Credit Risk Modeling Proposal
Incorporating a machine learning system for credit risk modeling into Citi’s existing loan
infrastructure could be beneficial for the company’s loan approval process. Since risk is
currently assessed largely by manual review, the addition of a machine learning algorithm could
provide a few core benefits:
● Efficiency: Automating the credit risk assessment process with a machine learning
system can save time and reduce the workload for loan officers. This can lead to faster
loan processing times and improved customer satisfaction.
● Accurate risk management: Machine learning models can analyze a large amount of
data in a short amount of time, often leading to more accurate predictions of credit risk.
This can help lenders to make better loan approval decisions and reduce the risk of
defaults, resulting in improved profitability and reduced losses.
● Adaptability to new data: Machine learning models can be easily updated with new
data, which means they can adapt rapidly to changes in the market.
Data Requirements
To create a credit risk modeling system, we’ll need to decide on the variables to include in our
system. Although we already have a lot of the data we need internally, we may also want to
obtain data from other sources, such as credit bureaus.
Some variables we’ll want to include in our credit risk model are loan amount, loan purpose,
employment category, education level, income, credit score, debt-to-income ratio, loan
repayment history, and outstanding loans. We may also want to explore other variables we can
create using feature engineering.
Data Outputs
There are a few different outputs that our credit risk model can produce to help lenders assess
whether to approve a loan. Two of the best options include:
● Credit risk score: A score ranging between 0 and 100, where 0 represents a low credit
risk and 100 represents a high credit risk.
● Probability of default: A probability between 0 and 1, representing the likelihood that a
loan applicant will default on their loan.
Rather than relying on a threshold score to automatically approve or deny a loan, loan officers
will use the model’s output to make informed loan approval decisions.
Architecture
There are a variety of model types and architectures that we can explore for use in this system.
A few common choices are neural networks, random forests, gradient boosting, and support
vector machines. We can also combine various models to increase the accuracy of our
predictions. We will want to assess which option performs best within the constraints of our
credit risk modeling system. The data output we ultimately choose will influence the best choice
for the model architecture, as well as the metrics we should use to assess model performance.
Risks and Challenges
When creating our credit risk modeling system, there are a few risks and challenges that we
need to keep in mind throughout our development process. Here are are a few examples:
● Ethical considerations: We must ensure that our credit risk modeling system is as
unbiased as possible. We should always be searching for sources of bias, with the goal
of creating an ethical risk modeling system.
● Data quality: Our model’s accuracy depends on having access to high-quality data
inputs. We need to keep an eye out for incorrect data, data shift with time, and data
leakage.
● Data drift: Data drift over time can negatively impact model performance. We need the
infrastructure to regularly retrain our model on new data.