Which one is more effective in the modeling of credit risk? Scorecards based on traditional methods or machine learning and artificial intelligence?

With the fervor surrounding AI in the present, this question is unavoidable. It’s also somewhat ridiculous. Although some of the new players in the fintech sector might have a stake in promoting AI strategies, the reality is that traditional methods of scoring, as well as AI, provide different benefits to credit risk modeling when you understand how to integrate them.

For instance, our scorecard technology, which is well-established and leverages the power of AI machines and AI capabilities to create more accurate credit risk models, which are algorithms that analyze the likelihood that customers will be able to pay on time.

How FICO Uses AI to Build Better Credit Risk Models

When FICO creates credit risk models for the likes of credit originations, the models are divided into segments by different types of customers, as well as different types of credit offerings require various models in order to evaluate their risk to credit.

In risk modeling that is based on traditional methods, customer segmentation is built in “hard” lines and broad factors, like new customers and. an existing customer. It isn’t able to capture the behavior of specific entities or the most effective methods of segmenting scoring models.

To create models using an organized way, the data scientists of our company often make use of AI or machine-learning techniques to find a better method to divide the scorecards. It allows us to use AI to enhance risk prediction without constructing “black box” models that do not provide risk management, regulators, and customers the needed information on how individuals can score as they do.

Our teams of data scientists are currently using methods like collaborative profiles to identify entity segmentation based on the customer’s behavior. We are able to segment customers into micro-segments, according to similarity rather than traditional segmentation methods that are based on hard business characteristics. For instance, collaborative profiles are based on behavior archetype distributionswhich could comprise archetypes pointing to those who are credit-worthy as opposed to. those who are at greater risk and have a history of any other credit-related issues within their credit history.

The ability to detect the subtle changes in behavior and integrate these within the credit risk models is a major advantage for FICO customers. Our model is built on a model that has been proven time and again, as well as scorecards, and enhances them using sophisticated AI technology that drives more efficient segments and features within models.

Another option is to employ AI as well as machine learning to “train” models to discover the most powerful predictive capabilities, and to discover new relationships between factors that are input to create an improved model. For instance the utilization of credit by consumers is an essential element within a credit model and delinquency is also a key component, however the nonlinear combination of these could yield better results when using an AI model. Then, you can incorporate these new data into a conventional scorecard model to guarantee explain ability by using the same methods that regulators are used to.

However, predictive power should not be at the expense of explanationability. Some say, “The bigger machine learning model must be more effective. It can learn more about the information.” More complex models that cannot be interpreted, particularly in the production environment, generally don’t do as well. They’ve accumulated noise about the model’s data set and biases, and won’t provide a robust performance when the model is put into production. So, the primary aspect of making use of AI responsibly is not just creating a scorecard or a machine learning model that’s capable of being understood and ethically acceptable, as well as ensuring that it is responsible in the manufacturing environment.

Improving Risk Management Performance with AI as well as Machine Learning

Two examples below demonstrate how you can increase performance by combining machine-learning and scorecard strategies to enhance traditional scorecards.

In developing a credit-card model of churn, FICO data scientists used machine learning to uncover an extremely strong interaction between frequency and recency of use. The choice to add this feature as a nonlinear input feature that could be easily interpretable way into a scorecard resulted in the improvement of (~10 percent) of the measure of lift, which is used to assess the effectiveness of models that use attrition. Furthermore, a further 15% improvement in performance was obtained by combining machine learning using a greater number of features related to the frequency and recency of events. These improved predictive capabilities could translate into significant returns on portfolio profits when used in the most precise retention strategies.

In a study to create an equity portfolio for homeowners with only a few data points, the absence of sufficient “bads” (poor-performing loans) in our database caused some issues. Through the creation of a machine learning score that was optimized with hyperparameters, our team of data scientists was successful in proving that we lost a substantial amount of signals using ta traditional scorecard. Machine learning enabled us to alter the model’s performance from a binary result to a continuous result. Combining this technology with the scorecard, we developed an extremely robust, strong, and palatable solution. We also saw an improvement of 20% in the performance of models (KS) in comparison to a standard scorecard model by itself (see the next section below).

Use AI as a Tool

FICO has been associated for a long time in the use of AI as well as machine learning in our approach to analytics. How long? We recently filed a fresh explanationable AI patent application in order to improve on the IP of an FICO explanationable AI patent that was granted back in 1998. That patent had already expired.

The years of experience in the market have established our method, which is quite different from what we would call the “move fast and break things” attitude we’ve seen from a few AI startups within the credit industry. They’re focusing on using a variety of kinds of data from alternative sources, including information obtained from social media sites, to assess the risk of credit.

Innovation is wonderful, but you don’t want to add a large number of new sources of data that may not be suitable for credit decision-making and may be easily altered (like Social Media data) in any existing AI model that produces a number that cannot be explained or implies bias. Why wouldn’t you? In the first place, the lenders in a variety of markets must be able to explain the way a customer’s score was calculated. If you aren’t sure of the connections that are being uncovered through this data, then don’t use the model.

For those who are credit professionals and aren’t sure about switching from scoring cards to machine learning, do not be afraid. For more than 5 years, FICO has consistently delivered research-based advancements in the field of Responsible AI, specifically, specially-designed machine learning algorithms to manage credit risk that are modeled after the advantages of scorecard technology. These neural networks that can be interpreted are accessible, understandable, ethical, and auditable.

FICO’s Responsible AI research demonstrates the capability to reveal and eliminate biases that result from information that is reflected in the model. Even if the data is intentionally biased towards a particular part of the population over another within the lab of research, the technique minimizes the pick-up of biased signals in relation to essential relationships that are crucial. Our method is patented and eliminates any data noise or irrelevant correlations, in which bias can manifest itself.

It’s not advisable to build a blind machine learning model for an area that is as important as credit risk. Interpretable models are able to consume diverse complex data sets much faster and with greater prudence as compared to traditional ML models. FICO’s interpretability algorithm can detect and eliminate biases in a model of machine learning, which helps in the practicalization of human-involved development methods based on palatability, ethics, and explanation rather than allowing these crucial requirements to be pure luck or worse, completely ignoring them altogether.

FICO’s ongoing efforts in Responsible AI are extensively explored through my ongoing collection of LinkedIn Live talks with other experts from around the world’s AI community.

The Power of Scorecards

Scorecards can be a very effective instrument because, just like AI it is possible to incorporate non-linearity within the input layer and also make use of various characteristics that can be prescient in various ways for various subpopulations by using scorecard ensembles that are segmented. In contrast to other forms of non-constrained AI, scorecards provide transparency and accessibility. This is a central issue in any discussion on AI or credit riskYou must be certain you know how the decision was taken to make a sound decision when lending. A majority of AI technology remains a “black box” and can’t provide a solution to a client’s question, “How did I get this score?”

However, if I can make use of machine learning to reveal the most powerful and accurate potential features in credit risk, I’ll use them to directly integrate them into a scoring model that users are accustomed to (and regulatory authorities are used to evaluating). This will ensure transparency while also improving the accuracy of predictions, and ultimately the management of risk.

For more information about the latest developments on the horizon in AI or machine learning, be sure to check out my LinkedIn page and Twitter feed, which are constantly updated with my most recent ideas about the state of analytics as well as AI.