Cohort and vintage delinquency monitoring for a US bank
Concurrent and ever-delinquency matrices, vintage curves and early-warning indicators for Axos Bank — turning a monthly regulatory reporting grind into a portfolio view that surfaces deterioration while it is still early.
+25–30%
delinquency prediction accuracy
−50%
regulatory reporting time
Context
At Quinte Financial Technologies I worked on portfolio risk analytics for Axos Bank, a US commercial bank. The recurring question in any lending book is deceptively simple: is credit quality getting better or worse?
Calendar-month reporting answers that badly. A book that is growing quickly looks healthy on a headline delinquency rate simply because new, not-yet-delinquent accounts dilute the denominator. Deterioration hides inside growth.
Problem
Two problems, one structural and one operational.
The structural problem: the reporting could not distinguish “our lending is getting worse” from “we lent more recently”. A single portfolio-level delinquency percentage moves for reasons that have nothing to do with credit quality — mix, growth rate, seasonality. Acting on it means acting on noise.
The operational problem: regulatory reporting consumed a large share of every monthly cycle. Assembling the same figures by hand each month left little time for interpretation, and manual assembly is exactly where reporting errors enter.
Approach
Stop measuring the portfolio by calendar date, and start measuring it by account age.
Vintage cohorts. Group accounts by origination month, then track each cohort’s performance against months-on-book. Every cohort is compared at the same age, so an August cohort at month six is compared with a March cohort at month six. Deterioration in underwriting quality becomes directly visible as later cohorts tracking above earlier ones at the same maturity — usually months before it reaches the headline number.
Concurrent versus ever delinquency. These answer genuinely different questions and are routinely conflated:
- Concurrent — the share of a cohort sitting in a given delinquency bucket right now. This is the current state of the book, and it improves when borrowers cure.
- Ever — the share of a cohort that has at any point touched that bucket. This never improves, and it is the honest measure of underwriting quality, because it counts borrowers who went delinquent and recovered.
A cohort with low concurrent and high ever delinquency is one where collections are working but the credit decision was not. The opposite pattern means the opposite. Reported as a matrix — cohort against months-on-book, for each DPD bucket, in both concurrent and ever form — that distinction stops being invisible.
Early-warning indicators. Leading signals that move before an account books a formal delinquency, so intervention can happen while it still changes the outcome.
What I built
- Concurrent and ever-delinquency matrices by DPD bucket across cohort and months-on-book, with cohort lifetime tracked side by side.
- Vintage curves comparing origination cohorts at equal maturity, making underwriting drift legible early.
- Early-warning indicator monitoring feeding proactive collections rather than post-hoc reporting.
- Automated regulatory reporting on SQL Server surfaced through Tableau, replacing recurring manual assembly.
- Rule-based governance monitoring translating credit and compliance policy into executable validation checks — the same policy-to-code discipline as my lending work, applied to controls.
Outcome
- Delinquency prediction accuracy improved 25–30%, largely because cohort-and-age framing compares like with like where calendar aggregates could not.
- Regulatory reporting time cut by 50% across monthly cycles, moving effort from assembly to interpretation.
- Alongside the governance monitoring: audit scores up 15% and risk incidents down 20%.
Tools
SQL, SQL Server, Tableau, Excel.
Where else this applies
Any book where accounts age and can deteriorate: subscription churn by signup cohort, accounts-receivable ageing, insurance claims development, warranty and returns curves. Cohort-versus-calendar reasoning is the transferable idea — compare like with like by age, not by date.