The Gig Economy Problem: Why Standard Credit Models Fail Hustlers and Anchors

Why Do Standard Credit Models Systematically Undervalue Gig Workers?

India has approximately 450 million workers in informal or gig employment. This includes:

• Platform gig workers (Swiggy, Ola, Urban Company)
• Freelancers and consultants
• Self-employed tradespeople
• Small business owners

Traditional credit models are calibrated on loan performance data from a customer base that skews heavily salaried. The models have seen millions of examples of "salaried employee with regular income → paid back loan" and a much smaller number of "gig worker with variable income → paid back loan."

The volatility misread: When a gig worker's monthly income shows: ₹45K → ₹62K → ₹38K → ₹71K → ₹55K, the model reads high variance. High variance in salaried employees is indeed a risk signal (it might indicate job instability). But for gig workers, income variance with a positive trend is often a business growth signal.

The model was never trained to distinguish between:

• "Salaried income dropping from ₹45K to ₹38K" (risk signal: job loss)
• "Gig income dropping from ₹45K to ₹38K in a low month" (noise: seasonal variation)

Both look identical in the feature space, but carry fundamentally different credit risk implications.

What Is Decile-Level Drift and How Does It Expose Segment Bias?

How standard monitoring misses segment bias:

A model might show overall accuracy of 92% and PSI of 0.08 (stable). But:

• Accuracy for salaried employees: 95%
• Accuracy for gig workers: 76%

The 16-point accuracy gap is invisible in aggregate metrics. Regulators examining this model under the RBI FREE-AI fairness requirement would immediately flag it. But the bank's own monitoring might never detect it.

Decile-Level Drift methodology:

1. Segment the applicant population by employment type, income source, or other relevant segment attributes
2. For each segment, compute the approval rate and default rate by score decile
3. Compare the segment-level decile distributions to the overall population
4. Flag segments where the approval-to-default rate ratio diverges significantly from the population mean

What this reveals:

In a study of 8 Indian NBFC clients, Claris found that gig workers consistently clustered in the 5th–7th deciles of credit score distributions, with actual default rates 40% lower than the salaried population in the same deciles. This means they were being systematically assigned lower scores than their actual credit behaviour warranted — resulting in higher rejection rates and higher interest rates for this segment.

The business case: Correctly pricing gig worker credit represents a significant untapped market opportunity. Banks that recalibrate their models to correctly score this segment access 450M+ potential customers that competitors are systematically turning away.

How Can Banks Recalibrate Models to Correctly Score Gig Workers?

Intervention 1: Income normalisation

Instead of using raw monthly income, compute:

• 12-month income trend (linear regression slope)
• Normalised income volatility (coefficient of variation)
• Minimum guaranteed income floor (10th percentile of trailing 12-month income)

Replace raw income in the model with these three normalised signals. This preserves the income information while removing the volatility signal that the model was incorrectly interpreting as risk.

Intervention 2: Alternative data integration

UPI transaction data is particularly valuable for gig workers because it captures income patterns that don't appear in traditional banking data:

• Frequency of large inbound transfers (invoice payments)
• Regularity of outbound business expense transfers
• Merchant category diversity (indicator of business breadth)
• Cash flow coverage ratio (income vs recurring expenses)

These signals are available for any customer who uses UPI — which covers virtually all economically active Indians.

Intervention 3: Segment-specific calibration

Post-hoc calibration adjusts the model's score output for specific segments to correct for systematic bias without retraining:

1. Identify the segments with systematically different actual default rates vs predicted default rates
2. Fit a calibration function for each segment (typically a Platt scaling or isotonic regression)
3. Apply the calibration at scoring time, before the threshold is applied

This approach avoids the cost of full model retraining while still achieving fairness correction within a regulatory examination timeline.