• Python 3.11.7
• Libraries: pandas, matplotlib, seaborn, scikit-learn, optuna, xgboost, graphviz, shap, hydra-core

1. Clone the repository to your local machine.
2. Install the required dependencies using Poetry:

git clone https://github.com/frkangul/rpaa-banking.git
cd rpaa-banking
pyenv install 3.11.7
pyenv local 3.11.7
python -m venv .venv
source .venv/bin/activate
poetry install

Run the main training script after setting configurations from the config dir:

python train.py

This will launch a training pipeli and create result directory with the given name via configuration files.

• Improve the profitability of the loan portfolio

• A personal loan is profitable if the customer repays the installments. If a customer defaults and the bank is not able to recover the lended money, it causes a loss to the bank.

• Loan delinquencies are defined as the number of installments not paid after the due date. When the delinquency exceeds 3 months (=3 installments, 3 payments missed), the loan is defined as defaulted.

• The loan amount can be considered lost when a default occurs.
• For portfolio management only the first 12 installments can be considered. Delinquencies on later installments are not provided, and hence can be ignored.

To solve the recruiting case for the RPAA team, I will outline a plan to build a predictive model that aims to maximize the future profits of the bank's personal loan portfolio. The model will predict the likelihood of a default loan based on the provided datasets. Here's how I would approach the task:

The primary goal is to predict the probability of a loan becoming default within the first 12 installments. This will help the bank to make informed decisions on whether to approve a loan application, adjust the loan terms, or take preemptive actions to mitigate potential losses.

• loan_data.csv: Analyze the distribution of loan reasons, terms, requested amounts, and customer account numbers. Check for missing values and outliers.

• customer_data.csv: Examine customer demographics and their relationship with the bank. Convert birth_date and joined_BANK_date to age and length of relationship with the bank.

• customer_financials.csv: Calculate average salary, current account balance, saving account balance, and credit card balance for each customer over the provided period.

• loan_delinquencies.csv: Create a target variable indicating whether a loan has become default (1) or not (0) within the first 12 installments.

• Create new features such as debt-to-income ratio, average monthly balance, and credit utilization from the financials data.

• Encode categorical variables like loan_reason, gender, religion, and employment using one-hot encoding or label encoding.

• Consider interaction terms if they make economic sense (e.g., loan amount vs. salary).

Given the binary nature of the target variable (default or not), I would consider the following algorithms:

• Logistic Regression: As a baseline model due to its simplicity and interpretability.

• Gradient Boosting Machines (e.g., XGBoost (compatible with M1 series MACs)): For their performance in classification tasks.

• Split the data into training and validation sets.

• Standardize numerical features to have a mean of zero and a standard deviation of one.

• Train the models using the training set and perform hyperparameter tuning using cross-validation.

• Evaluate the models on the validation set using appropriate metrics (e.g., Recall and AUC precision-recall).

• Analyze the feature importances to understand the drivers of loan default.

• Select the model with the best balance between performance and interpretability.

• Prepare a presentation summarizing the approach, model performance, and key findings.
• Discuss potential next steps and improvements if more time were available, such as:
  • Incorporating additional data sources (e.g., credit scores, macroeconomic indicators).
  • Exploring more complex models or ensemble methods.
  • Conducting a cost-benefit analysis to optimize the threshold for classifying a loan as high risk.

Disclaimer: the data provided for the purpose of this exercise is in no way referred to actual bank clients, but it is rather based on simulations.

The data should be in data/.

Contains all the information about the all the loans disbursed in the years 2016, 2017 and 2018.

• loan_id: unique identifier of the loan
• cust_id: unique identifier of the customer that requested the loan
• date: month and year when the loan has been disbursed. Example: 2017-09-01 means that the loan has been requested and disbursed in the month of September 2017
• loan_reason: what is the loan meant for. Categories go to:
  • Car: Financing of a new vehicle
  • Housing: Related to any house related expense (renovation, new furniture…)
  • Financial: repayment of some other debt
  • Personal: Personal Reasons
• loan_term: number of installments (in months)
• requested_amount: total lended amount in Euros
• installment: monthly installment amount, in Euros
• number_accounts: estimated number of non debt related products that the customer has at the time of the loan application. The considered products include current accounts, saving accounts and investment accounts.
• number_loan_accounts: estimated number of debt-related products that the customer has at the time of the loan application. The considered products include other personal loans and mortgages.

Information about the customer, extracted from the database in January 2020:

• cust_id: unique identifier of the customer that requested the loan
• gender: gender of the customer (M for Male, F for Female)
• religion: if known, religion of the customer
• employment: employment sector for the customer
• postal_code: first two digits of the postal code, where the second is always set to 0
• birth_date: month and year of birth of the customer
• joined_BANK_date: month and year when the customer became a BANK client
• number_client_calls_to_BANK: total number of times the customer has reached BANK by phone
• number_client_calls_from_BANK: number of times the customer has been contacted by phone by BANK

Monthly information about product balances and detected salaries of the customer, for the years 2016, 2017 and 2018

• cust_id: unique identifier of the customer that requested the loan
• date: date when the date has been extracted.
• salary: detected salary relative to this date: If the date is 01-02-2017 (1st February 2017), this refers to the detected salary of the previous month (January 2017)
• current_account_balance: balance of all the current accounts on the extraction day at midnight. Example: If the date is 01-02-2017 (1st February 2017), this refers to the balance on all the current accounts on the 1st of February 2017, at 00:01 AM CET
• saving_acc_balance: balance of the saving accounts on the extraction day at midnight. Example: If the date is 01-02-2017 (1st February 2017), this refers to the balance on all the current accounts on the 1st of February 2017, at 00:01 AM CET
• credit_card_balance: due balance of the credit card on the extraction day at midnight. Example: If the date is 01-02-2017 (1st February 2017), this refers to the balance on all the current accounts on the 1st of February 2017, at 00:01 AM CET.

Contains the information of the latest deliquency on the installment. This data is provided only for the first 12 installments, in accordance with the banking rules.

• loan_id: unique identifier of the loan
• start_date: month and year when the delinquent installment was due
• end_date: month and year then the delinquent installment has been repaid. In case that the delinquent installment has not been repaid within 3 months, the date is set to the 4th month, no matter when and if the customer has repaid the delinquency.