credit-score-predictor

Using the sklearn Random Forest Library, I have created a model that predicts whether a users Credit Score is 'Poor', 'Standard' or 'Good' using a dataset that I found on Kaggle.

Pre-processing

This dataset was very messy, so a lot of cleaning needed to be done!

I realised that I only needed to read in the training dataset at first, and then just apply all those same changes to the test set when I needed to test the model.

To gain an overview of the dataset, I used 4 common commands:

1. head to visualise the first few rows of the dataset
2. shape to see how many rows and columns there are in the dataframe
3. describe to see the key statistics of each feature
4. info to see the data types of features and other details

I also used .isna().sum() to count the number of NaN rows. For rows that had loads of NaN entries, I simply removed these from my analysis as they weren't of any use.

Below are the rows that I removed along with reasoning:

1. ID, Customer_ID, Month, Name, SSN, Type_of_Loan: removed not due to data issues, but because the features are not good predictors of someone's credit score
2. Monthly_Inhand_Salary, Num_of_Delayed_Payment, Num_Credit_Inquiries, Credit_History_Age, Amount_invested_monthly, Monthly_Balance as they contained NaN rows. It wasn't clear whether to interpret NaN as 0 or missing data, so it wasn't fair to include them in training. I could've considered interpolation, but I wanted the training data to be as realistic as possible to get an accurate model.
3. Payment_of_Min_Amount as I didn't know how to format bool values for RandomForest

From a first glance, key issues that I noticed with entries in the CSV included underscores in numerical data which was easily fixed using str.replace(). Also, rows with negative values were removed as this is not possible, as well as duplicate rows. Another issue was pandas reading in the columns as incorrect data types, so I had to manually define each dtype.

Converting non-numerical data to labels

An issue that I ran into using this library with this specific dataset is that non-numerical columns are not directly supported. Following some StackOverflow searching, I found that the best solution to this problem is converting such columns to labels using sklearn's Label Encoder.

This was causing issues (with OneHotEncoder too!) so have removed categorical data for now.

The model

Testing

Improvements