Revolutionizing Credit: CASHe’s Approach to AI-Powered Underwriting
By Dikshant Agarwal: Head, Data Science, CASHe
This article delves into the journey of personal finance and credit management, emphasizing the transformative potential of technology in simplifying financial processes and enhancing financial well-being. It also underscores the importance of continuous innovation in the dynamic realm of credit management, positioning CASHe at the forefront of this paradigm shift.
Introduction
Growing up, I often accompanied my dad to the bank, observing how he managed his money. We’d sit in the manager’s office while our passbooks were updated and cheques for his business associates were processed. As I sipped on a graciously offered Coke or buttermilk, I listened to them discuss the state of the economy and the art of managing money, using terms like interest rates, FD returns, and tax-saving strategies. My dad expected me to learn these processes when I went to college. This sparked my initial insecurities about managing money, hoping technology would simplify these tasks over time. Fortunately, thanks to tech advancements like online banking, I haven’t filled as many passbooks or written as many cheques as I expected.
My first credit borrowing experience was a learning curve. After graduating from engineering college, I was offered an AmEx credit card but had no idea how it worked. Electronic payments like debit/credit cards felt insecure in the early 2010s when transacting with cash was the norm. A few months later, I used the AmEx card for an online purchase and didn’t pay the bill on time, unaware it would accrue interest and impact my credit history. This common experience highlights the widespread lack of financial literacy, making borrowing credit daunting. Today, culturally, we’re shifting from a lifestyle of scarcity to managing abundance, where borrowing done right can enhance life quality.
Solving for Credit at Convenience
At CASHe, we’re not just envisioning a future where credit is omnipresent and instant for everybody. We’re making it a reality. Our approach eliminates the need to visit any physical or digital avenue to get credit. It seamlessly happens in the background whenever and wherever you need it. In a world of 10-minute deliveries, entertainment at your fingertips, and a cab in a few clicks, we’re bringing the same level of convenience to credit.
Banks vs. FinTechs
Traditional banks have long struggled with inclusion, leaving over 80% of Indians outside the formal financial ecosystem. Their lengthy loan approval process has been a source of frustration for many. Credit underwriting, a crucial part of the customer journey, often becomes an operational bottleneck. However, the banking landscape is changing with the improving credit reporting infrastructure, increasing digital presence, and advances in ML and AI. CASHe, unlike traditional banks, leverages a broader range of data points processed by state-of-the-art AI/ML models to underwrite customers better and faster, revolutionizing the credit industry.
The broad categories of data points that a modern customer typically generates include:
Mobile device and app usage data: Variables like the number of finance, productivity, or shopping apps used, operating system, device brand, and app session data.
Customer demographic data: Current/permanent address, number of dependents, age, salary, employer/business profile, education level, etc.
Past borrowing data: Number of loans taken, credit cards used, active loans, recent inquiries, etc.
Transaction data: Recent high-value transactions, consistent bank balance or salary, necessary vs. lifestyle expenditure, etc.
References/community data: Risk profile of references and closest community members.
Macroeconomic factors: Indicators like interest rates, inflation, and unemployment at various regional and sectoral levels.
Personality/psychometric data: Indicators related to personality type, such as degrees of conscientiousness, extroversion, neuroticism, etc.
CASHe’s Underwriting Secret Sauce
Underwriting is simply the decision to give or not give a loan to a customer. For CASHe, getting this right is crucial. The complexities involved in underwriting in an immature and heavily regulated market like India are high and ever-changing. It’s also a high-rejection activity by nature, as resources (i.e., money) are finite, and a single bad month or quarter (like during Covid) can significantly affect future prospects.
The main components of underwriting include fraud identification, assessing existing debt and cash flows, and evaluating a customer’s intention and ability to repay. The raw material for underwriting is data. We’ve described various data sources contributing to final decision-making and how CASHe differs from traditional banks. These data sources are either submitted by the customer during onboarding, collected from third parties (like credit bureau agencies, KYC vendors, etc.), or generated as metadata while using the platform.
Evaluating a customer’s intention and ability to repay is where Big Data and Machine Learning shine, allowing us to create a consistent underwriting recipe for all customers. Defining the target variable is crucial for financial modeling and ML model training. For the loan default scenario, the definition of default is subjective. For someone in finance, the default might be NPA or DPD90+ (i.e., 90+ days past the due date), but NPA is a lagging indicator of portfolio performance. A customer who hasn’t paid their first installment is a more severe defaulter than one who has repaid 2/3 installments and defaulted on the third. Thus, based on business context and collections strategy, the target variable can be defined using other definitions of DPDs (1+, 30+, etc.) at various installment levels (first EMI, last EMI, any EMI, etc.).
The secret ingredient of our underwriting process is the feature engineering of raw data, extracting and refining it to create valuable indicators. For example, a customer’s device brand can indicate if users of a specific smartphone brand have a higher default rate. Additionally, metadata of the same device such as the device’s original launch/retail price, model release date, initial vs final OS state, and hardware and software specs provide more context about the customer’s psychology, purchasing preferences, digital savviness, and general lifestyle. Combined with other independent sources like banking/transaction data and past borrowing behavior, these interactions help the statistical model learn patterns better and boost prediction performance on new customers.
Another ingredient of the underwriting secret sauce is the modularity of different data sources and how they are consumed in our internal risk models across our different customer journeys — CASHe app, embedded journeys like GPay, IRCTC, Amazon, etc. or checkout finance at a nearby merchant. The underlying philosophy for underwriting is similar across these diverse journeys, even with the varying degrees of raw customer data available. This happens via an unsupervised + supervised way of training our models on much smaller datasets and extrapolating their learnings on a much larger base.
Proprietary Risk Scores: SLQ and GM
SLQ (Social Loan Quotient) and GM (Goodness Measure) are CASHe’s proprietary alternative data-based risk scores. Unlike conventional bureau scores based on past borrowings, these scores consider the customer’s digital and device footprint to create a risk profile. This approach has several advantages:
Authenticity: This data is hard to fake as it’s based on the customer’s device usage and past web activity, unlike manually aggregated bureau tradelines data.
Sensitivity: These scores are more responsive to changes in customer lifestyle, with sensitivity at the hourly/daily level compared to bureau scores, updated once every three months.
Inclusivity: They can score new-to-credit customers, avoiding the cold start problem of traditional bureau scores.
SLQ is a customer-facing alternative credit score transferable across geographies. Over time, it adapts to customers’ borrowing behavior in a new region, fine-tuning its scoring mechanism. GM is a region-specific, loan-level risk score designed to assist underwriters in evaluating the customer’s risk profile based on alternative data indicators. It sequentially learns from the customer’s past CASHe loans, carefully weighing them to assess risk on subsequent loans.
Conclusion
Since its inception, CASHe’s mission has been to leverage the power of big data and AI to underwrite customers better. We’re currently on the sixth generation of our AI-powered risk models and can confidently proclaim that our in-house risk system performs significantly better than any third-party solution. In the last few years, we’ve aligned the architecture and philosophy of our credit decisioning engine with evolving business and product requirements. In the unsecured lending market, where volumes are high and stakes are higher, this system must work as expected, if not better. Just as “search” has been the holy grail for Google, credit risk decisioning is for CASHe, and it’s a never-ending journey for us.
