Buy Now Pay Later (BNPL), a short-term interest-free consumer credit solution, is increasing in popularity in the US. However, despite explosive growth in sales volumes, none of the pure-play BNPL companies are profitable, including BNPL giants. Since the credit risk undertaken in BNPL is usually higher compared to other credit solutions, the traditional statistical models used for credit risk modelling will require adaptations for BNPL. Today, machine learning (ML) models have evolved significantly; they perform better and can predict Probability of Default (PD) and Expected Credit Loss (ECL) more accurately.

This blog discusses the reasons underlining higher credit risk in BNPL and approaches for better prediction, particularly how ML models can predict credit losses more accurately.

Significant credit risk at stake

While BNPL solutions are attractive to both merchants and consumers, BNPL providers are exposed to higher credit risk due to the following reasons: 

• Poor or no credit history: Though BNPL is offered to all segments of the society, it is popular among those who have poor or no credit history or are unfamiliar with credit and the consequences of default. This segment of customers poses a higher credit risk. BNPL borrowers with access to traditional credit are more likely to be highly indebted as per the recent survey from CFPB (Consumer Financial Protection Bureau). The survey sample included consumers with at least one traditional credit tradeline. BNPL borrowers exhibited higher measures of financial distress than non-BNPL borrowers and are more likely to have delinquencies in traditional credit products and lower credit scores.
• Behavioral risks: Customers are generally tempted to buy more than they can afford due to the instant credit decisioning facility. This impulse shopping causes a higher risk of default if customers do not manage their finances properly.

The above risks need to be considered while predicting credit losses. The impact of these risks can be better understood by incorporating the following measures in the credit loss calculation:

• Macro-economic factors tend to impact large swathes of population. This can help in forecasting the repayment capacity of BNPL borrowers, specifically those with poor or no credit history.
• Customers tend to leave footprints on their attitude and behaviors in social media conversations. Tracking such behaviors on social media can help in detecting high risk customers.

Influence of macro-economic factors

BNPL providers generally initiate a soft credit pull for credit decisioning. A customer’s credit tradelines and credit default data are among the important determining factors. Apart from these, macro-economic factors can also be a key determinant in evaluating the credit risk. For example, if unemployment rates are expected to increase, it can reflect as ‘higher credit risk’ in the credit loss estimation.

Considering these factors, researchers have built machine learning models to compute ECL for BNPL portfolios. They found that including macro-economic factors can predict credit losses more accurately. 

Social media for assessing behavioral risk

The repayment ability of a customer can change during the BNPL contract term due to personal, emotional, and psychological factors, such as the loss of a family member. Social media can be considered a powerful channel to gain such insights about customers. A customer’s attitude and behavior are also influenced by peers within their social circles. A research conducted to determine the impact of social media behavior on predicting default probability revealed that social media behavior data did produce more accurate results.

This research can be extended for BNPL as well. The below mentioned social media data can be used in building ML models specifically for BNPL. This can help in determining the changing behavior of BNPL customers while computing ECL for the contract term.

1. Number of social media platforms used by customers
2. Total count and demography of their followers, profiles they are following
3. Active posts, content quality and behavior in the posts
4. Endorsements or sponsorships
5. Group behavior on credit and payments (delinquents or defaults)

Closing thoughts

With increasing losses being reported by BNPL providers, it is highly critical to build robust mechanisms to forecast expected credit loss more accurately. This will help in refining the credit decision-making process. Financial data can be augmented with non-financial data like social media behavior and macro-economic factors to accurately forecast credit loss. Using the above mentioned machine learning models for forecasting will enable more accurate credit risk provisioning.