Credit Score Analysis

It is an analysis used to determine the credit worthiness of an individual or company

Credit scoring measures the probability of a person (loanee) defaulting on their loan payments. This is used to make loans, credit cards, and mortgage lending decisions.

Credit Score Analysis

There is a similar term called Credit Ratings: This is a measurement of the creditworthiness of a group/ organization such as businesses, large corporations, and governments. It is also used for decision-making in asset-backed securities.

A higher credit rating suggests a higher possibility of debt payment and vice versa. Therefore, these ratings have a big influence on potential investor decisions. 

It also indicates less risk of default on the part of the debt taker and helps the debt-giving entity to precaution and make their due diligence.

The difference between Credit Scoring and Credit Rating is as follows:

The major difference between the two terms lies in whom they are used. Credit Scoring is used for individuals or MSMEs, while Credit Rating is used for large businesses, corporations, or governments.  

There are various types of scoring strategies. Some are automated, and different mathematical models are trained based on the available datasets. And some models have predefined factors with a set of weights for their factors. 

Automated models involve Machine Learning models - Classification, Regression. After these credit scoring models, we will look at various challenges, Opportunities, and Risks involved in the described credit scoring models. 

Then we will move on to more traditional credit scoring methods such as FICO and the Vantage Scoring model

Before moving on to various models, let us look at a simple Machine Learning model and how it is interlinked with mathematics, factors, etc. The generalized Additive Model is a linear model like linear regression. The only difference is that it can model non-linear features as well. 

If we think about linear regression in mathematical terms, it is more like the weighted average of various features. 

However, in the case of Generalized Additive models, the weights multiplied with features need not be constant. Instead, they can be functions of each feature. 

Similar to beta in the Linear Regression or CAPM model, we use a spline function that can have non-linear relationships. The sum of these splines forms the Generalized Additive Model.

Traditional models

Generalized Additive Models are more linear models. So let us understand how to do these GAMs work.

Z = s0x0 + s1x1+...+snxn

Here,

Z is the output value/credit score we are looking for, and, 

S is the coefficient similar to parameters of linear regression.

The difference between the two is S is a smooth function. This brings non-linearity to the function. 

There are various ways GAM fitting methods for the same. One of the most common algorithms is the backfitting algorithm. It also helps bring flexibility to the function as the number of parameters increases. 

These smooth functions are called spline. Splines are defined as smooth functions (polynomial functions, to be specific) that cover a small range. 

Credit Scoring Methods

Let us first have a look at some of the automated Credit Scoring methodologies:

1. Linear Regression: 

As one can observe that the CAPM model tries to find factors to model a linear relationship, Linear Regression is used for modeling CAPM and extended CAPM. This model helps explore various risk premia. 

The sum of squared residuals is minimized to find the model's parameters or coefficients. 

2. Discriminant Analysis:

Discriminant Analysis is more of a classification model. What this means is, Discriminant Analysis deals with categorical data. Therefore, it tries to find a mathematical relationship that measures whether an individual would be able to repay the loan. 

The two categories are default and non-default. This also works on companies and understands the possibility of a firm going bankrupt. One of the most widely used models is the Altman Z-Score model. 

The model is as follows:

Z = 1.2X1 + 1.4X2 + 3.3X3 + 0.6X4 + 1.0X5

Where,

X1 = working capital/total assets

X2 = retained earnings/total assets

X3 = EBIT/total assets

X4 = Market value of equity/book value of total liabilities

X5 = sales /total assets

3. Probit Analysis and Logistic Regression

Logistic regression, similar to Discriminant Analysis, is used for classification purposes. 

However, it deals with continuous data. Logistic regression is used to measure the probability of an event. 

Instead of just binary values, one can measure the probability from 0 to 1 of someone defaulting on a loan. Probit Analysis is also known as probability unit analysis. It is a regression-type model. This model analyzes binomial response variables.

4. Judgment-based models

These models use intuitive judgment and opinions for decision-making. This is useful in cases of lack of historical data.

Supervised Learning Models

Supervised Learning techniques are a type of Machine Learning algorithm. For example, word supervised indicates that the algorithms take a data set with input and the corresponding output and calculate mapping between the two(input and output). This function can then predict others' behaviors(understanding credit score).

Supervised Learning Models

These algorithms can again be used to predict the possibility of default. In addition, they can act as regression as well as classification problems. Here are some of the algorithms used for credit scoring.

1. Decision Trees:

As the name suggests, Decision trees consist of tree-shaped diagrams that calculate the statistical probability. Each branch breaks the dataset into smaller subsets. Each subset has certain commonalities among its elements.

2. Random Forests:

These are an amalgamation of many decision trees. This algorithm trains on different samples of the data set and hence are used to take care of overfitting the dataset. This is an ensemble method for regression and classification. 

Ensemble methods are the method that use multiple learning algorithms to improve prediction performance.

3. Gradient boosting:

Like Random forests, gradient boosting is also an ensemble method for classification and regression. The framework of this algorithm focuses on the optimization of an objective function. 

4. Deep Neural Networks:

These algorithms don't work on any objective function or a mathematical equation. Instead, they try to find patterns in the dataset. They use perceptrons, similar to neuron cells, for finding patterns. 

Each perceptron finds a pattern using a simple mathematical equation in the form weight*input + bias. Each of these perceptrons has activation functions adding to the complexity.

These algorithms work as a black-box model. For example, it is impossible to find why a perceptron's weights and biases have a certain value. Furthermore, since the algorithm has many perceptrons, the algorithm can be highly complex, mapping with non-linear patterns.

Unsupervised Learning Models - How do they work?

Unsupervised Learning techniques are one of the types of machine learning algorithms. In these types of algorithms, input is given. However, there is no output for training. 

Algorithms try to find patterns between the dataset by forming clusters or groups among themselves. Then, the patterns are captured based on the probability densities. 

Unsupervised Learning Models

Here are a few unsupervised algorithms. Each of these algorithms uses different approaches for finding patterns. Let us have a look at them.

1. Clustering:

Clustering algorithms form clusters in the dataset to identify various parameters for the output group. These are called clusters and not classes because clustering labels the dataset, and it is not just a classification where data is already labeled. 

The two most used clustering algorithms are K-Means Clustering and Hierarchical Clustering.

a.  K-Means Clustering:

K-Means clustering is used to form K clusters of the complete dataset. The algorithm selects k points as its centroid and tries to form a cluster by giving all the remaining points a centroid(centroid at the closest distance). 

Then centroids are reallocated based on the sum of the distance between points. 

b.  Hierarchical Clustering:

The algorithm is similar to K-means clustering because it also forms clusters based on the Euclidean distance. It initially appoints each dataset as a cluster. Then the two nearest points form a cluster, and so on. 

This goes on until the convergence.

2. Reinforcement Learning:

As the word suggests, reinforcement learning is based on rewarding and punishing behaviors according to the need. This algorithm is based on trial and error. 

It is neither a supervised nor unsupervised learning technique since learning is based on feedback from the environment.   

3. Natural Language Processing:

Natural Language Processing is a subdivision of Machine Learning that aims at understanding the language by converting the human language into a mathematical form or Machine language. 

It can analyze, understand, and generate human language. 

4. Blockchain for Decentralized Credit Scoring:

Blockchain, as we know, is used for higher security as it uses encryptions and hash functions. Therefore, blockchain has a lot of scope for use in Credit Scoring database storage. 

However, the large database storing increases costs significantly, i.e., cost per transaction increases significantly. 

Opportunities

Let us have a look at various Opportunities in the credit:

Financial Access and Inclusion 

There are a lot of untapped markets in the form of rural areas, where many people don't have bank accounts. In addition, many MSMEs/businesses are without external financing in the form of credit or loans because of a lack of collateral and credit history.

Opportunities

Automation of Processes

As we saw above, many models can be automated, reducing the cost and increasing the efficiency of credit selection(removing all the bias) for credit service providers. This will help increase the customer base.

Enhanced Consumer Experience

Since the credit scoring models have been evolving, there has been an increase in the space for customization of financial products for clients. This provides a better-personalized customer experience. 

Better Predictive Models, More Accurate Decisions

As more data are trained, more factors are explored there has been an increase in the accuracy of predictive models.

Here are some of the risks associated with credit scoring models:

Fairness

Since the data used is biased, as it depends on the decisions taken in the past by loan givers, there is always a possibility of bias in the predictive models. Hence it isn't easy to have an unbiased model for prediction. 

Interpretability

The use of complex algorithms reduces the interpretability in mathematical form. This leads to a lack of transparency, which further leads to difficulty in challenging a credit score or the resulting credit decision. 

This lack of interpretability may butterfly effect and lead to macro-level risk. 

Accountability

With the increase in data usage and the longer data value chains, determining who is accountable and responsible for the data accuracy and quality may be challenging. 

Along with these challenges, there is an increasing risk of Data Privacy and Data Security. 

Unintended Consequences - How do they affect the models?

Another classic challenge with innovative algorithms is their tendency to overfit the data. This occurs when the algorithm has adjusted too closely to a specific training data set to the extent that it cannot make accurate predictions on new data.

Selection bias may occur if there are limitations in the data used for model development. In addition, if there are limitations in the methodology used to develop the models, then statistical bias may occur. 

If historical data are used where social bias was prominent, the algorithm may enforce and amplify the social bias. 

Let us look at various challenges in automating credit scoring. 

Potential for Discrimination

There is always a possibility of discrimination. The algorithms in themself might not be corrupted(discriminate). However, the dataset used by Machine Learning algorithms might be as the database user will have discrimination from the historical discrimination by humans. 

Consumer Protection

Protecting privacy(data privacy) is one of the major concerns in many fields. Credit Scoring also needs strict policies, and security risks should be monitored and managed. The firms must also maintain transparency(right to information) for the individuals or stakeholders. 

Model Governance

A lack of skills to carry out audits or control and manage the use of credit scoring models has been observed. Model Governance is required as both the parties - the bank and the client- are at risk of bias.

Conventional Credit Scoring Models

Some of the conventional Credit Score Models are - FICO Scoring Model, Vantage Score Model, TransRisk, Experian's National Equivalency Score, Credit Xpert Credit Score, and CE Credit Score. These methods have a predefined set of factors to determine the score.

Methodologies

Along with predefined factors and an associated weight with them (just like a Linear Regression Model). Though these do not require a large dataset to train the model, they need months of credit history for scoring to score each person.

 The total scores generally range from 300 to 800. The higher the score, the higher the chance of approval of the loan application. Each model has its factors. 

Many independent organizations provide your credit score. Since there have been cases of Data breaches, people have an option to freeze them. The term used is a credit freeze. These models overcome the issue of bias based on historical data. 

Experian, Equifax, TransUnion, and Innovis are some of the credit reporting agencies. FICO has been a dominant player in providing credit scores. Experian, Equifax, and TransUnion introduced the Vantage Model to overcome this dominance. 

Let us have a look at the two most widely used models. 

FICO Scoring Model

FICO stands for Fair Isaac Corporation. This company evaluated the credit scores of lending companies. There are many models under the category of the FICO scoring model. However, all of them have a high percentage of commonality. 

FICO

They use the following five metrics to calculate a credit score:

  1. Payment History
  2. Credit Utilization
  3. Credit History
  4. Types of Credits
  5. New credit

1. Payment History

The weightage of this metric is 35%. Payment history is scored based on your history of declaring bankruptcy or foreclosure. A late payments history is not preferred.

2. Credit Utilization

The weightage of Credit Utilization is 30%. Many people utilize their complete credit limits. F CO's score of credit utilization uses 30% as a threshold. So I  a person has 30% or less credit utilization, their Credit Utilization will get a high score. 

3. Credit History

This factor makes up 15% of the total score. T e longer a person has used a credit card, the higher their credit history score will be.

4. Types of credit

This metric accounts for 10% of the total score. The value of this metric is based on a variety of credit forms used. The higher the types of credit used, the higher the score. However, suddenly taking up various types of credits can also be a red flag, as it increases the possibility of default. 

5. New Credit

New credit accounts for the remaining 10% of the score. This metric suggests that taking up loans or credits one after the other is a red flag and advice to spread the credit process over time.

Vantage Score Model - Conventional or Machine Learning Algorithm?

Experian, Equifax, and TransUnion introduced this model. FIC  and Vantage Score Model have a lot of commonalities. The metrics used are quite similar. The major difference lies in the weightage allotted to various factors. Van age Score Model requires much less credit history than FICO score models, which require at least six months of data.

Let us have a look at various metrics used in Vantage Score Model. These are the approximate percentages in a Vantage score model. Van age 4.0 uses Machine Learning to train the credit score model. 

Payment History: This metric accounts for about 40% of the total weightage. It reflects the time delay in past payments. Hen e, late payments drastically affect credit scores.

Age and type of credit: Age and type of credit weigh about 21% of the total score  Sim lar to how types of credit increase the score in FICO mode, types of credit also increase scores in the Vantage Score model. 

The point of view is if a person can handle varied credits and on-time payments, then they will be able to handle the given credit reasonably.

Credit Utilization: This factor accounts for 20% of the total score. Cre it Utilization = Balance/Available Credit. It is suggested to keep it below 30%

Total Balances: This accounts for 11%

Recent Behavior(5%): This metric deals with the newly opened account. As there is not much data available on their credit payment history, their recent behavior is studied in detail for decision-making. 

Available Credit: This accounts for 3% of the total score.

Machine Learning Package Course

Everything You Need To Master Algo Trading using Python

To Help You Thrive in One of the Most Future Proof Careers on Wall Street.

Learn More

Researched and authored by Punit Manjani | LinkedIn

Edited by Aditya Murarka | LinkedIn

Free Resources

To continue learning and advancing your career, check out these additional helpful WSO resources: