When a Model Crosses a Border

There is a tempting notion that internationalizing a financial model is chiefly a matter of exchanging currency symbols and substituting tax tables. The arithmetic generally travels well enough; the underlying assumptions are less accommodating.

Different tax jurisdictions

A recent investigation into the international applicability of two financial models, one concerned with shareholder liquidity and the other with employee stock options, took us through more than a dozen tax jurisdictions. Britain and much of continental Europe offered promising starting points, provided we accounted for differences in tax residence, equity plans, and the timing of taxable events. Elsewhere the questions became considerably more interesting.

In the Netherlands, annual taxation of investment returns changes the comparison between borrowing against shares and selling them. In Norway, Spain, and Switzerland, net-wealth taxes can introduce costs even when no transaction takes place. New Zealand’s foreign investment fund rules can produce taxable income without a sale or cash distribution. Australia presented promising opportunities for both models, though employee share schemes introduce their own peculiarities. Each jurisdiction suggested a different answer to the same question: what, precisely, are we comparing?

Markets and watermarks

Another investigation concerned the curious geography of financial markets themselves. Hyperliquid, a decentralized trading platform whose interface restricts American participants, nevertheless publishes market information through an accessible API. Prices, executed transactions, funding rates, open interest, and order-book activity offer opportunities to observe market sentiment without participating in the market. We are particularly interested in divergence: what can be learned when different markets assign different values to the same underlying proposition?

A third inquiry arose from the European Union’s AI Act, whose transparency provisions have introduced new requirements for identifying machine-generated content. Language models are beginning to incorporate statistical watermarks into their output, raising questions about provenance, detection, and the treatment of synthetic text as it travels between systems. We are investigating whether these markings might eventually influence search visibility, email delivery, or other mechanisms governing the circulation of information. Here, experimentation is more useful than speculation.

The arithmetic generally travels well enough; the underlying assumptions are less accommodating.

Revisiting the model

These are rather different problems, yet each illustrates a familiar difficulty. A system designed within one set of assumptions encounters another, and what once appeared to be a minor implementation detail becomes fundamental to the model’s behavior.

Sometimes the solution is a new parameter. Sometimes it requires an entirely different model. The interesting work lies in discovering which is which.