Context
A city average could hide the decision that mattered
The same city could produce different customer profiles and credit outcomes for different automotive brands. A city-level average could therefore conceal both a promising segment and a deteriorating one.
A second problem was timing. Repayment outcomes take time to mature. By the time a change became visible in delinquency, the customer mix entering the portfolio might already have shifted.
I needed a framework that could answer two questions together: where was the business opportunity, and what did the available evidence say about the quality of that opportunity?
And the answer had to lead somewhere. Regional evaluation was not simply a reporting exercise: its purpose was to inform how lending policy should differ across markets with different risk profiles and business opportunities. Each quarter ran the same cycle.
- Quarterly evaluation of brand–city segments.
- Regional classification.
- Design of differentiated policy combinations.
- Management approval of the recommendations.
- Communication of changes and expected impact to business teams.
- Improvement targets and the next quarterly review, which returns to evaluation.
Labels under each stage show who held it. The analysis and recommendations were mine; policy changes were decided through management approval; business teams were informed of the impact rather than asked to approve risk policy.
Approach
01Define the unit of analysis around the business decision
I evaluated brand–city combinations rather than relying on city aggregates alone. This made it possible to compare a local segment with its brand context and investigate changes that a broader average might obscure.
02Read each region from two sides
I designed a two-factor framework that looks at every segment from the outside and the inside: market potential and credit quality. Together they place a segment in one of four zones, and its lending contribution shows how much that zone matters to the business.
- 01Market potential factor
Built mainly on regional sales volume. Segments are tiered using the 80/20 principle, with a minimum floor so very small markets are not over-read.
- 02Credit quality factor
Centred on the quality of recently acquired customers, supplemented by medium- and long-term risk performance. Indicators are scored by interpolation and weighted by how recent they are.
- 03One-vote veto
Regions with suspected fraud, or cities with weak operating compliance, go straight into the Control zone — whatever their score.
Circles are fictional segments; their size reflects business contribution. The zones frame management questions — they are not automatic lending decisions. A flagged segment bypasses the scores and moves straight to Control.
03Bring recent customer signals into the review
I incorporated recent customer-mix indicators alongside more mature risk outcomes. This helped distinguish what had already happened in the portfolio from what might be changing in new business.
A segment with healthy historical repayment performance could still deserve attention if the share of higher-quality applicants was falling. Conversely, a single fluctuation in a small sample should not be treated as conclusive deterioration.
The credit-quality factor therefore combined three time windows, counting back from the observation point. The newest window — who is coming in now and how they rank — carries the most weight; realised risk from older cohorts is still counted, but with less. The framework also included handling rules for limited performance history, thin samples and suspected operating anomalies.
Bar height shows relative weight, not exact values. The newest window describes the customers entering the portfolio now; the oldest shows how earlier cohorts actually performed.
Technical noteSeparate the indicator from the interpretation+
Here is a brand–city segment and a review period. These expressions explain the published example, not the full internal score. The applicant cohort and quality classification must be consistent across periods for the comparison to be meaningful. Internal definitions and score weights are not reproduced.
| Signal | Useful question | Interpretation limit |
|---|---|---|
| Mature repayment performance | What happened to earlier lending cohorts? | Outcomes arrive with a delay |
| Recent customer mix | Is the incoming portfolio changing? | A mix shift is not itself a default forecast |
| Market potential | Is there a meaningful business opportunity? | Opportunity does not establish acceptable risk |
| Lending scale | Where would investigation have greater business relevance? | Size alone should not determine the response |
Policy
04Translate classifications into differentiated policy
The classifications informed a combination of policy adjustments across several stages of the lending decision — eligibility, down-payment requirements, collateral conditions and other risk controls. The change for a region was a package, not a single parameter.
Stronger-performing regions could qualify for more favourable treatment, while weaker or higher-risk regions could be subject to tighter conditions.
I translated each evaluation into regional policy recommendations — which regions, what proposed treatment and the expected business implications — for management approval. Once decisions were agreed, I explained the changes and their likely effect on conversion and business indicators to the relevant business teams. Their role was to understand and plan for the change; the policy decision itself sat with management.
- Stronger-performing regions: more favourable treatment may apply.
- Regions requiring improvement: improvement targets, then review.
- Higher-risk or flagged regions: a tighter policy combination.
- Each becomes a recommendation with its expected business impact, submitted for management approval.
Directions, not rules: a region’s treatment combined several levers, and no change took effect without management approval. Internal thresholds and the specific rules for each region are not shown.
05Give underperforming regions a path to improvement
Not every region showing elevated risk required immediate tightening.
For regions requiring improvement, we set specific targets — improving customer quality or reducing risk indicators — and assessed their progress over subsequent quarterly reviews.
If a region met its targets, the planned tightening could be avoided where justified. If performance stayed below expectations, stricter treatment could be considered, again subject to approval. This turned a classification from a verdict into a recurring decision process, connecting regional analysis, differentiated policy, business impact and performance monitoring.
- A quarterly classification identifies a region requiring improvement.
- Improvement targets are set on customer quality or risk indicators.
- After an observation period, the region is reviewed against the agreed targets.
- If targets are met, planned tightening can be avoided where justified.
- If targets are missed, tighter policy treatment is considered, subject to approval.
Illustrative path for a region that falls short. No internal thresholds are shown, and neither branch is an automatic system rule: both outcomes return to the quarterly review, and any tightening is still decided through approval.
06Turn a flag into a question for investigation
Behind a target there was usually a question. For a segment with a declining share of higher-quality applicants, the score was the beginning of the analysis. I investigated the change and worked with colleagues to direct business follow-up on possible explanations, including competition, lead allocation, delivery delays and operating practices.
These were hypotheses to check, not causes inferred from a chart alone. Subsequent reviews examined whether customer composition or risk performance moved back toward the relevant benchmark.
- A customer-mix change triggers segment diagnosis.
- Diagnosis raises two kinds of question: competition or lead allocation, and delivery or operating practices.
- Both are checked through business verification, leading to follow-up actions.
- The next-quarter review returns to diagnosis.
Analysis generated targeted questions for the business. Follow-up evidence determined which explanations were supported and what needed further review.
Diagnostic example: when a previously strong segment starts changing
A recurring pattern was a decline in the share of higher-quality applicants within a previously strong segment. The important question was whether this reflected a broader brand trend or a local change requiring investigation. The example below demonstrates the reading logic using synthetic data; it is not an individual customer, city or brand record.
| Review period | Example segment: higher-quality applicant share | Illustrative brand benchmark |
|---|---|---|
| Q1 | 32% | 31% |
| Q2 | 25% | 30% |
| Q3 | 19% | 30% |
The example segment falls while the benchmark is broadly stable. That pattern motivates local investigation; it does not establish the cause or prove that future defaults will rise.
Monitoring
Following decisions through, month by month
The quarterly framework set priorities. Between reviews, I monitored portfolio performance over time and across brands, channels and cities — looking for where risk was rising before it settled into the averages.
When a cohort showed unusual deterioration, I went down to the individual cases to understand what sat behind the number: which customers they were, how they had come in, and what happened after approval. Those findings fed into policy review.
A change was not the end of the work. After a policy or process change, I kept tracking the affected segments to see whether performance moved the way the change intended — and flagged it when it did not.
- Monthly monitoring across time, brands, channels and cities.
- An unusual cohort or segment is identified.
- Individual cases are examined.
- Findings feed policy review.
- A change is implemented.
- The affected segments are tracked after the change, which returns to monthly monitoring.
Monitoring cuts by time, brand, channel and city. Case-level review explains a deterioration before it informs a policy change; tracking continues after implementation.
Outcome
A repeatable basis for quarterly review
The documented review covered 463 brand–city combinations across three core brands. I owned the framework, indicator and scoring design, quarterly analysis, monthly monitoring and subsequent review. Business communication was shared with colleagues, while local teams carried out investigation and follow-up actions.
In subsequent reviews, I observed cases where risk returned toward the overall benchmark or customer composition recovered toward historical or brand reference levels. These observations informed ongoing review; they do not isolate the framework’s causal effect.