In Home Credit, we decide whether to approve or reject tens of thousands of applications daily. Our goal is to make this process as fast and as objective as possible. Hence, we eliminated all the manual underwriting. We make our decisions on clients fully automatically. Our underwriting team makes decisions on how the underwriting strategy should look, but not on each client individually.
The aim of our underwriting strategy is not to have the lowest possible risk but to align the risk precisely with the shareholders' risk appetite while sustaining our business. The most important aspect is how to uphold responsible lending principles in our business to avoid imposing a financial burden on our customers.
The main tool we have for managing risk is scorecards. These are statistical/machine learning models that predict a client’s future payment behavior based on data we have about the client. We can build these models because we have historical data on how different clients paid us in the past and their data at the time of application. Many years ago, the models were mostly built on application fields, internal payment behavior and Credit Bureau data. Now, with smart usage of internal and external data, we know much more about the clients than we used to, enabling us to reduce credit risk and offer better pricing to less risky customers. We see how clients behave in our app and similar, more advanced internal data. We cooperate with e-commerce, telco, and other partners who provide us with aggregated scores on clients. Here, I need to stress out we do not get any raw data about the clients from the partners, only a number from 0 to 1 representing the client’s probability of not paying us.
Even though the scorecard is the strongest tool in our arsenal, it is still just giving us the probability of a client not paying us (score), but it does not give us any final decision on whom to approve or reject. We need to define what score is good for approval, and we call these setting cutoffs. One single cutoff to reject/approve is not sufficient as the score only predicts risk but does not include factors like profitability, partner relationship or expected market behavior on certain client segments. All these we handle by setting different cutoffs for different segments of our clients.
As a failsafe and for regulatory purposes, we also use what we call 'Hard checks', which are simple one-dimensional rules that, if clients meet them, we reject them regardless of how good their other data is. This is much less surgical than a scorecard, so we try to use it as little as possible, but it has certain benefits. Namely, it helps us include anti-crisis rules, as scorecard might not have learned this from its’ training data that did not include all the historical crises, and it can help manage fraud because when we have a reason to believe we are dealing with fraudsters, other data should not be considered as they might be falsified or belonging to a different person. An example of the anti-crisis rule would be the number of loans in a credit bureau. If clients have too many, this is not good. Why does the scorecard not catch this behavior? The reason is simply when there is no crisis, these people tend to pay well, and having access to finance is a good sign. Plus, some of them sadly cover one loan with another loan, so, during a good economy, they indeed are paying well, but during a crisis, these people are very likely to start defaulting.
The last tools we have at our disposal are limit management and down payment. We calculate clients' affordability, and for clients where we see their debt is too high or don't have the ability to pay a high annuity, we decrease the credit limit available to them.
How do we know when to use which tool? Mostly, we try to either predict or analyze existing risk increases and try to understand which risk tool to tweak depending on the root cause of the problem. Sometimes, the problems can be solved not by tweaking risk tools but rather by changes in business processes like product design, fixing mobile app issues or sales (change training of Sales assistant, change bonus scheme, close fraudulent shops) and so on.
