The data was always the wall
Building a credit model has never really been about the algorithm. It has been about the data you needed before you could start. A PD model wants enough realised defaults to learn from, and defaults accumulate slowly, over years and across a full cycle. For an established mortgage or consumer book, that history exists, but for a new product, a young originator, or an asset class no one has lent into at scale, it does not. You either wait until enough borrowers default to model the ones who will, which is a strange thing to have to wait for, or you do not underwrite the segment at all.
That was the wall. Not which gradient-boosted tree won, but whether you had enough matured cohorts to build anything a risk committee would sign. It is why the newest and fastest-growing corners of asset-based finance have been the hardest to model, and why capital has stayed on the sidelines of the very asset classes where borrower demand is growing quickest.
Tabular foundation models move that wall. A model pre-trained on millions of datasets arrives already knowing what tabular structure looks like, so it needs far less of your own history to produce a calibrated score, and it produces it in minutes instead of months. Once the data requirement falls that far, the question changes. It stops being “do we have enough defaults to model this” and becomes “which data is worth acquiring, and what can we now underwrite.” The model turns into the commodity. The context becomes the product: the data you feed the model, the signals you engineer into it, the asset classes you can bring into scope, and the governance that lets a committee trust what comes out. We have spent years building that layer for asset-based finance, and it is already in production.
What actually changed
A tabular foundation model like TabPFN is a transformer pre-trained on millions of synthetic tabular datasets. It has already learned what tabular structure looks like, so it does not learn your problem by training on it. It reads your labelled rows as context, the way a language model reads a prompt, and predicts the rest in a single forward pass. No gradient descent on your data, no hyperparameter search, no retraining when the tape updates.
Two consequences matter for credit. These models are strongest exactly where conventional machine learning is weakest, on small datasets, which is what most asset-based portfolios are. And they return calibrated probabilities without the separate calibration step a tree-based PD model usually needs, which matters when the output feeds a regulatory risk classification or a capital calculation rather than just a ranking.
The performance is no longer something you take on faith. On TabArena, the public benchmark that now anchors the field, the top single models are all tabular foundation models. The best tuned gradient-boosted tree sits well outside that leading group, with the first tree appearing only around seventh place. A few years ago a tuned GBDT was the undisputed default. That is a fast inversion.
The limits are just as real, and worth stating plainly, so we read like people who deploy these rather than people selling them. The current generation is built for small-to-medium data, roughly tens of thousands of rows, though the newest models are pushing that ceiling and some now reach far larger sets. Benchmark rankings shift with evaluation protocol and ensembling, not architecture alone. And a transformer pre-trained on synthetic data is harder to defend in front of a model-risk committee than a decade-proven XGBoost. We use them as a fast experimentation layer and inside ensembles, and we do the validation work rather than assume it away.
Cheap models change what you can underwrite
The speed is easy to file under convenience, a shortcut that trims a data-science sprint, and that reading misses what it actually does. When fitting and evaluating a candidate model falls from a multi-month project to something you can run in an afternoon, the economics of the decision above change. A data source that could never have justified six months of engineering is suddenly worth a one-week pilot, and an asset class that never had the default history to support a model of its own becomes something you can price.
That matters most in the places conventional modelling has always struggled. BNPL portfolios, revenue-based finance, public-administration receivables, solar lease pools, these are young asset classes with thin files and short histories that have never run through a full downturn, the segments where a gradient-boosted tree has too little to learn from and where, until recently, the honest answer to “can we model this” was usually no. They are also, not by coincidence, among the fastest-growing corners of asset-based finance, which puts the modelling gap directly on top of where the market is expanding.
A foundation model changes that answer without loosening your risk discipline. You still validate it, still stress it against a downturn, still defend the thing to a committee, but you reach that conversation in days instead of after a year of accumulating losses to learn from. The consequence is strategic rather than technical. The set of deals a firm can credibly underwrite widens, and it widens toward exactly the segments that were off-limits because they were new or too hard to proxy.
If the model is the commodity, the edge is the context
If a pre-trained model gives everyone the same fast, calibrated baseline, then the baseline is not where anyone’s advantage lives, and everything shifts to what you feed the model and how you shape it. This is the part that “just download TabPFN” quietly skips. A foundation model hands you a score the moment you point it at data, but the quality of that score is decided almost entirely by the inputs, and the inputs are where the real work has always been.
For us that work has two sides. The first is knowing which signals move risk in a given asset class and getting them into a state a model can use, and it rarely involves the exotic data people imagine. It runs through cash flows from bank accounts and payment processors, which have lifted predictive accuracy by more than 40% in some consumer-credit applications, and the operational data already sitting inside origination and servicing systems: how long underwriting took, how often a borrower called collections, the seasonality of repayment. It also runs through the context that never reaches the loan tape, and this is where the map redraws itself with every book. For receivables it is obligor payment behaviour, for revenue-based finance it is platform revenue, for solar it is production readings, and the macroeconomic and sector layer that frames all of them, national statistics, central-bank series, industry indicators, looks different in every market you operate in. We have built that mapping across a range of asset classes and in markets whose data has almost nothing in common, from US consumer credit to EU SME lending, and the skill that carries over is not any single dataset but knowing which sources hold signal for a given market and asset class and where to pull them. That is years of accumulated work, not a weekend of API calls.
The second side is less visible and much harder to copy, and it is knowing what a tabular foundation model responds to. These are not gradient-boosted trees with a friendlier interface. The feature engineering that helps them differs from what helps a GBDT, the handling of categoricals matters, and the choice of what goes into the context window moves the output in ways the parameters governing it only reward if you have spent real time with the internals rather than treating the model as a black box that either works or doesn’t. Most teams reaching for these models are meeting them for the first time. We are not, and the distance between a foundation model used naively and one used well is wide enough to be the whole difference between a score a committee signs and one it sends back.
What this looks like in production
An argument like this is only worth anything if it runs on a real portfolio. For us it already does, in production on client books today. You hand over a loan tape, even a thin one, and instead of the collect-clean-validate-train cycle we bring the context, the outside signals and the feature engineering behind them, and hand back a scored portfolio in a single pass, with no model to train on your side. The build follows the claim this article makes: the model is the commodity, the context is the product. One pre-trained base, context assembled per asset class, each book scored on its own terms rather than forced through a generic template.
A score on its own is not enough to act on. Every prediction breaks into named, quantified drivers a credit committee and an auditor can sign off, rather than a number they are asked to trust. Because those drivers include the alternative signals we engineer in, the explanation shows not just that a loan looks risky but what in the wider data made it so. Every step is logged. The model that clears an investment committee is the same one that survives the audit and the model-risk review that follow, because the trail from input to score is there to walk.
The model is only the starting point. The real advantage comes from the context around your data. See how Cardo AI helps you underwrite new asset classes with explainable, audit-ready AI. Book a demo.