Neural Network model designed to predict customer churn, optimizing for accuracy and computational efficiency.
Summary:
Is it possible to predict -which customers are highly likely to leave the bank within 6 months? I will design a ANN to answer the question.
A bank services dataset will be used to analyze customer behaviours and tendencies.
The following observations were extracted using essential statistical libraries against this bank's dataset:
-
The average Age of all Customers is 39 years.
-
75% of Customers have been with the bank for about 7 years with and average of 5 years.
-
The average Customer balance is 76.5k.
-
70% of all Customers have Credit Cards.
-
The average salary of Customers is 100k.
-
50% of Customers are from France and the remaining half are equally divided between Germany and Spain.
Recommendations:
- Right off the bat, bank offerings must include services in three languages: French, German and Spanish.
- There are twice as many French customers, therefore the marketing and customer support and customer retention budgets should reflect the same.
- Bots or AI assistants should provide culture and language sensitivity to reflect customer's distribution.
- Important: EU AI Act applies to all 3 countries and the bank must follow these regulations with strict compliance.
Additional Observations from the Exploratory Data Analysis (EDA):
- A fairly large number of Customers are keeping a 0 balance in their accounts.
- More precisely 3,617 out of 10,000 (36.2%) or a third of all customers. Recommendations:
- More research is required in this area. A budget for research should be added to understand contributing factors and corrective actions, RCCAs (Root Cause Corrective Actions).
- Incentives should be studied to provide value-add to these customers as a win-win for them and for the bank.
- A fairly large number of Customers are no longer active customers within the last 6 months.
- More precisely 2037 out of 7963 (26.6%) or a quarter of all customers leave. This is not acceptable.
- Incentives should be studied to provide value-add to these customers as a win-win for them and for the bank.
A Recommendation and Business Justification to design, construct and train a Neural Network Model to predict customer departures:
- A predictive NN model should be designed to anticipate Customer trends and provide opportunities to interject before is too late. (This project. :-) )
- Daily runs in production should provide 'alerts' whenever patterns are detected or treshold metrics are crossed with production or unseen data.
- The predictive model should be fine-tuned regularly to incorporate additional data values accounting for the preventative-actions and measure both their exit-recurrence effectiveness to the business and the efficiency of the model predictions.
- Transparency of Customer's Privacy protections should be auditable and test-demonstrated frequently as this model is predictive and could cause incorrect, even catastrophic communications to Customers and Regulators when biased.
Some examples from intermediate results: Only Model 0 anf Model 5 are shown here. The entire set of models 0--5 is in the python notebook in the repository.
Model 0 - Layers Summary:
Loss comparison for Training vs Validation data:
Model 5 - Layers Summary:
Loss comparison for Training vs Validation data:
Summary of all the models performance metrics:
After fine-tuning and optimizing the best trained model, it can be further used to make more accurate predictions:
