Loss Forecasting: Knowing What a Loan Book Will Lose Before It Loses It

Every lender that has grown quickly has had the same experience: the portfolio loss rate looked excellent, growth continued, and then losses climbed sharply without any obvious change in underwriting. Nothing went wrong. The book simply got old enough to default. Credit losses follow a maturation pattern, which means a portfolio's observed loss rate is largely a function of its average age — and a fast-growing book is structurally young. This is the single most common way lenders misread their own performance, and it's arithmetic rather than judgment. This guide covers vintage analysis, loss curves, roll rates, and how to forecast losses on a book you're still building.

Why growth hides losses

The mechanism, stated plainly, because everything else follows from it.

Loans don't default evenly across their lives. Defaults are near zero in the first months, rise to a peak somewhere in the first year or two depending on product, then decline as surviving borrowers demonstrate they can pay. A loan's probability of loss depends heavily on how long it has been outstanding.

Now consider a portfolio doubling annually. At any moment, a large share of balances were originated recently — sitting in the period where losses barely register. The portfolio loss rate, computed as losses divided by outstanding balances, is suppressed by the denominator being full of loans too young to have failed yet.

The consequences:

• A growing book always looks better than it is, and the faster it grows the better it looks.
• Losses appear to spike when growth slows, because the average age rises and the young-loan dilution stops. Nothing about underwriting changed.
• Portfolio-level loss rates cannot tell you whether recent underwriting is better or worse than past underwriting, which is the question that actually matters.

Which produces the central discipline: never evaluate credit performance at the portfolio level. Evaluate it by cohort, at the same age. Everything below is machinery for doing that.

Vintage analysis

Vintage analysis groups loans by origination period — typically monthly or quarterly cohorts — and tracks each cohort's performance by months on book rather than by calendar date.

The structure:

• Each vintage is a fixed set of loans originated in one period.
• Performance is measured at each age — cumulative losses at month three, month six, month twelve, and so on.
• Cohorts are compared at the same age, which is the entire point.

What this makes visible that portfolio metrics can't:

• Whether underwriting is improving or deteriorating — if the January cohort is losing more at month six than the prior June cohort did at month six, something changed, and you know when.
• Which changes worked. A policy change in March shows up as a difference between pre-March and post-March vintages, which is as close to a controlled experiment as lending offers.
• Channel and segment differences, when vintages are cut by acquisition source, product, or score band — frequently revealing that a blended number was concealing one deteriorating channel offset by another improving.
• Where the book actually is in its maturation, which drives the reserve.

The practical requirement is data discipline: every loan needs its origination date, its cohort assignment, and its performance tracked from origination. Lenders who didn't build this from the start face a painful reconstruction, which is an argument for doing it early even at small volume.

Loss curves and maturation

Plotting cumulative loss against months on book produces a loss curve — the shape describing how a cohort's losses accumulate over its life.

Typical characteristics:

• A near-flat early period where almost nothing has defaulted.
• A steepening section as the cohort passes through peak risk.
• A flattening tail as surviving borrowers prove out.
• An asymptote approximating lifetime loss.

The shape varies substantially by product. Short-duration products mature fast; longer-term installment products have extended curves; revolving products don't have a single origination-driven curve at all, since exposure changes continuously — which is why revolving portfolios need the behavioral monitoring in our early warning guide alongside vintage work.

Two features worth understanding:

Early-cycle defaults mean something different. Losses in the first few months frequently indicate fraud or gross underwriting failure rather than borrower distress — someone who never intended to pay, or who could never have paid. A vintage with elevated early losses is a fraud signal, not a credit signal, per our application fraud guide.

The curve shifts with the environment. Cohorts originated into deteriorating conditions perform worse than identical cohorts originated into good ones, which means the curve is not purely a function of underwriting — the cycle dynamics our credit cycle analysis tracks. Separating underwriting effects from environment effects is the hard part of the discipline and requires honesty about which you're observing.

Roll rates for the near term

Loss curves project the lifetime. Roll rates project the next few months, and they're more accurate for that horizon than any long-range model.

A roll rate is the percentage of balances moving from one delinquency bucket to the next in a period — current to 30, 30 to 60, 60 to 90, and onward to charge-off. Chaining them produces a near-term loss projection: apply the roll rates to current bucket balances and you get an estimate of what will charge off over the next several months.

Why they're valuable:

• They're leading. A change in the current-to-30 rate shows deterioration forming before it accumulates into visible balances — the timing advantage our collections metrics guide emphasizes.
• They're decomposable. A rise in early roll rates and a stable late-stage cure rate is a different problem from the reverse.
• They connect underwriting to collections. Late-stage roll rates are substantially a function of collections effectiveness, which means part of your loss forecast is a servicing variable rather than a credit one.

The caution: roll rates assume stable behavior. They project the recent past forward, so they miss turning points — which is exactly when you most want a forecast. Use them for the near term and loss curves for the lifetime, and treat divergence between the two as information.

Projecting incomplete vintages

The practical problem: most of your book hasn't finished maturing, so most of your loss is unobserved. The standard approach is extrapolation against seasoned cohorts.

1. Establish a reference curve from cohorts that have fully matured.
2. Measure where an incomplete vintage sits relative to the reference at the same age — for example, losing 20% more at month nine than the reference did at month nine.
3. Scale the reference curve by that ratio to project the incomplete vintage's lifetime loss.
4. Adjust for known differences in mix, policy, or environment between the cohorts.
5. Attach uncertainty that widens with how little of the vintage has matured.

Two failure modes to avoid. Over-reading young cohorts — a vintage at month three has almost no signal, and reacting strongly to it is reacting to noise. And assuming the reference curve still applies when the environment or the product has changed, which is how forecasts stay confidently wrong.

Forecasting without history

A new product has no vintages, and the honest answer is that your first forecast is a hypothesis.

The approach:

• Borrow a curve shape from the closest comparable product — similar duration, similar borrower profile, similar structure. The shape transfers better than the level.
• Anchor with early observations, updating as the first cohorts season.
• Hold wide uncertainty bands and say so internally, since a single-point forecast on a new product conveys false precision that then drives decisions.
• Originate deliberately. The first cohorts are the most informative data you will ever have about this product, which argues for controlled volume and careful tracking rather than fast growth on an assumed curve.
• Update rapidly, treating each month of new performance as substantial information rather than confirmation.

The failure this prevents is specific and common: scaling a new product quickly on an assumed loss rate, and discovering the actual rate only when the first cohorts mature — by which time you've originated many times the volume at the wrong price. This is the mechanism behind most credit blowups in new lending products, and the fix is sequencing rather than sophistication.

Reserving on expected loss

Reserving asks what the book will lose over its life, not what it has lost — which is a different question from the one delinquency answers.

The components:

Element	What it captures
Probability of default	How likely each exposure is to fail, by segment and age
Exposure at default	How much will be outstanding when it does
Loss given default	What you don't recover, net of collections and any collateral
Forward-looking adjustment	Expected conditions over the remaining life, not just historical experience

Two things worth emphasizing for smaller lenders. Recovery assumptions are frequently the weakest input, because they require knowing your own recovery rates by age — the decay curve in our collections framework — and most lenders assume rather than measure them. A loss given default assumption that's ten points wrong moves the reserve substantially.

And the forward-looking component is where judgment enters and where it should be documented. An expected-loss framework requires a view about conditions, that view is a choice, and the record of why you chose it is what makes the reserve defensible later — the same documentation discipline our monitoring guide applies to models generally.

Signals your forecast is wrong

Forecasts drift. The signals that yours has:

• Actual versus expected divergence by vintage — the primary control, tracked monthly, at the cohort level rather than in aggregate.
• Roll rates moving away from loss curve implications, meaning your near-term and lifetime views disagree.
• Early-cycle losses rising, which usually signals fraud or a channel problem rather than credit deterioration.
• Cure rates falling, which means delinquency is converting to loss at a higher rate than assumed — frequently a collections capacity issue rather than a credit one.
• Recovery rates below assumption, which raises loss given default across the whole book.
• Mix shift — the population you're originating has changed, which invalidates a curve built on the prior mix.
• Score distribution drift at constant approval rates, which means the same policy is now approving a different population, per our policy guide.

The single most useful control: an actual-versus-expected report by vintage, reviewed monthly by someone with authority to change policy. A forecast nobody compares against outcomes is a number, not a forecast.

Making it operational

1. Assign every loan a vintage at origination, and never overwrite it.
2. Build the loss curve from whatever mature cohorts you have, and rebuild it as more season.
3. Track actual versus expected by vintage monthly, with named ownership.
4. Cut vintages by segment — channel, product, score band — because blended curves conceal exactly the variation you need.
5. Run roll rates weekly for near-term visibility.
6. Measure recovery rates rather than assuming them.
7. Document the forward-looking view and revisit it on a schedule.
8. Report growth-adjusted loss rates to anyone making decisions, so nobody mistakes a young book for a good one.

That last item is the one with the most organizational value. A board or investor seeing a portfolio loss rate on a fast-growing book is seeing a number that will get worse for reasons unrelated to management quality. Presenting vintage curves alongside it — and explaining the maturation effect before it arrives — is the difference between a foreseeable trend and a credibility problem.

Frequently asked questions

---

This guide is educational and does not constitute accounting, audit, or regulatory advice. Allowance methodologies and disclosure requirements are governed by applicable accounting standards and supervisory expectations; consult qualified accounting and audit professionals regarding your reserve.