When a lender sells loans instead of holding them, it stops paying for its own mistakes. Economists call this the originate-to-distribute model. The cost shows up years later, as loans from particular origination years ("vintages") default more than their credit scores and loan-to-value ratios predicted. By the time those defaults are visible, regulators are already behind. This is the regulatory lag, or pacing problem. This note proposes a forecasting model that is interpretable by construction. It splits loan-level default risk into four additive parts: a borrower-observables part, a loan-age curve, a calendar-time shock, and a vintage effect specific to each origination channel. The quantity of interest is the gap in vintage effects between loans from brokers, who never fund the loan they arrange, and loans from retail lenders, who do. The model reads this gap from the first months of each vintage's repayment history, years before cumulative losses would show it.
The core idea is that hidden underwriting deterioration lives in the vintage axis of loan performance data, not in calendar time. Standard decomposition models cannot separate it from macroeconomic shocks. A channel contrast can. If the model can recover that contrast early and in an auditable form, it gives regulators a number that is available while the risky loans are still being written.
Core Claim
Loan default data has three entangled time axes. These are loan age (default risk rises and then falls as a loan "seasons"), calendar period (recessions and house-price shocks hit all loans at once), and vintage (the underwriting standards of the origination year). A classic difficulty is that period equals vintage plus age. Any linear trend can therefore be shifted freely between the three components, which is the age-period-cohort identification problem. Our claim is that a difference in vintage effects between two origination channels is identified, as long as the age curve and the period shock are shared across channels. The arbitrary linear trend cancels in the difference.
We model the log default hazard of loan $i$ at age $a$, calendar month $t$, vintage $v$, and channel $c$ as
$$\log \lambda_i(a,t) = f(x_i) + g(a) + h(t) + k_c(v),$$
where $x_i$ are credit attributes recorded at origination, $f$ is a constrained neural score, $g$ is a smooth age curve, $h$ is a calendar shock allowed to jump, and $k_c$ is a smooth vintage curve per channel. The decoupling index is $D(v) = k_{\text{broker}}(v) - k_{\text{retail}}(v)$. It measures how much worse broker-originated loans from vintage $v$ perform than retail loans with identical observable attributes, the same age, and the same calendar environment. We expect $D(v)$ to rise through vintages in which originate-to-distribute incentives intensified. We also expect it to be estimable from short seasoning windows, because it is pooled over many loans and needs no long default history.
Grounding in Prior Work
The motivation comes from the originate-to-distribute literature. A HOLISTIC VIEW TOWARDS THE BREAKING-OUT OF THE GLOBAL FINANCIAL CRISIS. argues narratively that securitization removed the incentive to screen borrowers. Bubble in the Real Estate Market and Its Risks to Macroeconomy: A Case Study of the United States argues that bundling loans into securities hid credit risk rather than reducing it. Two studies document the same pattern in Chinese fintech. The Analysis of Asset-Backed Security Backed by Online Consumer Loan shows consumer-loan securitization used to replenish lending capital. Correction to: Lessons from the rise and fall of Chinese peer-to-peer lending traces 60-fold growth in an unregulated sector before regulators shut it down. All four are retrospective and qualitative. None produces a quantity measurable while the risky loans are still being originated, and that is the gap we target.
The method draws on neural decomposition. N-BEATS: Neural basis expansion analysis for interpretable time series forecasting restricts sub-networks to polynomial and Fourier forms so that each forecast is an exact sum of components. We reuse that basis-restriction idea for $g$ and $k_c$. CoST: Contrastive Learning of Disentangled Seasonal-Trend Representations for Time Series Forecasting and Probabilistic Decomposition Transformer for Time Series Forecasting separate trend from seasonality along a single time axis. Our setting needs a decomposition across three axes linked by an exact linear identity. Those methods do not address this, and naive single-axis decomposition would confound vintage quality with recession timing.
The research file also flags a constraint. Smooth bases handle sharp regime breaks badly. We therefore keep $h$ as an unconstrained per-month effect rather than a smooth basis, since that is where shocks belong.
Two pieces of interpretability work frame the design. Interpretable Selective Learning in Credit Risk routes most borrowers to logistic regression and sends only the 1.4–2.3% of cases with non-linear interactions to a neural network. This supports our choice of a mostly additive $f$ with limited interaction capacity. Towards Transparent Time Series Forecasting argues that shape-level explanations serve decision-makers better than pointwise saliency. Our curves $D(v)$ and $h(t)$ are explanations of exactly that shape-level kind.
On regulatory lag, The Bottleneck in AI Governance and Oversight: Evidence from 1,419 State Bills quantifies delay inside the legislative process. Government Interventions to Avert Future Catastrophic AI Risks frames the evidence dilemma: regulators either act before there is data or wait until harm is done. We address the evidence side of that dilemma. The question is how early defensible evidence of incentive failure exists, not how fast legislatures act on it.
Research Program
Data. We build the model on the Freddie Mac single-family loan-level data that Transforming Credit Risk Analysis: A Time-Series-Driven ResE-BiLSTM Framework for Post-Loan Default Detection uses for post-loan default detection. That data records origination attributes and monthly performance. It also records an origination-channel field distinguishing retail lending from broker and correspondent channels; we assume this field is available and verify it at data intake.
All of these loans were eventually sold to Freddie Mac. The contrast is therefore not securitized versus held. It is between originators with more skin in the game (retail lenders who fund and service the relationship) and those with essentially none (brokers). This is a narrower but cleaner test of incentive decoupling. It avoids comparing loan populations that differ in many unobserved ways.
Fitting. The model is fit as a discrete-time hazard with these components:
- $f$: an additive network with a small interaction budget.
- $g$ and $k_c$: B-spline bases with smoothness penalties.
- $h$: free monthly effects.
Early estimation. For each vintage we re-estimate $D(v)$ using only the first $m$ months of performance, for $m$ of 6, 12, and 24. This mirrors what a supervisor would have known in real time.
Uncertainty. Uncertainty comes from a block bootstrap over originating sellers, because lender-level clustering is the main source of dependence.
Fintech replication. As a secondary, smaller study, we attempt the same construction on a consumer or peer-to-peer loan dataset with a channel or guarantee-status contrast. Candidates include the data underlying Multi-view GCN for loan default risk prediction. That paper warns of severe class imbalance and more than 50% missing attributes. We treat this arm as exploratory and report it only if the shared-age-curve assumption passes the diagnostic described below.
Baselines
Naive reference. The do-nothing comparison is the raw early-payment-default gap: the difference between broker and retail delinquency rates at month $m$ for each vintage, with no adjustment. This is what a supervisor could compute today from a spreadsheet. To count as an improvement, $D(v)$ must predict the eventual 36-month channel default gap out of sample better than this raw gap at the same $m$.
Strongest existing default predictor. This is ResE-BiLSTM from Transforming Credit Risk Analysis: A Time-Series-Driven ResE-BiLSTM Framework for Post-Loan Default Detection. It is a residual attention encoder with a bidirectional recurrent network and SHAP attributions. It reports the highest accuracy in 38 of 44 Freddie Mac cohorts and the highest F1 score in 39 of 44. We run it on identical loan samples and identical time splits, with the same hyperparameter search budget in GPU-hours as our model. Channel is included as an input, and we read a vintage-level channel attribution from mean SHAP differences.
Our model does not need to beat ResE-BiLSTM on loan-level accuracy; we accept a small loss. It must produce a channel signal that is (a) earlier, (b) more stable across random seeds, and (c) non-zero only when channels truly differ, as checked on placebos. A black box with post-hoc SHAP is not designed to deliver those three properties.
Classical regression. We also run a classical age-period-cohort hazard regression with fixed effects for age, month, and vintage-by-channel, plus linear credit controls. This is the econometrician's answer. If it matches our model on every check, the neural components add nothing and we will say so.
Ablations. Two ablations isolate the key ingredient:
- Letting the age curve $g$ differ by channel. This tests whether identification really comes from sharing it.
- Replacing the per-month $h$ with a smooth basis. This tests the claim that shocks must be allowed to jump.
Single-axis decomposition. Finally, we apply a seasonal-trend decomposition in the style of CoST to the aggregate monthly channel default-gap series. This shows what goes wrong when calendar time is the only axis.
Why not just…?
Why not just track early-payment defaults by vintage and channel? This is the most serious threat, because industry already does it. Its weakness is that early defaults mix three things: composition (brokers may serve lower-score borrowers), calendar timing (a vintage originated just before a downturn defaults early regardless of underwriting), and age (vintages observed at different ages are not comparable). Our model removes the first two by conditioning and the third by the shared age curve. If, after adjustment, the raw gap predicts the eventual gap just as well, the contribution collapses to "the simple metric suffices." That would be a useful negative result, and the first baseline comparison is designed to show it if true.
Why not just add channel to a strong black-box model and read SHAP? SHAP values for a feature in a recurrent model with attention depend on the model's learned interactions. Nothing forces them to separate a vintage effect from a calendar effect. When a channel's share of originations changes over time, SHAP can attribute calendar shocks to channel. The placebo test below checks directly whether ResE-BiLSTM's channel attribution reacts to fake channel splits.
Why not just run a fixed-effects age-period-cohort regression? This is a strong alternative. It shares our identification argument, since differences across channels cancel the linear indeterminacy. What it lacks is flexible handling of non-linear credit interactions in $f$, which Interpretable Selective Learning in Credit Risk finds matter for a small fraction of borrowers. It also lacks smoothing across vintages, which matters when $m$ is small and each vintage is noisy. If the regression matches us at every $m$, we would recommend it to regulators as the simpler tool. The neural version earns its place only if it wins at short seasoning windows.
Proposed Experiments
Planted-signal test. We first validate on semi-synthetic data. We take real Freddie Mac attributes and timing, simulate defaults from known $g$, $h$, and $k_c$ curves, and plant a broker-specific vintage ramp of chosen size and onset. The question is how many months of seasoning each method needs to detect the onset at a fixed false-alarm rate. This answers the identification question before any real-data claim.
Main real-data test. On real data we use rolling-origin evaluation. For each vintage $v$ we estimate $D(v)$ from the first $m$ months of performance, then score it against the realized 36-month adjusted channel default gap. Metrics are rank correlation and mean absolute error across vintages. We report how these change with $m$ and compare them against all baselines.
Lead time against dated events. We record the first vintage at which $D(v)$'s bootstrap interval excludes zero. We measure the lead time from that date to (a) the calendar month when the cumulative default gap for that vintage becomes statistically visible without adjustment, and (b) dated public supervisory actions, which we compile ourselves from official sources. This lead time is our quantitative measure of regulatory lag on the evidence side.
Placebo checks. Two placebos are decisive:
- We split retail loans randomly into two pseudo-channels; $D(v)$ must stay at zero.
- We shift the channel label by one vintage; the signal must degrade.
A further diagnostic tests whether channel-specific age curves fit significantly better. If they do, identification fails for that dataset and we report it.
Sanity check on general forecasting. To confirm that the basis restrictions do not cripple ordinary forecasting, we run the age-and-period part of the architecture on M4. There, N-BEATS: Neural basis expansion analysis for interpretable time series forecasting reports beating the competition winner by 3% and the statistical benchmark by 11%. We expect to be close to N-BEATS rather than to beat it.
References
- A HOLISTIC VIEW TOWARDS THE BREAKING-OUT OF THE GLOBAL FINANCIAL CRISIS.
- Bubble in the Real Estate Market and Its Risks to Macroeconomy: A Case Study of the United States
- The Analysis of Asset-Backed Security Backed by Online Consumer Loan
- Correction to: Lessons from the rise and fall of Chinese peer-to-peer lending
- N-BEATS: Neural basis expansion analysis for interpretable time series forecasting
- CoST: Contrastive Learning of Disentangled Seasonal-Trend Representations for Time Series Forecasting
- Probabilistic Decomposition Transformer for Time Series Forecasting
- Interpretable Selective Learning in Credit Risk
- Towards Transparent Time Series Forecasting
- The Bottleneck in AI Governance and Oversight: Evidence from 1,419 State Bills
- Government Interventions to Avert Future Catastrophic AI Risks
- Transforming Credit Risk Analysis: A Time-Series-Driven ResE-BiLSTM Framework for Post-Loan Default Detection
- Multi-view GCN for loan default risk prediction