Data Science & Analytics Virtual Intern at BCG :

Churn-Model-on-Data-Science-&-Analytics-Virtual-Intern-at-BCG

completion_certificate.pdf

__________________________________________

- First Task : Business Understanding & Hypothesis Framing

Understanding the business context and problem statement.

* PowerCo. needs BCG to help diagnose the source of churning SME customers.

* A fair hypothesis is that price changes affect customer churn.

* The head of the SME division is considering a 20% discount that is considered large enough to dissuade almost anyone from churning (especially those for whom price is the primary concern).

* i email the AD to erquest the data(Customer data,Churn data and Historical price data).

___________________________________________

- Second Task : Exploratory Data Analysis

Understanding the business through data.

* After the AD send the data, i Perform some exploratory data analysis. Look into the data types, data statistics, specific parameters, and variable distributions.

* Try to verify the hypothesis of price sensitivity being to some extent correlated with churn.

* Try to prepare a half-page summary or slide of key findings and add some suggestions for data augmentation

* After EDA, i found some conclusions :

• on average churned clients drives more gross and net margin

• consumption and forcasting between two groubs almost the same till 75%

• net margin will be a little bit higher than the current clients after 20% discount

_____________________________________________

- Third Task : Feature Engineering & Modelling

Uncovering signals within the data, predicting churn probability and evaluating model performance.

* Sub-Task 1 :

• Think of ways to evaluate a feature against a label.

• Think of ways to add new features which would complement the already existing ones.

• Think of feature granularity.

• Remove unnecessary features.

* Sub-Task 2 :

• Is this problem best represented as classification or regression?

• What kind of model performance do you think is appropriate?

• Most importantly how would you measure such a performance?

• How would you tie business metrics such as profits or savings to the model performance?

* After preparing data, i do some feature engineering, then i Start with a Dummy Model (np.rand) - Baseline Model with accuarcy 0.51 and F1_score 0.18

* I try a Simple Model (linear) with :

• Linear model

• calculate score

• calculate feature imporance

• Simple model with top 10/20 features

* Also i try a Simple Model with Balanced Dataset ( Upsampling[f1-score 0.21], Downsampling[f1-score 0.22] ).

* Then i try Complex and Explainable Modeles (Tree Based) :

• Decision Tree (Original_data[f1-score 0.12], Upsampling[f1-score 0.15], Downsampling[f1-score 0.19] ).

• Random Forest (Original_data[f1-score 0.05], Upsampling[f1-score 0.08], Downsampling[f1-score 0.19] ).

* Finally i try Deeper Modeles :

• XGBoost (Original_data[f1-score 0.04], Upsampling[f1-score 0.12], Downsampling[f1-score 0.22] ).

• CatBoost (Original_data[f1-score 0.03], Upsampling[f1-score 0.11], Downsampling[f1-score 0.22] ).

* The Best Model is : CatBoostClassifier With Downsampling data because the accuracy is higher.

___________________________________________

- Forth Task : Findings & Recommendations

 Presenting the results and giving recommended actions to the client to AD in draft an abstract (executive summary) of findings.