Where Applicants Come From Changes What They Are

Two applicants, identical files, identical scores. One arrived through a comparison marketplace, one through an existing relationship. They will not perform the same, and no variable in either file explains why. The channel selected them before you saw them — and because channel effects sit outside the data most models consume, the consequence shows up as unexplained underperformance or as apparent model decay that no amount of model work will fix. A great deal of what gets diagnosed as a failing model is a change in who is applying.

How channels select

Channel	Selects for	Typical effect
Existing relationship	People you already know	Best, and you have behavioural data
Direct brand	People seeking you specifically	Generally favourable
Comparison marketplace	Shoppers with alternatives	Depends on where you rank
Broker or intermediary	Whoever the broker places	Depends on their incentives
Embedded at point of sale	People buying something	Different population entirely
Prescreened offers	Those who responded	Response is itself a signal

Row six is the subtle one. A prescreened population is selected by you, and then the subset that responds selects itself — and responding to a credit offer correlates with needing credit, which is information the file doesn't contain.

Row one is why our renewal analysis finds behavioural data dominates application data. And row five connects to our distribution analysis: embedded credit reaches people who weren't seeking credit at all, which is a genuinely different population rather than the same one arriving differently.

Adverse selection specifically

The mechanism worth naming precisely, because it's the one that costs money.

Adverse selection occurs when a channel systematically sends you applicants other lenders declined or priced worse. Where it arises:

• Marketplaces where you rank behind others. If applicants see a ranked list and you're fourth, you receive the people the first three didn't want — a population filtered by other lenders' judgments, with no record of the filtering.
• Where your criteria are looser on a specific dimension, which per our selection analysis means you concentrate exactly the risk your competitors are avoiding.
• Where your process is faster and applicants who need money urgently self-select toward you.
• Where a broker places declined applications with whoever will take them.
• Where you're the last resort in any sense.

The first is the one most lenders underestimate. Being competitive on a marketplace is desirable; being marginally competitive is worse than being absent, because you receive volume filtered by everyone ahead of you and priced as though it weren't.

And the effect is invisible in the file. An applicant declined by three lenders looks identical to one who applied to you first — which is why this shows up as unexplained underperformance rather than as a risk factor anyone can point at.

Mix shift versus model decay

The diagnostic distinction that saves the most wasted effort.

Portfolio performance deteriorates. Two possible causes:

	Model decay	Mix shift
What changed	The relationship	Who is applying
Within-channel performance	Worse	Unchanged
Channel mix	Unchanged	Shifted
Right response	Model work	Channel management

Row two is the test and it takes one query. If each channel is performing exactly as it always did and the portfolio has deteriorated, the model is working correctly and more volume is coming from a worse source.

Per our monitoring analysis, population stability tracking catches this — but only if channel is among the dimensions monitored, and in many operations it isn't, because channel is treated as a marketing attribute rather than a risk one.

The wasted effort is real: a team rebuilding a model to fix a mix problem will produce a model that performs no better, because the model was never the issue. And per our validation analysis, a validation that doesn't examine population composition will conclude the model has degraded.

Measuring it

Straightforward once the data exists, and the data has to be captured at the time.

1. Tag every application with channel and source — this cannot be reconstructed later, which per our decision record analysis is the general rule for anything you'll want to analyze.
2. Record sub-source — which marketplace, which broker, which campaign, since aggregate channel hides wide variation within it.
3. Compare performance by channel at equivalent score bands. Equal scores performing unequally is the signature.
4. Compute a channel adjustment — how many points of score difference the channel effect is worth.
5. Track mix over time, as a monitored input.
6. Track acceptance rates by channel, since a channel whose applicants you decline heavily may be sending you a filtered population.
7. Watch for sub-source drift within a channel, which is where problems usually start.

Item three is the core analysis and most operations have never run it. It requires only channel tags and outcomes, and it frequently reveals differences large enough to change pricing or cutoffs by channel.

Item seven deserves attention because it's where deterioration begins. A channel performing well in aggregate can contain one sub-source that has changed substantially, and the aggregate hides it until the volume grows.

Launching a new channel

Where the risk is concentrated, because a new channel is an unmodelled population.

Per our cold start analysis, existing performance evidence doesn't transfer to a source you've never lent into — the model was fitted on a population this one isn't.

What to do:

• Start small, at volumes where a bad outcome is affordable.
• Apply a more conservative cutoff initially, and relax it on evidence rather than on hope.
• Hold a random sample below your cutoff if you can, per our reject inference analysis — otherwise you'll never learn where the right cutoff is for this channel.
• Wait for real performance before scaling, which takes longer than commercial pressure allows.
• Watch early indicators — first payment defaults surface a bad population far sooner than mature performance does.
• Agree volume ramps in advance with the partner, tied to performance.

The fifth is the practical early warning. A channel sending a materially worse population shows it in first-payment behaviour within weeks — long before any vintage matures — which makes it the metric to watch during a ramp.

When an intermediary chooses

The case needing separate treatment, because there's a party with their own incentives between you and the applicant.

A broker or partner decides where to place an application, and that decision is made in their interest rather than yours.

What follows:

• They learn your criteria and send what fits, which sounds efficient and means you receive applications selected to pass rather than to perform.
• They may present applications differently to improve acceptance.
• They place declines elsewhere, so your acceptance rate tells you little about the population they see.
• Their compensation drives behaviour, per our incentive analysis — paid on volume, they send volume.
• Their conduct may be attributed to you, which is a compliance matter as well as a risk one.

Managing it:

• Monitor performance by intermediary individually, never in aggregate.
• Watch for applications clustering just above cutoffs, which indicates criteria learning — and is the same structuring pattern our controls guide describes in another setting.
• Verify independently rather than relying on what's supplied, per our application fraud analysis.
• Tie compensation to performance, not only to volume.
• Reserve the right to stop, and use it.

The second is the most diagnostic single check available. A distribution of applications piling up immediately above a threshold didn't occur naturally, and it's visible in a histogram.

Whether to model it

A genuine question with arguments both ways.

For including channel: it captures real information about selection that file variables don't, it's predictive, and excluding a predictive variable means the model is wrong in a way you know about.

Against:

• It correlates with geography, targeting, and intermediary relationships in ways that can raise fair lending questions.
• It's gameable. An intermediary who learns that one route prices better will route through it.
• It's unstable — channel composition changes, so the relationship the model learned may not hold.
• It's hard to explain in a reason code, per our explainability analysis. "You applied through the wrong website" is not a reason an applicant can act on, which is a problem where specific reasons are required.

Many lenders handle it outside the model instead — channel-specific cutoffs, channel-specific pricing, and channel monitoring — which captures the effect without embedding it in an applicant-level score. Per our segmentation analysis, that's the segment-versus-variable question, and the answer here is frequently segment.

The explainability point is the one I'd weight most heavily. A variable that can't produce an actionable reason sits badly in a decision that requires giving reasons.

Fair lending

Not an afterthought, because channel correlates with things that matter.

Why it needs scrutiny:

• Channels correlate with geography, which per our testing analysis correlates with protected characteristics.
• Marketing targeting is a choice you made, so a channel's composition partly reflects your own decisions.
• Channel-specific pricing or cutoffs produce different treatment for equivalent applicants, which needs a documented justification.
• Intermediary conduct may be attributed to you, including how they steer applicants.
• Prescreen criteria determine who receives offers at all, which is an access decision before any application exists.

What to do:

1. Test for disparities by channel as well as in aggregate.
2. Document the business justification for any channel-based difference, at the time you make it.
3. Review targeting for distributional effects.
4. Monitor intermediary steering — where they route applicants and on what basis.
5. Include channel in fair lending review rather than treating it as marketing.

The third is the one most easily overlooked, because targeting sits with marketing and the fair lending function looks at decisions. But who receives an offer is a decision, and it's made before anything the credit function sees.

Frequently asked questions

---

This guide is educational and does not constitute legal or compliance advice. Fair lending requirements, the treatment of channel and marketing variables, adverse action reason requirements, and responsibility for intermediary conduct vary by product, institution, and jurisdiction. Consult qualified counsel and your compliance and model risk functions before adopting channel-based criteria or pricing.