Hardship and collections.

Which collections and hardship actions actually cause payment, for which customers? Luminnovate measures it on your own history, before anything changes in production.

The problem every collections team has

Actions go to customers for reasons, so comparing those who got an action with those who did not measures the reasons as much as the action. Calls go to those likely to pay, so calls look effective; hardship plans go to those already behind, so plans look harmful. Neither comparison says what the action did. Measuring that needs a method that corrects for who was chosen for each action, checked against a known answer, and it is the one we bring.

What Luminnovate does

Which contacts cause paymentThe effect of each contact by channel and segment, so effort goes where contact changes the outcome rather than to customers who would pay anyway.
Which arrangements holdArrangements shaped by the cause of hardship, which is already recorded, with the cure and recovery expected from each.
Value a strategy before trying itA proposed strategy valued on your past records, with an interval, before any change in production.
Identify hardship earlyEarly identification of customers at risk of hardship, from payment behaviour judged against each customer's own history and peers, with false alerts held to the team's capacity.

Hardship timelineOne customer, twelve monthsDays each repayment was late

Illustrative example with invented data.

Who this is for

Lenders' collections and hardship teams

Where in the cycle: Early arrears, hardship requests, and outreach before arrears
The question: Which customers to contact, how, and which hardship arrangement will hold
What wastes money: Contacting customers who would pay anyway; plans that break
What regulators expect: Spot early signs of financial difficulty and respond (Banking Code of Practice; ABA financial difficulty guideline)

Collection agencies and debt buyers

Where in the cycle: Late stage and charged off debt, placed or purchased ledgers
The question: Which accounts to work, which channel and offer, and what a ledger will return
What wastes money: Working accounts that will never pay, or would have paid unprompted
What regulators expect: Fair, flexible treatment of hardship and vulnerability (ASIC RG 96 and the ACCC debt collection guideline)

What is measured, and what a pilot measures

Impact Sandbox for hardship plans: an intensive plan for lasting hardship recommended, flagged as assumed until a pilot with a holdout measures it; the side by side; and the measured share of plans that went bad, 64.7% for lasting hardship against 31.8% for temporary shocks. — From Luminnovate's Impact Sandbox. Bad rates measured on real lending data from an open research dataset; dollar figures from illustrative assumptions, not a client's.

Two cases

Results from Luminnovate's work, labelled by source. Unlike the illustration above, these figures are not invented.

8.1Real email campaign data from an open randomised experiment

The method gets the known answer

estimated effect, against a true 8.2; comparing who got the email with who did not said 11.8

What was observed: A public experiment emailed customers at random, so the true effect is known. We made it look like a targeted campaign, emailing the most promising customers first: for them, the email added 8.2 website visits per 100 customers.
What a generic approach says: Compare customers who got the email with those who did not: 11.8, overstating the effect by 45%.
What the engine read: A method that corrects for who was chosen for an action estimated 8.1, within 0.5% of the truth averaged over 10 targeted samples; single samples ranged from 6.9 to 9.3, and the 90% interval contained the truth in all 10. Run where there was no effect, it found none.
What it meant for the decision: The effect of a contact can be measured on records where contacts were targeted, which describes every collections book. The outcome here is a website visit, not a payment: no public dataset records collections contacts, and your own data closes that gap.

Read the full case

64.7%Real lending data from an open research dataset

Which hardship plans hold

of plans for lasting hardship went bad, against 31.8% for temporary shocks

What was observed: 4,433 hardship plans, granted either for a temporary shock such as a natural disaster or for lasting hardship.
What a generic approach says: Treat every hardship plan alike.
What the engine read: Plans for lasting hardship went bad 64.7% of the time, against 31.8% for temporary shocks, comparing borrowers in similar circumstances before the plan.
What it meant for the decision: The reason for a plan deserves its own treatment path. Whether a plan changes the outcome is a separate question, answered by a measured trial.

Read the full case

What is shown, and what is not yet

Claim	Status
The method recovers known effects and refuses questions it cannot answer	Shown on a public randomised email experiment, where the outcome is a website visit
The cause of hardship predicts outcomes at the same arrears	Shown on real consumer loans (US, 2017); the like for like comparison's balance check was flagged as not fully balanced
Early warning flags most first misses at low false alerts	Shown on a real bank's loans, out of time
A contact strategy's effect on collections data	Not yet: no public dataset records contacts; your data closes this
Whether a hardship arrangement itself improves outcomes	Not yet: needs monthly payment history, which your data has

Checks built in, and what we will refuse

No estimate where customers who got an action have no comparable customers who did not: we say so rather than guess.
Every estimate shows how well the comparison is balanced, a test run where there should be no effect, and a 90% interval.
Decisions stay with your people; the tools rank, estimate and explain.

Where it is used: Banks and lenders, Energy and telco retailers, collection agencies and debt buyers
One case: The method gets the known answer (Real email campaign data from an open randomised experiment)
Trust: Regulatory context

A proof of concept: four weeks, your data, no change to production. You share 12 to 24 months of anonymised accounts, contacts and hardship arrangements; we return the effect of each action by segment, hardship triage by cause, and one alternative strategy valued on your records.

Book a conversation