credit score card

The Why

Imagine you are part of a financial institution which is in the loan lending business. Your objective is two-fold:

1.??????Attract new credit-worthy clients who need a loan

2.??????Ensure the existing clients pay back all the instalments regularly on time

But you are faced with three problems:

Before giving out the loan:

1.??????Once your brand is well-known, you’ll have no issues in finding people who need money. The challenge lies in selecting the right ones to grant a loan. You wouldn’t want a Vijay Mallya who takes a huge loan from you and then escapes to another country after defaulting. Even with a collateral, it is a loss-making business. Selecting credible clients and keeping out the potentially delinquent ones becomes the primary task of a lender.

2.??????Due to heavy competition, the opportunities for new businesses open for short time periods – you need to reduce the turnaround time involved in assessing the client’s profile. It is not just to avoid losing a good client altogether but also to reduce the per-unit processing cost of your institution.

After giving the loan:

3.??????To control your losses, you need a robust system that monitors all the clients who are in the process of repayment and raise a red flag in advance whenever they are about to delay any payment or about to default.

This calls for building an automated risk scorecard that can process thousands of clients within a few minutes and distinguishes a “good” client from a “bad” one. “Good” being defined as a low risk client who has a low probability to default on his obligations.???????????

The How (Basic concepts)

There are two types of scorecards that one can build to tackle the above problems.

1.??????Acquisition or application scorecard

A successful money lending business does not grant a loan to anyone who walks in the door. Every applicant is evaluated according to a set of criteria. This is called underwriting. Application scorecards are used alongside this to evaluate the risk of a proposal during the credit approval process for reaching a “approve/don’t approve/refer to higher authority” decision.

2.??????Behaviour scorecard

Once a loan has been disbursed, you must regularly re-assess the riskiness of the client. Due to the length of the repayment period (For instance, 1-5 years for automobiles, 20-25 years for house loans) the initial assessment may prove obsolete and regular assessment of the client’s probability of default is necessary. Therefore, behaviour scorecards are used to determine the client’s riskiness based on most recent financial information, including repayment behaviour and overall relationship with your institution.

The How (Modelling process)

Following are the steps involved in the modelling process:

1.??????Data

Assess availability and quality of data by performing an exploratory data analysis (EDA)

• ?% capture
• Quartiles, mean, median values
• Outlier / missing value treatment
• Highlight exclusions (Eg. Certain time period / regions)

2.??????Target definition

Define your “good”, “bad” and “indeterminate” cases:

• By industry standards, the clients who have defaulted or are at a 90-day delinquency (missed 3 EMI’s) are generally tagged as “bad”. This information is stored in the target variable, denoted as “1”, which suggests an occurrence of the event that we are trying to model.
• Clients who have completed all payments successfully (loan has matured) are tagged as “good”. We can consider adding those who have completed 90% / 95% of their payments as well here.
• Those are who still in the early stages of loan repayment and could potentially turn to be good or bad are tagged as “indeterminate”; they are usually not considered for the model building process.

3.??????Define sample and performance window

To predict the performance of future clients, we need to gather data for the approved and disbursed loans from a specific time frame (Sample window), and then monitor their performance for another specific length of time to determine if they were good or bad (Performance window). ?

For the sample window, one must be careful to not go too far back into the past, because approval criteria, business and market conditions may have differed significantly. However, it cannot be too recent too as the clients need to be given enough time to assess their performance.

For the sake of an example, let’s say today is the 1st of January 2018. Via past repayment performance analysis, we see that 80% of the loan defaults in your portfolio happen within a year and hence 12 months is established to be the performance window.

This means that our sample window can comprise of all the loans disbursed in Jan – Dec 2016 and every loan is given a rolling performance period of 12 months (up to Dec 2017) to assess their repayment outcome. The independent or input variables could comprise of all the information collected from the client at the proposal stage.

The above example makes sense if we are building an application scorecard. For a behaviour scorecard, the sampling process is tricky. The sample window should comprise of “live” clients i.e. they are in the process of repayment; accounts are chosen at one point in time. The performance window could be the next 1 to 3 months to assess the outcome, while the independent variables would mainly be derived from their behaviour over the past 6 to 12-month period.

4.??????Segmentation

In some instances, using several scorecards for a portfolio provides a better result than using a single scorecard for everyone. Segmentation can be experience-based (heuristic) basis demographics, portfolio type, source of business (online/offline), applicant type (new/existing); or it can be through statistical techniques like clustering.

5.??????Selecting the independent variables

Post EDA, selecting the right variables as input parameters for the model is an important task. Few guidelines to follow:

• Choose variables with high information value
• Not highly correlated with the target variable or other independent variables
• Spend some time to determine if the variables are to be used as absolute numbers or ratios (For eg. No of dependents can be an absolute number; overdue could be used as % of ‘to be received’ amount instead of absolute value)
• As a rule of thumb, ignore variables with >30% missing values unless there are compelling business reasons to include them

6.??????Modelling techniques

There are several mathematical techniques to predict the risk scores – Logistic regression, neural networks, decision trees, and so on. Most appropriate technique to be used can depend on quality of data, type of target outcome (binary or continuous), ease of implementation, interpretability of results, and finally the actual model results (AUC, Gini, Accuracy, TPR, precision etc.). Generally, high TPR (True positive rate or sensitivity) is the metric to chase here instead of precision as the proportion of defaulters is considerably low.

In my personal experience, the machine learning techniques gave better results for behaviour scorecards while regression method worked better for application scorecards.

7.??????Scorecard development

The probability scores obtained from the modelling process are enough to assist the business managers in the decision-making process. However, converting these scores into a point-based scorecard makes sense from a business and IT perspective for following reasons:

• Easy to understand - It helps any business user including the operations team or external collection agencies to understand the credit-worthiness of a client.
• Easy to implement – By having integer score points for each attribute, calculating total score by simple addition is much simpler and transparent than by using a formula.

Score points are calculated using a defined minimum/maximum scale (Think CIBIL Score) with a specified odds ratio at a certain point and specified rate of change of odds. Score points are another way to denote the scorecard, they do not affect its predictive power, and hence are not mandatory.

The What

Now that you have the risk scorecards in hand, what do you do. For high-risk applicants, some of the strategies are:

1.??????Reject the loan proposal if risk is too high

2.??????Charge higher interest rate for medium risk

3.??????Ask the client to provide a higher down payment or deposit for mortgages or automobile loans

4.??????Charge higher premium on insurance policies

5.??????In telecommunication industry, you could offer prepaid services instead of postpaid, or block international calling for high-risk clients

6.??????Higher number of checks before approval of loan

On an ongoing basis, the scorecards could be used for:

1.??????Identifying good clients for up-sell and cross-sell

2.??????Increasing credit limits on credit cards and lines of credit

3.??????Deciding whether to give a top-up loan or reissue an expired credit card

4.??????Directing high-risk accounts to more stringent collection methods or outsourcing them to a collection agency

5.??????Deciding when and from whom to repossess asset (automobile, expensive goods, house etc.)

6.??????Put an account into a “watch list” for potentially fraudulent activity

As a final note, do ensure that the scorecard developed is applicable to the current population on a regular basis, else re-calibration is necessary. There are various ways to create a stability index, but a simple way to establish the stability of a scorecard is to monitor the % distribution of the population in different score bands – this should not deviate much from the expected/historical distribution.