Context
The decision behind a down payment
A down payment is the share of a vehicle’s price that a customer pays upfront. The remaining amount is financed by the lender. Requiring more upfront cash reduces the amount financed, but can also make it harder for a customer to complete the purchase.
The policy question was therefore specific: where could we offer less restrictive terms, and where did the evidence justify keeping tighter requirements?
As part of a major policy revision, two rating tools were available: a comprehensive credit rating, which had come into the business through data testing I carried out earlier, and a repayment-capacity grade. My responsibility was to evaluate how they could support the lending decision, design the policy matrix built on them and validate its implementation. Final policy decisions required management approval.
Key insight: historical risk was shaped by the existing policy
A segment that performed well under a high down-payment requirement would not necessarily perform the same way under a lower requirement. Lowering the down payment could change the amount financed and the customer’s repayment burden. It could also change which applicants proceeded with the loan.
This meant that historical risk was not a fixed property that I could simply carry into the new policy. I needed to consider two changes separately: the expected number of completed loans and the risk assumptions for those loans.
The question was not only “Which customers receive different terms?” It was also “Who is likely to complete a loan under those terms?”
| Input | Question it answers | Role in the framework |
|---|---|---|
| Comprehensive credit rating | What does the customer’s credit profile indicate? | First axis of the matrix; differentiates credit behaviour |
| Repayment-capacity grade | How strong is the evidence of repayment ability? | Second axis; adds information beyond credit history |
| Vehicle price / financing exposure | How much financing would the proposed terms imply? | Splits each cell by financing amount |
- The comprehensive credit rating and repayment-capacity grade form the two axes of a rating matrix.
- Each cell is split by vehicle price, and its risk is compared with the overall portfolio.
- Cells are grouped into risk tiers that map to policy terms, alongside additional risk rules.
- The resulting terms go through scenario evaluation and review.
The two ratings provided complementary inputs. My work turned them into a matrix, split it by financing exposure, tiered it against the portfolio and connected it to scenario evaluation. Analytical structure only, not production rules.
Approach
01Evaluate what the new ratings added
I used historical borrowers with a full twelve instalments of repayment performance to evaluate risk differentiation. Alongside measures such as KS and AUC, I examined the intersection of old and new ratings: did the new information distinguish risk within groups that previously looked similar?
The sample retained bad outcomes more heavily than good outcomes. I therefore accounted for the sampling design when estimating portfolio risk, rather than treating the raw sample’s bad-outcome share as the population rate.
This established a basis for segmentation. It did not, by itself, determine the appropriate down-payment requirement.
Technical noteAccounting for outcome-based sampling+
Here, and are sampled bad- and good-outcome counts, and , are their sampling probabilities. This simplified expression assumes random sampling within each outcome group; more complex sampling requires the relevant observation-level weights. Production sampling ratios are omitted.
02Rebuild the policy matrix
I rebuilt the matrix behind the down-payment requirement and set the direction of evaluating the policy through it. Its two axes were rating tools: the comprehensive credit rating — a data product I had evaluated and brought in through earlier testing — and the repayment-capacity grade.
Within the matrix, I split customers further by vehicle price, so the same rating combination could carry a different requirement when the amount financed was larger. I then compared each cell’s risk with the overall portfolio and grouped cells into risk tiers, each tied to a down-payment requirement. Additional controls stayed outside the main matrix.
- 1Two rating tools as axes
- 2Split by vehicle price
- 3Compare each cell with the portfolio
- 4Group into risk tiers
Cell risk ÷ portfolio risk
- Tier 1Clearly below the portfolio → lower down payment
- Tier 2Close to the portfolio → standard requirement
- Tier 3Above the portfolio → higher down payment
- Limited historySupplementary evidence before relaxing
Hover a cell to read it.
Grades, ratios and tiers are fictional and only show the logic: two rating axes, a vehicle-price split, each cell’s risk read against the portfolio, then tiers. Hatched cells have limited repayment history and were assessed with supplementary evidence.
Some intersections contained limited repayment history. I considered supplementary evidence from other products and external credit information when assessing these groups. A low observed risk rate in a small cell was not sufficient evidence for aggressive relaxation.
I also checked whether adjacent policy bands formed a coherent structure. The resulting policy had to be explainable and implementable, as well as consistent with the risk evidence.
03Adjust for conversion and risk
Historical performance could not be transferred directly to the proposed policy. Changing a customer’s required down payment could affect both their likelihood of completing a loan and the risk profile of the loans that resulted.
Building on the method used in the previous policy review, I estimated two sets of historical adjustment factors for each down-payment tier:
- 01Conversion factor
How likely is an approved applicant to complete a loan? Derived from observed completion rates — booked loans as a share of approved applicants — in each tier.
- 02Risk factor
Are the customers who complete a loan riskier or safer than everyone approved? Estimated from customers who also held another lending product at the bank, comparing the repayment performance of booked and approved customers on that product.
- 03Transition adjustment
For each move from an old tier to a new one, the ratio of conversion factors adjusts the expected number of completed loans, and the ratio of risk factors adjusts the expected bad-loan rate — separately.
The adjusted results were then aggregated across customer segments to compare alternative policy scenarios. This carried historical behavioural and risk differences into the evaluation, rather than treating the original portfolio’s performance as an unchanged forecast.
- Approved applicants and booked loans by tier give a conversion factor.
- For customers who also hold another loan at the bank, the bad rate of booked versus approved customers gives a risk factor.
- For each old-to-new tier transition, the ratio of new to old conversion factors adjusts expected completions, and the ratio of risk factors adjusts the expected bad rate.
- The adjusted groups are combined in a scenario simulation.
Two separate questions, two separate factors, both measured per down-payment tier. Conversion adjusts how many loans a transition group is expected to complete; risk adjusts the bad-loan rate expected of them. Illustrative structure only — no production coefficients or cut-offs.
Technical noteHow the two factors are defined+
For each down-payment tier g, the conversion factor is the share of approved applicants who went on to complete a loan:
The risk factor uses customers who also held another lending product at the bank. On that product both groups have repayment history, so I could compare the bad rate of customers who completed a car loan with that of everyone approved in the same tier:
A value above one means the customers who complete a loan in that tier tend to be riskier than the approved population; below one, safer. For a move from old tier i to proposed tier j, the two ratios become the adjustments used in the portfolio calculation under Figure 4:
Production coefficients, customer counts and tier cut-offs are omitted.
04Compare old and proposed terms on the same customers
I applied the old and proposed rules to a common customer sample, then organised the results into an old-to-new transition matrix. Each cell represented a group moving from one requirement to another.
This distinguished relaxation, tightening and unchanged terms. It also made the analysis more informative than a single average down payment, which could conceal offsetting changes across different groups.
I then applied the conversion and risk adjustments from step 03 to each transition group and reweighted the results by estimated completed loans to assess the proposed portfolio. These were scenario estimates made before rollout to support the decision — not observed outcomes.
| Old requirement ↓ / Proposed → | A · Lower | B · Middle | C · Higher |
|---|---|---|---|
| A · Lower | Unchanged | Tighter | Tighter |
| B · Middle | Less restrictive | Unchanged | Tighter |
| C · Higher | Less restrictive | Less restrictive | Unchanged |
Bands A, B and C are fictional, ordered from lower to higher required down payment. Each customer belongs to one transition cell; the cell determines which scenario adjustments are relevant. A change in policy terms does not guarantee a completed loan. No customer counts or production cut-offs are shown.
Technical noteA simplified portfolio calculation+
Let denote the old band and the proposed band. Expected completions and risk assumptions are adjusted separately, then combined into a count-weighted portfolio estimate. This is a simplified explanation of the scenario logic, not a transcription of production formulas.
| Symbol | Meaning |
|---|---|
| Baseline completed-loan count represented by the transition group | |
| Conversion adjustment for expected completions, cj / ci | |
| Historical risk rate, accounting for sampling where relevant | |
| Risk adjustment to the historical bad rate, kj / ki | |
| Count-weighted scenario risk rate across the resulting portfolio |
The two adjustment factors answer different questions. More completions can change the weight of a group even when its risk assumption stays unchanged. Adjusting the risk assumption can change the portfolio estimate even when the number of completions stays fixed. Neither adjustment identifies a causal effect on its own.
Technical noteWorked example: why reweighting matters+
| Transition group | Baseline completions | Completion factor | Estimated completions | Historical risk | Risk factor | Scenario risk |
|---|---|---|---|---|---|---|
| Less restrictive | 100 | 1.20 | 120 | 1.00% | 1.20 | 1.20% |
| Unchanged | 200 | 1.00 | 200 | 2.00% | 1.00 | 2.00% |
| Tighter | 100 | 0.80 | 80 | 4.00% | 0.90 | 3.60% |
| Total | 400 | — | 400 | — | — | — |
All numbers are synthetic and unrelated to the reported project outcome.
Total expected completions remain unchanged in this example, yet the portfolio risk estimate changes because the mix and group-level assumptions change. The example illustrates the arithmetic; it is not evidence that a relaxation necessarily improves portfolio risk.
05Translate the analysis into executable rules
I wrote the business requirements and independently reproduced the proposed logic in SQL. I compared customer-level outputs with the strategy engine’s batch-test results, investigating differences in decision paths, derived fields and decimal precision.
The rollout expanded in stages. Monitoring included preliminary approval rates, approved down-payment distributions, telephone-verification rates, collateral requirements and implementation stability. Early operational monitoring helped check the rollout; it could not replace mature repayment outcomes.
- Policy proposalSegment-level evidence, scenario assumptions and decision trade-offs
- RequirementsTranslate the agreed framework into executable business rules
- Batch validationCompare independent SQL outputs with engine decisions
- Phased rolloutInspect financing terms and operational indicators as exposure expands10%50%100%
- Ongoing monitoringSeparate early operating signals from maturing credit performance
The rollout progressed from 10% to 50% to full deployment. These percentages describe deployment scope, not a measured conversion effect.
OutcomeReported outcome
More applicants qualified for lower upfront requirements
For the same applicants, the share approved for a down payment of 20% or less was 8 percentage points higher under the new policy than under a simulation of the old policy.
This result demonstrates a change in the financing terms made available. It does not establish the size of a realised conversion lift or a long-term reduction in credit risk. The scenario estimates supported the decision before rollout; the same-applicant comparison described the policy’s effect on approved terms.