Banking is a valuable business. At least one study has found that bankers contribute $1 of value for every $7 they extract from an economy. You can see why people want to be bankers! But to be a banker, you have to have bank customers. You have to attract new customers and retain the ones you have. This repository contains a service that predicts the probability that a bank customer will be retained.
This project was inspired by and its dataset drawn from the bank churn Kaggle competition.
The code for this project is split across several different directories which are largely self-contained:
data - This directory contains data used for training. It also contains data (test.csv) which can be used to generate submissions for the associated Kaggle competition.
notebooks - This directory contains a Jupyter notebook for exploring and preprocessing data. The notebook also contains code exploring several different models (Logistic Regression, a Random Forest Classifier) and tuning their hyperparameters. This folder contains its own Pipenv files, isolating data and model development work from the deployed prediction service.
scripts - This directory contains preprocess.py, which contains functions for preprocessing data, which are importedd by the prediction service and train.py.
service - Dockerized Flask app that exposes the predicition service via a /predict endpoint.
kaggle - Contains a Python script that utilizes the /predict endpoint to generate a Kaggle submission file.
To experiment with the notebook:
cd notebooks
pipenv install
jupyter notebook
To train the model:
cd scripts
python train.py
To run the prediction service:
docker build -t app .
docker run -p 5000:5000 app
To generate the kaggle submission:
pipenv install
python generate_submission.py