% of all churn is attributed to customer that do NOT have tech support Only 17% of customers with tech support churn Only 18% of all churn can be attributed to customers with tech support

Features I am moving to modeling With

• Churn is incredibly important as our target feature

Features I'm not moving to modeling with

• Tenure
• Monthly Charges
• Tech Support

Modeling

• Accuracy is our evaluation metric

• Our Target feature Churn, splits the data 27% Churn, 73% Active

• Simply guessing Active for every customer, we could achieve an accuracy of 73%

• Therefore 73% will be the baseline accuracy used for this project

• Models will be developed and evaluated using three different model types and various hyperparameter configurations

  • Decision Tree
  • Random Forest
  • KNN

• Models will be evaluated on train and validate data

• The model that performs the best will ultimately be the one and only model evaluated on our test data

Comparing Models

• Decision Tree, Random Forest, and KNN models all performed above the Baseline of 73%

• The KNN model performed slightly better on train data than it did on the validate data which may be a sign of overfit.

• Because the results of the Decision Tree, Random Forest, and KNN models were all very similar and above Baseline, we could proceed to test with any of these models.

• Random Forest however, is the best model that retained high performance across both train and validate data and will likely perform well above Baseline on the Test data.

Conclusions

Reccomendations

• Consider implementing incentives for increased Tech Support

Next Steps

• Decision Tree focused on other driving features above Tech Support
  • Investigate further into these features
  • Try running models with less features to isolate cause of predictions