LuminoTrial designs trials on real lending decisions, runs them against a matched control group, and reports the measured effect in both risk and commercial terms. Where a result is not conclusive, it says so.
Discuss LuminoTrial →Any decision applied to some customers and not others can be trialled: a cutoff, a price, a limit, a treatment, a channel rule.
Refer bands, decline rules, verification requirements and exception authority.
Rate cards, limit assignment, renewal increases and pre-approved offers.
Contact timing, channel, hardship pathways and early arrears handling.
Whether a new signal, metric or memo actually changes what underwriters decide.
Three parts, used in order. A risk or strategy lead can run all three; a modeller can go underneath any of them.
Choose what is changing and on whom: a cutoff, a price, a limit rule, a collections treatment. The module then picks the measurement method the book and the calendar actually allow, rather than the one you hoped for.
It will not start a trial that cannot detect the effect you are looking for.
One register across underwriting, pricing, collections and growth, with an owner against each trial. Trials that share a population are flagged and held, so two teams cannot quietly contaminate each other’s results.
The register becomes the evidence that changes were tested before they were adopted.
The measured effect, the uncertainty around it, the guardrail metrics that did not move, and what the result is worth at your volumes. Where the result is not conclusive, it says so.
A null result is an outcome you paid to learn, not a failed project.
New applications split 50/50 at decision time. Both groups live through the same market and the same policy drift, so nothing else has to be held constant.
Live for everyone from a date, measured against a segment left alone. Needs a clean record of everything else that changed in the window.
Applications either side of the current threshold are alike apart from which side they fell. That discontinuity already sits in your history.
Illustrative example. Method availability depends on volumes and what else is changing.
Illustrative example. Overlap checks run on the population, not the team.
Roughly 340 additional accounts cured per year before they reach 60 days, against no measurable increase in complaints or early repayment.
Illustrative example. Effects and volumes depend on the book and the design.
The same discipline applies whether the trial confirms the change or kills it.
Hypothesis, population, groups, metrics, guardrails, sample size and minimum run.
design / refer_band_cutoffWhat is changing, on whom, what would count as success, and how long the trial must run before the answer means anything.
Whether the trial can detect the effect you are looking for, agreed before it starts.
Every trial across teams, with owner, status, population and overlap checks.
trials / allFour running, one at read out, one held because it shares a population with the cutoff trial already in flight.
Whether two teams are quietly contaminating each other’s results.
Measured effect with uncertainty, guardrail metrics, and the commercial translation.
readout / contact_timingCure rate up 4.2 points with the interval stated, complaints unmoved, translated into accounts cured per year at your volumes.
Whether the change is worth adopting, and what it is worth in money.
What was adopted, for which population, and what was explicitly not tested.
decision / adopt_broker_onlyAdopted for the broker channel; direct channel and low balance accounts untested and left unchanged.
Whether the change can be defended later, to your board or your regulator.
Illustrative specimens. Trials, metrics and measured effects depend on your book, your volumes and the design agreed at the outset.
Not every question needs a trial. Where the evidence already exists in history, LuminoEvidence answers it faster and at no operational risk.
Something you can apply to some customers and withhold from others without treating anyone unfairly.
Sufficient flow through the population to detect the size of effect that would change your mind.
A result that arrives within a timeframe you can wait for, not one that takes three years to mature.
Someone who has committed to act on the answer, in either direction, before the trial begins.
The honest answers, including the ones that send you somewhere other than a trial.
Anything you can apply to some customers and withhold from others without treating anyone unfairly: refer bands and decline rules, verification requirements, rate cards, limit assignment, renewal increases, collections contact timing, and whether a new metric in the memo changes what underwriters decide. Each object carries the unit it must be randomised on, which is not always the application. Broker facing changes split at broker level because brokers compare terms; memo changes split at team level because underwriters learn.
Usually more useful, not less. At mid market volumes a fast outcome like approval rate or conversion can be trialled live in a few months, while a rare outcome like 90 day arrears cannot be detected this side of two years. So the method adapts: where the live book cannot answer the question, the answer comes from your history instead, most often from the discontinuity your existing cutoff already creates. The power calculation is done before anything starts, and it is allowed to conclude that a trial is not worth running.
It breaks any before and after comparison, which is why it is handled first. Policy versions, rule sets and override behaviour are read out of the decision records themselves, so changes are detected rather than self reported. The result is a change ledger showing what moved and when, including the changes nobody declared. Trials are then measured against a window you can actually defend, and the ledger is useful on its own for anyone who has been asked what changed in credit policy last quarter.
Yes, and the register exists mainly to make that safe. Trials are tracked independently, and any two that touch the same population are flagged and held until one clears. Without that, two teams testing different changes on the same borrowers will each read the other’s effect as their own, which is the most common way a trial programme quietly produces wrong answers.
For a retrospective measurement: decisions with their scores or policy version, the action taken, and the outcome that followed, for a period long enough for outcomes to have matured. For a live trial: the ability to apply the change to a defined group and to record which group each case fell into. Nothing needs to be moved into a new system, and a first measurement can run on an extract rather than an integration.
LuminoEvidence asks whether a signal predicts an outcome in data you already hold. LuminoTrial asks whether a change you make produces a different outcome. The first is about knowledge; the second is about cause. Where the history can answer the causal question, LuminoTrial reads it from the history and you avoid the operational cost entirely.
Your risk, portfolio or strategy lead designs and reads trials without needing a statistician in the room, and every assumption and check sits underneath for your modelling and validation teams to challenge. Deployment, calibration and the first two or three trials are run alongside your team so the method is understood before it is relied on.
Name one change you have argued about. We will tell you whether your book can measure it live, or from history.