Sigmoidal Success Story


Reduction in the default rate


Rise in credit scoring process rate


Faster Feature Engineering

What was the business problem?

Many potential borrowers look good on paper and are classified as creditworthy, but are in fact risky. This is because traditional methods can give credit to consumers who, e.g., have a large number of reward credit cards but sign for them to get bonuses (the so-called “card churners”), and are not, in fact, profitable for the issuer.

On the other hand, many of those who could turn out to be creditworthy get rejected by the traditional credit model. This can be caused by factors like payment history, types of credit used, and recent credit inquiries. Therefore, it will deny credits to consumers without a good credit history, no matter what is their actual current situation.

What savings and profits does the client achieve?

Our AI solution made the process of credit scoring much more accurate due to quicker and broader data analysis. The credit scoring process has been sped up by more than 45% with fewer cases going to manual verification. The system allowed our client to detect high-risk borrowers more effectively, reducing the default rate by almost 15%. 

The accuracy improvement in risk mitigation helped better assess risky borrowers, who look good at first sight, effectively increasing clients’ revenue.

How did we accomplish it?

As lending is a big data problem, machine learning is a great tool to leverage. The more data you have about an individual borrower and similar borrowers (and their history of loan repayment), the better you can assess if someone is genuinely creditworthy.

Our tool offers a granular and individualized approach. Among the others, it can consider factors like current income, employment opportunities, and one’s ability to earn. The amount of this data is so vast that it is almost impossible for humans to go through it. AI, on the other hand, can analyze all of these data sources together to create a coherent decision, uncovering hidden patterns.

The solution can also complete regulatory demands to provide grounds for credit decisions automatically


Data Preparation

The bank already held the most valuable data: loan repayment track records. This is crucial to understanding the market dynamics. To build an effective model, the bank needed to share this data (anonymized) with us (an AI company).

To make the scoring more accurate, we used alternative data. Currently, there are no regulatory challenges in using diverse sets of data for credit scoring. However, some raise ethical and privacy concerns in this matter. Nevertheless, such criteria as the applicant’s income, career, education, or utility and rental payments, although it took time to gather, were valuable for effective credit scoring.

This first step was, therefore, collecting all the available data and integrating it into our model.


Feature Engineering

In this case, one of the most challenging parts of the project was Feature Engineering. It is the process of using domain knowledge of the data to extract features from raw data and to create new ones based on data insights. If feature engineering is done correctly, it increases the predictive power of machine learning algorithms.

The process was quite a time-consuming one and took most of our engineering team’s time when building the initial ML pipeline.



The problem with traditional credit scoring is that it reinforces historical biases. If the system is to be free of discrimination the data that is used needs to be debiased. The data required to identify and remove discrimination include gender, age, race, disability, national origin, religion, sexual orientation, marital status, and more.

After each iteration, the client got a detailed performance report, so that he could regularly track insights and statistics.

How did we boost the project with Sigmoidal DNA?

Design Authority

Sigmoidal considered model as a scalable container – collaboration and scaling on a cluster becomes much easier. This solution guarantees reproducibility allowing the creation of complex machine learning software.

Product-focus Method

Sigmoidal technical leadership enabled us to understand the entire project scope clearly. This maximized the long-term deployment and maintenance of the project.

Deployment Strategy

Sigmoidal focused on versions management. We increased the accuracy of the dataset in the crucial part – updating and tinkering with different parts of the model. With versioning management, our engineers scoped out the best model and its tradeoffs.

What technical stack has been designed and implemented?