---

Database

Link: https://www.kaggle.com/upadorprofzs/credit-risk

License: none

---

Used libraries

• Pandas
• Matplotlib
• Seaborn
• Sklearn
• Pickle

---

Description

This database shows a list of customers with income, age, loan size, and whether they have paid off the loan or not.

We have the average 45.33K in incomes, 44.44K in loans and an average age of 40 years.

With this plot we can visualize that there are higher chances to the customers pay it when the loan is lower

We have more loans with smaller amounts in this database

We can see that most people paid[0] the loan comparing to people who don't pay[1]

---

Changing wrong or non-existent values

With credit_risk.loc[credit_risk['age'] < 0] we discover that we have 3 registers where the age are negative. To change this I decided to replace these values with the average age of the whole database. We also had missing values, I fixed that too by replacing those values with the average age of the whole database.

---

Dividing predictors and class

I split the database between predictor attributes, now as the variable x, and classifier attributes, now with the variable y.

---

Tranforming the values with standard scaler

I used the standard scale to normalize the database. It basically standardize the database so there isn't too many different values that can hinder the machine learning algorithm. Example: a value of 20000 compared to 10 may seem "better" to the algorithm so this standardization is done.

---

Dividing the database between train and test

After the standard scaler I split the base between training and testing for the algorithm to train and then test it. I used 25% of the database for training.

---

Using neural networks to predict if the customer will or not pay the loan

After pre-processing the data we will create, train and test the neural network. I used the following values for MLPClassifier from sklearn:

max_iter=1750, verbose=True, solver='adam', activation='tanh', hidden_layer_sizes=(25, 25)

We train the algorithm with the database, and create the variable for the neural network to predict whether the client will pay or not pay. We've got an excellent result, 99.67% accuracy score.

Classification report:

              precision    recall  f1-score   support

           0       1.00      1.00      1.00      1292
           1       0.99      0.99      0.99       208

    accuracy                           1.00      1500
   macro avg       0.99      0.99      0.99      1500
weighted avg       1.00      1.00      1.00      1500

In the end I saved the classifier, you can find it and use it with the file neural_network_creditrisk.sav.