In most of the world, the people you most want to lend to are invisible to a traditional bureau. Thin file, no file, first-time borrowers, they are the engine of digital and emerging-market lending, and a bureau score either does not exist for them or simply returns ‘insufficient history.’ A naïve lender reads that as ‘decline’ and walks away from the entire market. The opportunity is to score those customers well using the data you actually have — and the craft is doing it without quietly building a model that looks brilliant in development and falls apart in production.
This is the full build: where the predictive signal actually lives, why cash-flow data is the highest-value source you are probably underusing, the pipeline stage by stage, the two places alternate-data models silently go wrong, and what explainability and governance now require. It is written for the person who has to ship the model and defend it, not just admire the ROC curve.
The data spectrum: signal isn’t evenly spread
‘Alternate data’ is a spectrum, and the predictive signal is concentrated, not uniform. Cash-flow data, income regularity, balance volatility, days at zero before payday, the ratio of essential to discretionary spend, carries the most signal by a distance, because it is the closest thing you have to a direct read on ability to pay. Telco, utility, and rent payments prove payment discipline. Platform and e-commerce history is strong if you own the platform. Device and behavioural footprints are weaker, noisier, and far more privacy-sensitive, and should be treated with care.
Figure 1: Cash-flow data carries the most signal; device/footprint the least.
Source | Signal | Example features | Watch-out |
Cash-flow / bank transactions | Very high | Income regularity, balance-to-zero days | Needs consent + clean parsing |
Telco, utility & rent | High | On-time payment history | Coverage varies by market |
Platform / e-commerce | High* | Purchase & repayment history | *Only if you own the platform |
Bureau thin-file | Medium | Sparse trade lines, inquiries | Often too thin alone |
Device & behavioral | Low–med | App usage, digital footprint | Noisy, privacy-sensitive |
Start with cash-flow data. For a thin-file borrower, the bank statement out-predicts the bureau.
The reason to start with cash-flow data is simple: it shows actual ability to pay in near-real-time, not a stale summary of past borrowing. The craft is turning raw, messy transactions into features a model can use, classifying income, detecting salary regularity, measuring how often the balance hits zero before payday, flagging overdrafts and gambling spend.
Figure 2: The same statement, turned into predictive features.
The build pipeline, where alt-data models go wrong
The pipeline looks like any scorecard, but two stages, leakage/bias and governance, are where alternate-data models quietly fail. Build it as a loop, not a line: monitoring feeds back into features and retraining, because alt-data drifts faster than traditional bureau data.
Figure 3: Build it as a loop, not a line.
Stage | What you do | Trap to avoid |
Target & sample | Define ‘bad’, set performance window | Cold start — needs a champion phase |
Feature engineering | Build strong cash-flow features | Volume over signal |
Leakage & bias check | Drop decision-time-unavailable features | Selection bias; skip reject inference |
Build (explainable) | Scorecard or monotonic model | Black box you can’t defend |
Validate & govern | Out-of-time, by-segment, fairness | Ship-and-forget; no monitoring |
The bias you can’t see: leakage and reject inference
Two silent killers live in the middle of that pipeline. The first is leakage, features that won’t actually be available at the moment of decision, or that encode the outcome. A model trained on leaked features posts spectacular development metrics and then collapses in production. The second is selection bias: your through-the-door population was shaped by your old policy, so a model trained only on approved-and-observed accounts is learning from a skewed slice of reality. Reject inference, estimating how rejected applicants would have performed, is how you correct for it. Skip it and you will systematically misjudge the very segment you are trying to expand into.
Explainability, fairness, and governance
Because regulators increasingly treat credit scoring as ‘high-risk,’ explainability and fairness are not optional polish, they are table stakes. You must be able to produce a concrete decline reason for every applicant and demonstrate that the model does not create disparate impact on protected groups. That favours scorecards and monotonic, constrained models over opaque ones, and it makes documentation, independent validation, and ongoing monitoring part of the build rather than an afterthought.
Check | What it is | Why it matters |
Out-of-time validation | Test on a later period | Proves it holds, not just fits |
By-segment performance | Check across key segments | Catches hidden weakness |
Disparate-impact testing | Outcome fairness across groups | Regulatory + ethical necessity |
Stability monitoring (PSI) | Track population shift | Models drift; catch decay early |
The traps that catch people
Trap | The fix |
“More data = better model” | Prioritise a few high-signal features |
No clean consent | Permissioned, documented data only |
Ignoring drift | Monitor stability; retrain |
Black-box model | Use explainable / monotonic models |
No reject inference | Correct for through-the-door bias |
If you only do one thing
Don’t wait to build a full alternate-data model. Add a few cash-flow features as overlays and reason codes to your existing scorecard first. It is the fastest path to measurable lift, it builds the data pipeline and the consent plumbing you will need anyway, and it gives the team the muscle to do the full rebuild with confidence later. Lift first, elegance second.
Next: IFRS 9 / ECL, explained for the people who run it.
Views are my own and do not represent my employer.