Artificial Intelligence and machine learning in credit risk assessment

The provision of credit is a critical driver of economic growth. However, despite robust regulations and strong fundamentals, the Indian economy suffers from an acute credit gap. A good proxy for this gap is the credit-to-Gross Domestic Product (GDP) ratio which stands at 50% for India compared to 177% for China. The impact of this gap is acute for micro, small and medium enterprises (MSME) and nano-SME borrowers as the current banking infrastructure does not adequately reach them citing high operational costs and difficulty in underwriting. This is where the most impactful opportunity for Artificial Intelligence (AI) and machine learning (ML) in credit provision and decisioning lies.

As per ICRA estimates, in financial year (FY) 2024, we saw a 16% growth in credit with demand being led by unsecured loans of small value. While this growth rate is healthy, it led to concerns about poor lending practices such as over-indebtedness, sub-par underwriting causing the regulator (Reserve Bank of India) to tighten lending norms. This tightening will most likely depress credit growth rates to between 11-12% in FY25 and underscores the importance of risk management in the context of small loans i.e. at extremely low costs.

To understand and measure risk i.e. the creditworthiness of a borrower, we need to assess two things: Ability to repay and willingness to repay.

AI models offer a versatile toolkit for various stages of the customer lifecycle within financial institutions. These applications broadly fall into several categories:

· Credit decisioning: Employing AI/ML techniques in credit decisioning involves utilising supervised or unsupervised learning algorithms. For instance, leveraging ML to analyse credit bureau reports can unveil insights into incorrectly reported loans, specific repayment structures like bullet repayments, default trends across different regions and professions, as well as income distributions within districts and states. Such analysis aids in gauging a user's ability to repay.

· Fraud and bad actor detection: By scrutinising user behaviour during loan applications, including interactions with the application, copy-paste tendencies, data correction frequencies, and changes in connectivity, potential red flags can be identified. On the KYC front, assessing the integrity of user data across various sources helps uncover fraudulent borrowers and assess their willingness to repay.

· Early warning signs: Post loan disbursal, financial institutions must monitor repayment patterns closely. Scrutinising bureau data and employing ML techniques enable the identification of risks, facilitating proactive measures for successful collections.

· Operational efficiency: Intelligent systems can streamline operational workflows by learning and automating actions typically performed by operations teams. Implementation of ML techniques significantly reduces turnaround time (TAT) and minimises error rates resulting from manual interventions.

· Improvement in collection efficiency: In a lending institution, effective collections are paramount. AI models can identify repayment patterns, preferred modes of repayment, and user interactions with communications, enabling proactive issue resolution in collections.

Selecting the appropriate AI/ML algorithm hinges on business nature and the quality of collected data. For institutions dealing with unstructured data, unsupervised learning offers valuable insights. Clustering or association algorithms are viable choices for generating models in this context. Conversely, supervised learning is more apt for established financial institutions, leveraging collective intelligence from user data. Regression and classification are the primary algorithm types utilised in such models.

Two credit sub-sectors are likely to witness the significant AI linked uptake in the coming years. First, women borrowers who are already outpacing men in credit demand especially for small business loans. Women borrowers typically have less traditional underwriting data available at the time of application but more than adequate alternate data in the form of savings + spends, group savings etc. With custom AI/ML tools, not only can prevalent underwriting gender biases be uncovered and eliminated, they can also lead to better alternative data-based underwriting.

The second sub-sector comprises rural and semi-urban borrowers where risk assessment often needs to capture data well beyond the individual borrower such as household income dynamics, seasonality of inflows etc. which is again ideal for AI based models to learn from and deploy.

Overall, the power of AI/ML tools to transform how and to whom credit is delivered is especially relevant and important for India’s growth story.

This article is authored by Mohit Gupta, co-founder, IndiaP2P.