Why Your Credit Bureau Score Is a Rearview Mirror
The bureau score is the single most successful risk artifact in Indian lending. It compresses years of borrower behaviour into one integer, it is cheap to pull, and it is remarkably well-calibrated for the population it was built on. None of that is in dispute here. The argument is narrower and more uncomfortable: for a large, growing, and commercially important segment of India's credit-eligible population, the score is structurally a rearview mirror. It tells you where the borrower has been with high fidelity and tells you where they are going only to the extent that the future resembles the past — and only for borrowers whose past was densely recorded in the first place.
The answer is not to throw the score away. The score is a high-precision feature. The answer is to stop treating a single backward-looking signal as a decision, and to start governing multi-signal decisions. This post is about why that distinction matters technically, not rhetorically.
1. What a Bureau Score Actually Encodes
A CIBIL or Experian score is a function over a structured feature vector derived almost entirely from tradeline data: the accounts a borrower holds, their sanctioned limits, outstanding balances, and a month-by-month repayment string. From that substrate, the scoring model materialises a handful of feature families. Payment history is encoded as days-past-due (DPD) buckets — 30/60/90+ DPD flags and their recency. Credit utilisation is the ratio of revolving balance to sanctioned limit, computed per account and aggregated. Credit mix captures the secured-versus-unsecured composition. Length of history is the age of the oldest tradeline and the average age across accounts. And enquiry behaviour is the velocity of hard pulls in a trailing window, which proxies credit hunger.
Every one of those features is a function of realised credit events. That is the crux. The model is, in effect, estimating P(default within 12–24 months | S), where S is a state vector summarising past behaviour. The architecture leans on a first-order Markov assumption: that the future is conditionally independent of the deep history given the current state. Formally, P(X_{t+1} | X_t, X_{t-1}, …, X_0) ≈ P(X_{t+1} | S_t), where S_t — the score — is treated as a sufficient statistic for everything that came before. This is what makes the score compact and stable. It is also what makes it backward-looking by construction: no tradeline means no feature mass, and no feature mass means the state vector is mostly undefined.
That undefined region is the thin-file problem, and it should be stated precisely. Let a borrower's history be a vector h ∈ ℝ^d over the d coordinates the scoring model expects. For a thin-file borrower, the effective dimensionality d_eff ≪ d, because most coordinates are null. The score function f: ℝ^d → [300, 900] still returns a number, but the posterior variance on the underlying default probability is large: Var(P(default | h)) is high precisely because f is extrapolating from a sparse input. A 720 produced from fourteen years of seasoned tradelines and a 720 produced from one eight-month personal loan are not the same epistemic object, even though they print identically.
The scale of the affected population is not marginal. World Bank data puts India's credit-eligible cohort (ages 18–80) at roughly 1,036 million as of December 2024, yet only about 27% — around 277 million — actively use formal credit. By TransUnion CIBIL's accounting, roughly 451 million Indians have limited or no formal credit footprint, and approximately 99 million consumers became new-to-credit in a single year (FY24). These are not borrowers the score evaluates badly. They are borrowers the score cannot evaluate well, because the state vector it depends on does not yet exist.
2. Where Bureau Scores Systematically Fail
The failure is segment-structured, not random, which is what makes it a governance problem rather than a noise problem.
Gig workers carry income volatility the tradeline layer was never designed to see. The bureau records obligations and repayments, never income. A delivery rider with lumpy weekly earnings may show utilisation spikes that look like distress but are in fact liquidity-timing artifacts. Self-employed borrowers transact substantially in cash and through current accounts; their actual revenue and repayment capacity flow through channels invisible to the bureau. Rural borrowers sit in regions of low formal-credit penetration, producing sparse or absent tradelines. And new-to-credit individuals are the canonical cold-start case: f(h) is either undefined or high-variance, and roughly 99 million people enter this state every year.
These segments expose a deeper, quieter defect: score compression. The map from latent creditworthiness to the 300–900 scale is calibrated on a population overwhelmingly composed of salaried urban borrowers. Calibration is not segment-invariant. A given score bucket does not carry a constant default rate across segments — only across the segment that dominated the training distribution.
Consider a 680, which in CIBIL's own tiering sits exactly at the subprime/near-prime boundary. Treat the true 12-month default probability as a random variable across the population of borrowers who all map to 680, since the score is a lossy compression and many distinct borrowers collapse onto the same value.
For a salaried urban borrower, P(default | score = 680, segment = salaried) is roughly unimodal and tight — concentrated around an expected PD of about 3.2%, with most mass between 2.5% and 4.5%. The score was fit on this population, so here it behaves as a near-sufficient statistic and does what it claims.
For a gig or self-employed rural borrower, P(default | score = 680, segment = gig) is generated by a far more heterogeneous set. Some reached 680 with thin-but-clean files and strong cash buffers (true PD ≈ 2%). Some show utilisation spikes that are income-timing noise masking genuine capacity. Some are authentically fragile (true PD ≈ 11–14%). The result is a bimodal, right-skewed distribution: a marginal expected PD near 6.5% — roughly double the salaried case — and a materially fatter tail.
So E[PD | 680, gig] ≈ 2 × E[PD | 680, salaried], with the tails diverging even more. Now apply a single policy: approve if score ≥ 680, expecting pool PD ≤ 4%. That rule is true for the salaried pool and false for the gig pool. Holding the score threshold constant silently holds risk non-constant across the population the rule governs. You either over-reject creditworthy gig borrowers by tightening globally, or you under-reserve and mis-price by not. The constant-score cutoff encodes an assumption of segment-invariant calibration that the data refutes. That is the governance failure — the decision rule's realised risk is not constant across the borrowers it is applied to, and nothing in the score surfaces this.
3. India's Alternative Data Infrastructure
What changed is that the missing signals are now retrievable through regulated, consent-based plumbing. The Account Aggregator (AA) framework has moved from pilot to backbone: per Department of Financial Services figures, more than 2.88 billion financial accounts were enabled for sharing and roughly 284.6 million accounts linked as of 31 March 2026, with AA-enabled lending crossing ₹1.6 lakh crore in FY25. Licensed NBFC-AAs operate as consent-managed "blind pipes," and the in-scope Financial Information Provider categories now include banks, NBFCs, AMCs, insurers, and the GST Network.
This unlocks three forward-looking data streams. Bank transaction (CASA) data exposes salary or income credits, EMI debit regularity, balance trajectories, and overdraft behaviour. GST return data exposes declared turnover and, crucially, filing regularity across GSTR-1 and GSTR-3B. Digital payments (UPI) expose transaction frequency and the merchant-versus-P2P composition of flows. Unlike tradelines, these describe cash flow and income — the variables the bureau structurally omits.
From these, specific, defensible features can be engineered:
monthly_income_volatility_index— the coefficient of variation of income-classified credits over a trailing 12 months: σ_income / μ_income. High values flag earnings instability that a clean tradeline can mask.overdraft_utilisation_rate— time-weighted drawn overdraft against sanctioned OD limit (or the share of days the account sits in overdraft), a direct read on chronic liquidity stress.GST_filing_consistency_score— on-time filings divided by due filings over the trailing eight periods, recency-weighted; late or missed GSTR-3B filings are a leading indicator of business strain.cash_flow_stability_ratio— a buffer/runway measure such as minimum month-end balance divided by mean monthly outflow, or median net inflow over its standard deviation.
The discipline that separates this from data maximalism is signal validation. The only question worth asking is whether an alternative feature adds predictive information beyond the bureau score, not in general. Compute incremental AUC, ΔAUC = AUC(bureau + alt) − AUC(bureau-only), and test it with DeLong. But the headline number is a trap if read at the population level: the gain must be stratified by segment and should concentrate in the thin-file and underserved strata — exactly where the score has high posterior variance. A flat ΔAUC across all segments usually means the alternative feature is re-encoding something the bureau already captures.
This is where SHAP interaction analysis earns its place. Examine the interaction values Φ_{ij} between alternative features and the bureau score itself. If cash_flow_stability_ratio contributes most precisely where the bureau score is thin or low-confidence, that confirms it is filling the information gap rather than duplicating it. Conversely, a large main effect that overlaps the score's own attribution with little interaction is a redundancy signal — drop it. Orthogonal information, demonstrated per segment, is the bar.
4. Why Governance Is the Missing Piece
Here is the failure mode the industry is walking into: bolting alternative signals onto the model without governing them does not fix the rearview-mirror problem — it adds three new ones.
First, disparate impact through non-traditional proxies. UPI merchant patterns, transaction geolocation, and device metadata can act as proxies for sensitive attributes — region, gender, community — even when those attributes are never an input. A model can be formally attribute-blind and still produce a segmented adverse-outcome distribution. The DPDP Act 2023 governs the personal-data dimension of this; the lending-fairness dimension is on the lender. Second, opacity: gradient-boosted models over a couple of hundred cash-flow features are not natively explainable, yet RBI's Digital Lending Guidelines expect borrowers to be told why they were declined. Third, absence of recourse: a borrower rejected on cash_flow_stability_ratio needs a truthful, actionable adverse-action reason, not a black-box verdict.
RBI's FREE-AI report (the Framework for Responsible and Ethical Enablement of Artificial Intelligence, released 13 August 2025, with seven guiding sutras and 26 recommendations across six pillars) frames exactly this tension — "People First" and bias detection alongside enablement. Governance is what operationalises it. Concretely, governed multi-signal decision-making has three technical components.
Cluster-stratified threshold calibration. Segment the population into behavioural clusters — a SOM-plus-K-Means pass over the combined feature space works well — and calibrate the decision threshold per cluster to hold a constant target PD, rather than imposing one score cutoff everywhere. This attacks score compression at its root: you equalise E[PD | approve] across clusters instead of equalising the score. The 680 problem from Section 2 dissolves, because the gig cluster and the salaried cluster are admitted at thresholds tuned to the same realised risk, not the same integer.
SHAP-based local explainability per segment. Generate per-decision attributions, surface the top contributors as adverse-action reasons, and validate within each cluster that those reasons are stable across similar applicants and actionable by the borrower. Explainability that only holds on average is not recourse.
CFDI monitoring across signal combinations. Disparate impact is not a one-time pre-launch check; it drifts as the signal mix and thresholds shift in production. Continuous CFDI monitoring tracks fairness across signal combinations over time, flagging when a particular feature interaction begins producing a segmented outcome — feeding directly into the kind of incident-reporting discipline FREE-AI envisions.
The bureau score does not leave the model. It stays exactly where it belongs: a high-precision feature for the population it was built to evaluate. Governance is simply what makes it safe to extend the decision vector to the other 451 million — the borrowers for whom the rearview mirror was never going to be enough.