Data Science Series: Navigating complex credit risk models to give consumers relevant and actionable advice
Author: Stephen Miller, Data & Analytics Innovation Leader, Europe, Equifax
With the advent of GDPR and upon advice from UK regulators, there’s an increasing expectation on financial institutions using consumer data for credit risk modelling to explain how automated systems make decisions. However, explanations of credit risk models and decisions don’t necessarily translate into an ordered sequence of actions a consumer can take to improve their score. The increasing use of machine learning models also makes it more difficult to generate model explanations that can be translated into actionable consumer behaviour. This creates new challenges when it comes to explicitly navigating customers to a desired credit score.
Credit risk scores can be a confusing and cloudy subject for consumers - they’re often shocked to learn that a credit score was used in a decision instead of the credit score. Multiple generic credit risk scores exist, and these can be both useful and beneficial for consumers to learn more about the profiles of people that are assessed as being sound credit risks. However, lending decisions are often based on a product-specific application scorecard, and understanding the specific score is more useful to a consumer looking to be approved for a particular product, such as a mortgage.
Suppose a consumer wants to reach a given credit score or an approval threshold. There are products in the market that provide generic advice to consumers looking to improve their credit score, such as keeping credit utilisation below 30% or not taking out a mortgage more than three times your income. But this advice may not be relevant to a particular application scorecard, or to an individual consumer’s profile, and following the advice may be infeasible for some consumers, such as those with little to no credit history or trying to repair their credit file.
Other products simulate a consumer’s own credit score using ‘what if’ scenarios, intended to provide the consumer with the knowledge of how their score would change as a result of certain actions, such as applying for a new credit card. These solutions are more relevant to the individual, but they do not generally provide a path to reach a user-specified score, nor do they provide a sequence of actions that are necessarily achievable in a fixed period of time.
In the USA, regulation requires that the key factors that impact a credit score must be provided alongside the credit score. These factors are the items on the credit report that have the largest negative impact on the score, but in many instances are not factors the consumer can change in the short term. For example, a default or heavy search activity cannot be instantly erased. Therefore, there is a potential gap between what is causing the largest negative impact to a credit score versus what the consumer can do to improve their score.
What is needed is an approach that provides the consumer with a sequence of feasible actions, achievable over a period of time, tailored to their specific credit profile, that will enable the consumer to reach a desired score threshold.
The biggest challenge in developing this algorithm is to properly define what the feasible actions are for a particular consumer, given their circumstances, to overcome the pitfalls of trial-and-error credit score simulators or generic advice. Secondly, the algorithm must generate a sequence of such actions that provide an optimal path across a potentially complex scoring surface, such as one generated by machine learning algorithms, capturing non-linearities and interactions.
Overcoming these challenges was central to the work of our Data Science Lab to generate optimal, feasible paths that consumers can follow to improve their credit scores. The paths are built from reasonable and appropriate monthly steps, based on what individuals with similar profiles have been able to achieve. When followed closely by the consumer, the paths produce the desired score increase. The advantage of this approach over credit score simulators that rely on trial and error and ‘what if’ scenarios is that the steps proposed are more likely to be achievable for the consumer, in addition to being effective. The algorithm can be applied to any score, requiring only access to a representative sample of anonymised consumer credit profiles and scores over a short period of time. Access to the scoring function itself is beneficial, but not essential.
The optimal paths algorithm provides feasible, actionable, and impactful recommendations to the consumer, ensuring they have the best possible opportunity to make the changes they need to improve their score.