Scorecard is not only a machine learning algorithm, but also a generic modeling framework used to build a model for assessing credit risks. In scorecard modeling, the original data is processed by data binning and feature engineering, and then is used to build a linear model.
Scorecard modeling is typically used in credit assessment scenarios, such as for credit card applications and loan disbursements. It is also used in other industries for scoring, including customer service scoring and Alipay credit scoring. This project shows how to use the financial component on Alibaba Cloud Machine Learning Platform for AI to build a scorecard model.
The following dataset contains client information, such as gender, education, marital status, and age, payment history, and credit card billing statements. The payment_next_month column (goal field) indicates the probability of a client paying off their credit card debt, as shown in the following figure. A value of 1 indicates that the client will likely pay off the debt and a value of 0 indicates that the client will not likely pay off the debt.
The dataset contains 30,000 entries. You can download the dataset from https://www.kaggle.com/uciml/default-of-credit-card-clients-dataset.
The following figure shows the workflow of this project:
The procedure includes the following major steps:
Split the input data into two parts: one for model training and one for prediction result assessment.
Data binning is similar to onehot encoding. It is a process of grouping the input data into data classes (bins). The data values in each bin are replaced by a value, which is the representative of the bin. As shown in the following figure, the binning component groups the age values into a number of age intervals:
As shown in the following figure, after data binning, each field falls into multiple intervals:
Population stability index
Population stability index (PSI) is an important metric to identify a shift in the population for credit scorecards, for example, the changes in the population within two months. A PSI value smaller than 0.1 indicates insignificant changes. A PSI value between 0.1 and 0.25 indicates minor changes. A PSI value larger than 0.25 indicates major changes in the population.
By comparing the stability of the population before data split, after data split, and after data binning, the model calculates the final PSI values for all features as follows:
The following figure shows the scorecard training results:
The purpose of using the scorecard is to use normalized scores to indicate the weights of the features in the model.
- Unscaled: represents the original weight.
- Scaled: an index that indicates the amount of points that a feature gains or loses. For example, if the pay_0 feature falls into the (-1,0] bin, the feature gains 29 points. If the pay_0 feature falls into the (0，1] bin, the feature loses 27 points.
- Importance: represents the influence of each health exam indicator on the prediction results. The larger the value is, the greater influence the indicator has.
In this project, the modeling results refer to the credit scores calculated for all clients, as shown in the following figure:
You can use the credit card billing statements of your clients to train a scorecard model to calculate credit scores for all the clients. The credit scores can be used in loans or other credit dependent financial transactions for assessment.