It is no secret that Artificial Intelligence (AI) improves credit underwriting. It allows creditors taking advantage of the vastly more data and switching from point-in-time solutions to change-over-time solutions – resulting in credit models with higher accuracy than ever before. With a traditional, point-in-time, model is the applicant’s number of historical payment remarks calculated and measured. With an AI, change-over-time, model is it possible to understand the trends and spot the differences between applicants who have gained payment remarks recently compared to the ones who got them several years ago. Hence, AI allows creditors to identify the difference between the applicants who are on an upgoing and downgoing trend in their creditworthiness. Applicants who otherwise would have been seen as the same. This is only one example of how AI allows creditors to identify risk more accurately.
So, why hasn’t AI been highly adopted in credit underwriting?
Due to a regulated industry and since AI-models has historically been black-box solutions.
Demystifying the Black Box
We at Evispot have decided to transform credit underwriting by using the power of AI combined with transparency and explainability. Our mission is to deliver solutions which elevate creditors to new heights – our customers should be able to meet the demands of transparency while still having the accuracy and performance of AI.
At Evispot we have demystified the black-box, through three steps: transparency, explainability and provability.
Figure 1: Breaking down the black box in three steps.
Transparency, explainability and probability are our three technical key indicators helping us and our customers understand the full spectrum of how the model reasons, predicts, and – simply put – works. Transparency is about the AI-models high-level performance, the holistic view of its characteristics to help the creditors understand both the impact of the AI-model and the reasoning behind the model. Explainability is about helping the creditor understand each decision separately. Lastly, provability is used as a quality measurement of each prediction, to ensure the highest possible quality in each credit decision.
Introducing Evispot Traits
The solution we have developed is called Evispot Traits. When Evispot Traits sees an AI-model it returns a transparent and explainable description of the model. An AI-model is trained on a set amount of historical loans, good and bad ones with a set number of variables for each loan. The model has learnt itself which variables and combinations of variables that are significant for good and bad payers. Evispot Traits calculates how much a specific variable contributed to an actual historical event. This allows us to create a score of each variable’s contribution to each historical loan.
This number is called Evispot Score, it varies between -1 to 1 and each historical application score is plotted as a dot (see figure 2 below). The horizontal position of the dot is the impact of that feature for that specific application. A dot skewed towards the left – a negative value – increases the probability of default. While a dot skewed to the right – a positive value – decreases the probability of default. By calculating and plotting the Evispot Score for each variable and for each loan it is possible to calculate and visualise the overall model performance. Evispot Traits makes it possible to granularly understand how the AI-model reasons and how it determines the risk of each loan application.
Figure 2: Illustrating three variables that the model is using for predicting. Please note that the models that a full overview of the model’s Evispot score is listed at the end of this post.
By looking closer to the variable Annual income, it can be seen that the AI-model has seen four distinguished intervals. Missing value, below 34800, between 34800 – 218000 and above 218000. By looking at the Evispot score it can be clearly seen that applicants with missing value or less annual income than 34800 are defined as the riskiest group, while above 34800 is seen as a less risky group. By plotting Evispot Score it is possible to visually identify the models characteristic. By using this method it is possible to understand how age, salary or other variables correlate with risk. It helps us to better understand and trust the model performance and robustness.
Understanding Each Decision
When a new applicant enters, we calculate the Evispot score for this applicant and thereby give you as a creditor the possibility to truly understand the model and the reasoning behind each decision.
Figure 3: Illustrating how the model has reasoned when predicting a new application (red dots), which helps us understand why the model predicted the way it did.
By using the same example as used in figure 2, a comparison between the incoming applicant with the historical applicants can be made. By plotting the new application with red dots it can be seen that this applicant’s Annual income is skewed towards high risk. The applicant’s Account Balance, as well as Delinquencies, are skewed to low risk. The Evispot score of every variable gives us a complete picture of how and why the model predicts as it does. All in all, the benefits of Evispot Traits allows you as a creditor to gain the benefits of AI in your credit decision scoring while still meeting the demands of transparency and explainability.
Interested in learning more about Evispot Traits?
Please contact firstname.lastname@example.org or call +46(0)70-2332335
The Model’s Characteristics and Evispot Score for all Variables