Someone ran our withholding model against two ETF pairs. One switch doesn’t pay.

1. Two gates before any arithmetic

The study will not compare a pair until it clears two tests, and the second one is the interesting one.

Gate one: the same published index. Where two funds track different things, the gap between them contains exposure as well as cost, and a fee-and-withholding comparison quietly books the one as the other. VT against a FTSE All-World UCITS is not a wrapper switch; it is a different index — VT vs VWRA is that pair in full, with the index gap and the wrapper gap kept apart.

Gate two: how much of the dividend stream is actually US-source. The 15% fund-level rate used throughout is the US–Ireland treaty rate on US-source dividends. It says nothing about a dividend paid by a Japanese or British company, which is a different rate on both sides. That puts three candidate pairs in three different positions:

That third row is the part most comparisons skip. A global fund is exactly where the two-layer withholding argument is most often made and least computable, because it needs per-source-country dividend data that neither our model nor most published ones carry. The study declines to claim a number rather than estimating one.

2. The condition, in one line

For a position of value V paying annual distributions D, with the investor’s US treaty rate r, a fund-level rate L1 of 15%, and ΔTER the Irish fee minus the US fee:

annual difference = D × (r − L1) − V × ΔTER

Substitute D = V·y, where y is the distribution yield, and the position value cancels:

the switch is favourable ⇔ y × (r − L1) > ΔTER

Two things follow immediately, and both are easy to miss when the argument is made in prose.

The withholding term scales with the yield. The fee term does not — it is charged on the whole position, every year, whether the fund distributes anything or not. So the case for the Irish wrapper is strongest on a high-yielding index and weakest on a low-yielding one, and there is a crossover yield where the two cancel.

And the study deliberately writes the condition as a multiplication rather than as a crossover yield y* = ΔTER / (r − L1). The division form is undefined at r = 15% and silently flips its own inequality below it — which is precisely the region where the results below stop being obvious.

3. Two pairs, opposite signs

Published expense ratios, read 12 September 2026. No proxy, no model — six numbers off six fund pages:

The third row is worth a pause on its own, because it needs no tax assumption at all. The standard framing — US funds are cheap, Irish funds carry a fee premium you accept in exchange for a withholding saving — is simply false on the MSCI World pair. A reader applying the general rule would go looking for a trade-off that does not exist.

Now the annual comparison at the US statutory 30% rate, on a $100,000 position:

Same wrapper argument, same treaty rate, opposite answers.

The Nasdaq-100 pair is the one worth dwelling on. A withholding-only view reports $63.58 saved and stops there. The fee difference is nearly twice as large in the other direction. The crossover yield is 0.80% against a fund yielding 0.4238% — not a marginal call, a gap of close to two to one. Any tool that shows the withholding saving without the fee difference reports this switch as a benefit when, on these two layers, it is a cost.

The S&P 500 pair clears its crossover with a 4.0× margin, which sounds comfortable until you notice what the margin is made of. ΔTER is a difference between two very small numbers. A two-basis-point error in either fee moves the crossover yield by 0.13pp. The fee gap would have to widen from 4bp to about 16bp for the sign to change — real headroom, but headroom measured in basis points, not in percent.

Position value cancels out of the condition, so the sign of each result is independent of the $100,000 figure. The dollar amounts are not.

4. Below the statutory rate, both go negative

The statutory 30% applies where no treaty rate does. Most people reading this are not there. The same two pairs at a 15% treaty rate — Germany, the Netherlands, South Africa — and at 10% — Japan, China, Bulgaria, Romania, Mexico:

At 15%, the withholding term for the S&P 500 pair is not small. It is exactly zero, for every yield, because the investor’s rate and the fund-level rate are the same rate applied to the same income, and every S&P 500 dividend is US-source by index rule. What is left is the fee difference and nothing else.

Below 15% the term goes negative. The Irish fund suffers 15% inside itself on income the investor would have had 10% taken from directly. The wrapper now costs withholding rather than saving it, and the fee gap is charged on top.

This is the finding least likely to be already known, and the one our own product was hiding. More on that in section 6.

5. Switching an existing position: the tax you pay today dominates

Everything above assumes the money is already in the Irish fund. For new money that is the whole story — there is no sale, so no capital-gains charge and no sell-side friction. For an existing position there is a one-off cost:

capital-gains charge = V × embedded gain × CGT rate   |   friction = V × 0.5%

On $100,000 with a 40% embedded gain at a 15% rate, that is $6,000 + $500 = $6,500 — against $119 a year. The simple payback, dividing the one-off cost by the annual difference:

IVV → CSPX at a 30% treaty rate, distributions reinvested and untaxed. Simple benchmark — the appropriate treatment for a holder who would never otherwise sell.

Read the left column and the bottom row together. With no embedded gain the switch pays for itself in four years. With a 60% gain and a 25% rate it does not pay for itself inside a human lifetime. The spread between the capital-gains columns is larger than any difference between the pairs. For a holder sitting on a large embedded gain, their own tax position matters more than which pair they are switching — and the withholding argument that started the conversation is a rounding error against it.

The study also runs an acceleration basis, which compares both paths after tax at a horizon where both are sold, on the grounds that a holder who would eventually sell anyway has moved the tax forward rather than created it. That basis is usually assumed to be the friendlier of the two. With distributions untaxed it is longer in every cell where the capital-gains rate is above zero — 16.5 years against 4.2 at a 0% gain and a 25% rate, 179.8 against 130.3 at the bottom right. The tax paid today stops compounding, and the accumulating fund converts dividends into capital gain that is taxed on sale.

For the Nasdaq-100 pair there is no simple payback figure anywhere, at any gain or rate, because the annual difference is negative before the switching cost is counted. On the acceleration basis with distributions untaxed, no horizon exists within a thousand years in any cell.

6. What it found in our engine

Two things, both ours, both in the study.

We floored the withholding saving at zero. Our engine computes the per-position saving as the annual dividends multiplied by the difference between the investor’s rate and the 15% fund-level rate — and then clamps the result at zero. For anyone on a treaty rate below 15%, that difference is negative and the clamp turns a real penalty into a displayed “no benefit”. The signed figures are the −$93.00 and −$141.19 in the table in section 4.

Checking our own source while writing this, the consequence is slightly worse than the study states. A clamped-to-zero saving also fails the test that admits a pair into the net comparison, so for a 10%-treaty investor the position drops out of the aggregate altogether — and with it the fee difference, which is real, positive and charged every year regardless of any treaty. The reader was shown neither half of the penalty.

It is worth being precise about what was and was not broken. The net figure — withholding saved minus the extra fund fee — has never been clamped in our engine, and the payback horizon has always divided the one-off cost by that net figure rather than by the withholding saving alone. Those were deliberate choices, and the study’s central point about fee-blind withholding tools is not a description of what we ship. The zero floor on the component is the defect, and it bites exactly one group: investors whose treaty rate is better than Ireland’s.

The cost-basis treatment differed from ours, and ours was wrong. Every US fund in these pairs distributes and every Irish one accumulates. An investor reinvesting distributions from a distributing fund in a taxable account buys new shares, and those shares carry their own cost basis. Our supplied implementation held the stay path’s basis fixed while compounding net distributions into the terminal value — which taxes those distributions a second time as capital gain at the eventual sale. The accumulating side was modelled with a fixed basis, which is correct there, so the asymmetry was one-sided and ran in favour of switching. The study adds reinvested distributions to the stay path’s basis, reimplemented our version under its own discretisation to rule out method as the cause, and reproduced our figures to within two to four percent. The remaining gap is the basis handling. It is ours to fix.

We are also carrying the study’s §5 point, which it declares out of scope rather than solving: a flat 0.5% proportional switching cost has no fixed commission, no per-order minimum and no FX spread in it. On a $500 switch at $3 a side that is 1.2% against a 0.5% allowance, so small-ticket payback figures understate the cost. Closing it needs a fixed-plus-proportional split our model does not currently carry.

7. What would falsify it

The study states its own reversal conditions in advance, which is the part we would ask other people to copy. Among them:

• The S&P 500 fee gap widening to 16bp or more, from either a CSPX increase or an IVV cut. That reverses IVV → CSPX at the statutory rate. A narrowing gap does not falsify the result; it strengthens it.
• Any published QQQ distribution total above roughly $5.72 per share at the current price. That puts the Nasdaq-100 pair above its crossover and reverses the negative result.
• Constituent-level data showing the Nasdaq-100’s non-US dividend share above approximately 88.7%, at an assumed 0.30 withholding differential on the non-US portion. Below that share — or at any differential of 0.15 or less, where no share flips it — the stated result stands. This is a conditional bound, not a measurement, and the study says so.
• Any domestic tax treatment that distinguishes accumulating from distributing funds. This reverses results in either direction and is out of scope throughout.

The Irish fees in both the study and our own pipeline come from justETF. That is consistent, not independent, and the study flags an issuer KID disagreeing with justETF as a falsifier for that reason.

What we take from it

The useful result is not “Irish wrappers are good” or “Irish wrappers are oversold”. It is that the question does not have a general answer, and that the two inputs which decide it — the index’s dividend yield and the pair’s fee gap — are both published numbers that anyone can look up for their own holdings. On the S&P 500 at the statutory rate the switch pays. On the Nasdaq-100 it does not. On the MSCI World pair the fee gap runs the other way entirely and nobody has the data to price the tax side.

For an existing position, none of that is the binding constraint anyway. The embedded gain and the capital-gains rate are, by a wide margin.

Handing a model to someone who will publish what it says rather than what you would like it to say is an uncomfortable and extremely efficient way to find its defects. We would do it again, and the two above are on the list to fix.

Read the full study at QuantRoutine →