As India’s fintech lenders scale their unsecured credit portfolios, many encounter a common roadblock: How do you approve more customers without triggering a spike in defaults? The answer may lie in an often-overlooked tool—Reject Inference. While conventional credit scoring models rely solely on past approved customers, reject inference helps fintechs mine insights from those they previously turned down, unlocking new lending potential while maintaining risk discipline.
Background and Context
In the digital lending ecosystem, particularly among Buy Now, Pay Later (BNPL), personal loan, and small-ticket credit products, many lenders rely on application scorecards that are built exclusively on disbursed loan data. This approach introduces selection bias, limiting the model’s ability to generalize and often excluding creditworthy individuals who were rejected under earlier, stricter policies.
Reject inference helps address this gap. By incorporating data from rejected applicants—especially those who later secured loans from other lenders—it enables fintechs to develop more holistic risk models and expand approval rates without compromising credit quality.
Why Traditional Models Fall Short
Models trained only on approved borrowers tend to:
Overestimate credit quality by learning only from a lower-risk group.
Miss good customers who were wrongly rejected (Type II error).
Perform poorly in real-world deployment, especially when approval policies shift.
As fintechs aim to increase “swap-ins” (new customer approvals), the lack of learning from rejected segments stifles both scalability and profitability.
How Reject Inference Works
Reject inference leverages observed repayment performance of previously rejected applicants—either directly (if they later borrowed from another lender) or via predictive modelling.
Key fintech-specific reject inference steps include:
1. Segmenting Rejections
Demographic Rejects: Excluded due to age, pin code, etc.
Derogatory Rejects: Severe bureau flags; usually excluded.
Policy Rejects: Score-based rejections; prime candidates for inference.
2. Building the Modelling Base
Create a dataset using:
Disbursed Loans
Reject Proxy Group: Rejected by you but funded elsewhere.
Reject Non-Proxy Group: Rejected, no follow-on borrowing.
3. Leveraging Bureau Data
Check if rejected applicants took a loan elsewhere (BNPL, PL, etc.).
Observe bureau behaviour post-rejection using permitted historical pulls (within RBI timelines).
4. Training the Known Good-Bad Model (KGB)
Use actual observed performance from disbursed and proxy rejected segments to build a robust initial model.
5. Simulating Outcomes for Non-Proxy Rejects
Use probabilistic scoring to simulate their repayment likelihood.
Create synthetic labels using model probabilities to represent “good” or “bad” borrowers.
6. Final Training and Sampling Strategy
Combine real and synthetic data, apply TTD (Through-The-Door) weights, and re-sample to ensure balanced representation from disbursed and reject segments.
Expert Viewpoints
Rajeev Chandrasekhar, Digital Economy Strategist:
“Reject inference will be the next key differentiator in risk strategy. It gives you a crystal ball for policy shifts—those who crack it will scale responsibly.”
Anjali Menon, Chief Risk Officer at a Fintech NBFC:
“Without reject inference, we were only scratching the surface. Now, we’re uncovering high-performing segments in Tier III cities that were unjustly declined under legacy filters.”
Industry Implications
Fintech lenders: Enables safer expansion into new borrower cohorts with limited traditional data.
NBFCs: Offers deeper segmentation of borderline applicants to balance growth and asset quality.
Regulators: With tighter scrutiny on digital lending, reject inference ensures credit decisions remain data-driven, not discriminatory.
Investors: Better risk prediction translates to higher portfolio stability and performance predictability.
Limitations and Constraints
Data Availability: Bureau reports for rejected customers may be missing or misaligned.
Consent Frameworks: RBI restricts long-term storage of bureau pulls for rejected users.
Data Leakage Risks: Using post-application bureau variables can introduce bias if not properly scrubbed.
To mitigate these:
Focus on short-term outcome models (3M-6M bad definition).
Develop platform-only underwriting models for bureau-dark populations.
Incorporate early delinquency signals instead of long-term write-offs.
Conclusion
Reject inference isn’t just a modelling trick—it’s a strategic enabler for responsible fintech growth. As regulatory frameworks mature and the demand for inclusion deepens, fintechs using reject inference will be better positioned to safely approve more, while keeping losses contained. For lenders aiming to navigate India’s expanding digital credit landscape, this technique could be the most underappreciated lever of risk-smart scalability.
