Who Gets What: Segmenting a Delinquent Portfolio | HL Hunt
Who Gets What: Segmenting a Delinquent Portfolio
Almost every collections operation segments by balance and days past due. Those are the fields that always exist, which is why they're used — not because they predict what will work. Two accounts at the same balance and the same stage can be a customer who forgot and a household in serious difficulty, and the treatment that resolves one is the wrong treatment for the other. Segmenting by what actually differs changes recovery and cost simultaneously, and establishing that it works requires a holdout rather than a comparison between segments.
What you'll learn
Why the standard segmentation is weak
| Dimension | Tells you | Doesn't tell you |
|---|---|---|
| Balance | What's recoverable | Whether it will be |
| Days past due | How long it's been | Why |
| Product | What the obligation is | The person's situation |
| Risk score | Likelihood of loss | What would change it |
Read the right column. None of the standard dimensions tells you what treatment would work — they tell you what the account is, not what the person needs.
Days past due is the most misleading because it looks causal. Per our cure analysis, most accounts at an early stage recover — so an early-stage segment is mostly people who are fine, and treating it as a risk category over-treats the majority.
And risk score, despite being the most sophisticated input, answers the wrong question. Per our measurement analysis, predicting who will default is different from predicting who will respond to a treatment — and the second is what a segmentation decision requires.
Five reasons an account isn't paying
The segmentation that matters, and each needs a different response.
| Cause | What resolves it | What's wasted on it |
|---|---|---|
| Oversight | A reminder | Anything more |
| Timing | A date change | Pressure |
| Hardship | An arrangement | Demands for full payment |
| Dispute | Resolving the substance | Collection activity entirely |
| Decision not to pay | Escalation | Gentle reminders |
The right column is where the money goes. An operation applying one sequence to all five is applying the wrong treatment four times out of five — and the wrong treatment isn't neutral, it costs money and generates complaints.
What distinguishes them:
- Oversight: good payment history, no prior difficulty, responds immediately to first contact.
- Timing: per our timing analysis, a recurring pattern of late-then-paid, frequently around the same point in a month.
- Hardship: per our disclosure analysis, signals on contact, or deteriorating behaviour across obligations.
- Dispute: per our dispute analysis, it may arrive as a complaint rather than as a formal dispute — and must be recognized as one anyway.
- Decision: capacity present, contact avoided, no difficulty indicated.
The second is the cheapest win available and almost nobody segments for it. A customer who pays every month, four days late, doesn't need collections — they need a due date that matches when money arrives, per our arrangements guide, and that removes them from the portfolio permanently.
Cannot pay versus will not pay
The distinction the whole discipline turns on, and it's genuinely hard.
These look identical in the data — no payment, limited engagement, similar aging — and respond to opposite treatments. Pressure on someone who cannot pay produces complaints and no recovery; an arrangement offered to someone who simply won't pay is a discount for non-cooperation.
What helps distinguish them:
- Behaviour across other obligations, where visible — someone paying everything else has capacity.
- Engagement pattern. Per our inbound analysis, someone who calls you has a problem rather than an intention.
- Response to a genuine offer. Offering a realistic arrangement is itself diagnostic — those who cannot pay engage with it, those who will not pay mostly don't.
- What changed and when, since a clean history followed by an abrupt stop suggests an event.
The third is the practical test and it's underused. An operation that offers an arrangement early learns which segment an account is in and resolves the cannot-pay cases in the same step — which is cheaper than any amount of investigation.
And per our framing analysis, the distinction is worth handling carefully. Operations tend to assume refusal where hardship is more likely, because refusal justifies escalation and hardship requires accommodation.
Treatment cost as a constraint
Where balance actually belongs in the scheme — not as the primary division but as a limit.
Per our cost analysis, treatment cost is largely fixed per contact while balance varies, so below a threshold an expensive sequence loses money regardless of outcome.
Which gives a two-dimensional structure:
| Small balance | Large balance | |
|---|---|---|
| Oversight / timing | Digital, one touch | Digital, one touch |
| Hardship | Self-service arrangement | Agent, with authority |
| Dispute | Route to resolution | Route to resolution |
| Decision not to pay | Write off or digital only | Escalate |
The reason column drives the approach and the balance column drives the channel, which is a cleaner structure than either dimension alone.
And the bottom-left cell is the honest one. A small balance where someone has decided not to pay is not worth pursuing expensively — the arithmetic in our cost analysis says so plainly, and an operation working those accounts through an agent sequence is choosing a loss.
Proving it works
The methodological point, and it's where most segmentation claims fail.
Segments show different outcomes because they contain different people. A segment you treat gently and that recovers well may recover well because it contains people who were going to pay anyway — which per our measurement analysis is the selection problem in its standard form.
Comparing segments tells you nothing about whether the treatment caused the difference. Only a holdout within a segment does:
- Within each segment, randomly withhold the treatment from a sample.
- Compare treated against held-out, within the segment.
- The difference is what the treatment added.
- Net it against treatment cost.
- Keep the holdout running, since a one-off measurement expires.
Operations running this for the first time reliably find that some treatments add little — per our cure analysis, a substantial share of early-stage recovery happens without any intervention, and the operation has been crediting itself for it.
The objection that a holdout means forgoing recovery is worth answering: on a small random sample, for a bounded period, the cost is trivial and the information determines how the entire portfolio is worked. An operation unwilling to run one is choosing not to know whether its largest cost line does anything.
Boundaries
The practical problem with any segmentation, per our segmentation analysis.
Accounts just either side of a cutoff are nearly identical and receive different treatment. Which matters for two reasons:
- Fairness. Two similar customers treated very differently on the basis of a threshold is hard to defend, particularly where the difference is between an arrangement and escalation.
- It's where the segmentation is most likely wrong. The cases nearest a boundary are the ones the rule classifies least confidently.
What to do:
- Avoid cliff edges where possible — graduate treatment rather than switching it abruptly.
- Examine boundary accounts specifically when reviewing, since that's where misclassification concentrates.
- Allow reclassification when new information arrives — an account segmented as refusal that discloses hardship should move immediately, per our disclosure analysis.
- Make the reclassification path work in both directions and easily, since a segment assignment made on thin evidence shouldn't be sticky.
The third is the one to build first. A segmentation without a reclassification path locks accounts into a treatment chosen before anyone knew anything — which per our disclosure analysis is exactly the propagation failure that produces the worst conduct outcomes.
Building a workable scheme
- Start with the reason dimension, using what you can observe.
- Add balance as a channel constraint, not as a primary split.
- Define a genuinely distinct treatment per segment — if two segments get the same treatment they're one segment.
- Check each has enough volume to measure.
- Build the reclassification path.
- Set up holdouts within each.
- Report net of cost by segment, per our cost analysis — never gross.
- Review quarterly.
Items three and four together are the discipline that keeps schemes usable. Most elaborate segmentations contain segments that receive materially identical treatment or that are too small to evaluate — both are complexity without information, and a smaller scheme with measured results outperforms an intricate one nobody can assess.
And per our testing analysis, changes to a segment's treatment should be tested against the existing one rather than deployed on reasoning — the arguments for a new approach are always persuasive and frequently wrong.
Why it stops fitting
A segmentation is fitted to a population at a moment, and both move.
What changes:
- The portfolio's composition, as origination criteria or channel mix shift — per our channel analysis, that changes who is in each segment without changing any rule.
- Economic conditions, which shift the balance between hardship and other causes.
- Channel effectiveness, as people's responsiveness to contact methods changes.
- Your own treatments, which alter the population reaching later stages.
The last is the subtle one. Improving early-stage treatment means the accounts reaching later stages are the harder residual — so a later segment's performance can deteriorate because an earlier treatment improved, which reads as failure and is the opposite.
What to monitor:
- Segment sizes over time.
- Net recovery by segment, against the holdout.
- Migration between segments.
- Complaint rate by segment, per our complaint analysis — a segment generating disproportionate complaints is probably misclassified rather than difficult.
Segment by what determines the response
HL Hunt AI Debt Collection assigns treatment by observed cause rather than by balance and aging alone, supports reclassification on new information, runs holdouts within segments, and reports recovery net of treatment cost.
Frequently asked questions
They describe the account rather than the situation, and the situation determines which treatment works. They're used because they're always available.
By why the account isn't paying — oversight, timing, hardship, dispute, or decision — with balance as a constraint on treatment cost rather than the primary split.
A holdout within each segment. Comparing segments proves nothing, since they contain different people and would differ anyway.
As many as have genuinely distinct treatments and enough volume to measure. Two segments with the same treatment are one segment.
Key takeaways
- Standard dimensions tell you what the account is, not what the person needs — and treatment is chosen for the second.
- Five causes need five responses; one sequence is the wrong treatment for most of the portfolio.
- A customer who pays four days late every month needs a date change, not collections — and that removes them permanently.
- Offering a realistic arrangement early is the cheapest way to distinguish cannot-pay from will-not-pay, and it resolves one of them.
- Comparing segments proves nothing; only a holdout within a segment shows what the treatment added.
- A later segment can deteriorate because an earlier treatment improved — that reads as failure and is the opposite.
Know which of your treatments is doing anything
Get started with HL Hunt AI Debt Collection for cause-based segmentation, cost-aware channel assignment, built-in holdout measurement, and net-of-cost reporting by segment.
This guide is educational and does not constitute legal or compliance advice. Collection conduct requirements, contact rules, dispute handling obligations, and requirements relating to consumers in vulnerable circumstances apply to every segment regardless of how accounts are classified, and vary by jurisdiction. Consult qualified counsel about the requirements applicable to your operation, and your compliance function before differentiating treatment.