The Business Problem
Lending somewhere with no track record
A model that spots failing companies has to be built from companies that already failed. That works fine until the next set of customers is unfamiliar: a country the bank has just entered, an industry it has never lent to, a loan book it just bought from someone else, or a group of suppliers it suddenly depends on.
The accounts are available in all of those cases, so the inputs aren’t the problem. The outcomes are. You cannot learn what failure looks like from companies that have never been recorded failing.
The usual shortcut is to borrow a model: build it where the outcomes are known, then point it somewhere new. The catch is checking it. To prove the borrowed model worked, you need exactly the failure records whose absence sent you borrowing in the first place.
How this project gets around that
Both datasets here record what actually happened to every company. So the Polish outcomes can be hidden, used to mark the model afterwards, and never fed into it — the check a real bank cannot perform on itself.
Learn Where The Answers Exist
Build the model on Taiwanese companies, where it’s on record which ones went under.
Judge The New Country
Score Polish companies. The model never sees a single Polish outcome.
Mark The Homework
Only then reveal what happened to those Polish companies, and measure how much was lost.
The one thing this can’t separate
The Taiwanese companies are stock-market listed. The Polish ones are mostly private. Listed companies tend to be larger, and they’re obliged to publish far more about themselves.
So the country isn’t the only thing changing between the two datasets — the type of company, the accounting rules and the reporting standards all move at the same time. If the model does worse abroad, no design using these two datasets can tell you which of those caused it.
What the setup can measure is the thing a lender actually faces: a messy, realistic change of population, of the kind that happens whenever a model gets reused somewhere it wasn’t built. The question being answered is not “does Taiwanese bankruptcy theory apply in Poland” but does a model survive being moved?