Portfolio Lab in your own currency.

Seven currencies, two kinds

5 currencies have an edition of their own: US dollars, sterling, euros, Canadian dollars, rand. Picking one of those changes the numbers. Australian dollars and yen have no edition, so picking one changes the symbol on the money and leaves the assumptions in dollars. The app marks that on the currency control and again below the stage rail, because it is the thing a reader is most likely to get wrong.

The class counts differ because the editions differ. J.P. Morgan’s dollar book carries 57 of the classes Portfolio Lab prices and their sterling book carries 73, so 18 rows have no sterling figure. Those rows keep their weight and are marked on the sheet rather than converted at today’s exchange rate, so a portfolio switched to sterling and back comes home the way its owner left it.

What changes in the return

Comparing every asset J.P. Morgan publish in two editions at once shows what they do to a return when they restate it. They add one constant. Across the 20 unhedged lines the sterling and dollar editions share, the gap is the same to four decimal places for every one of them, US large cap and gold and emerging market debt alike, and the spread across those lines measures 0.0000.

That constant is a currency forecast. J.P. Morgan expect sterling to be worth more dollars in ten years than it is today, so a sterling investor earns less on the same global basket, and the sterling edition sits 0.60 points below the dollar one. They expect the rand to weaken, so a South African investor earns +1.50 points more on the identical holdings. The hedged column is smaller because a hedged share class has swapped the currency move for the interest rate differential between the two countries. It also rests on much less evidence: only 3 hedged lines appear in both the rand and dollar editions, and their gaps differ by 0.00, because two hedged lines with the same name can hold different bonds.

What changes in the risk

Volatility moves per asset. Hold an unhedged foreign fund and you carry the fund’s own risk and the currency’s risk together, and how much of the second you feel depends on whether the two move at the same time. A dollar asset that falls when the dollar rises against your currency is partly self-hedging, and the two risks work against each other.

A currency-hedged share class is the exception. The hedge removes the exchange rate leg, so the volatility a sterling investor sees is the dollar volatility, and J.P. Morgan’s own editions agree on hedged volatilities to a few hundredths of a point. Where a reader ticks the hedged variant of a row, the assumptions table, the optimiser, the backtest and the printed report all move to that share class together.

Whether a hedged share class exists for a reader to buy is a separate question from whether J.P. Morgan publish a hedged figure, and the app keeps a register of it. Each hedged line in each currency is graded: verified, with a named share class; standard, where hedged classes are routine for that asset in that market; not found, where a search turned up nothing; or no vehicle at all. Where the register cannot vouch for a line, the assumptions page says so under the hedge choice, with the date the register was checked, and that the unhedged figure is the one a reader can buy today. Not found is a lead to chase, never a finding that no product exists.

Where the correlations come from

Every correlation in every currency is measured, from each asset’s own monthly returns restated into the reader’s currency, over J.P. Morgan’s own stated risk window of 2006-07 to 2025-06. Not one cell is borrowed from another asset.

That is worth stating plainly because it used to be otherwise. A base-currency matrix was J.P. Morgan’s published one with the rows they do not publish filled in by proxy: a sterling line took the correlation column of whichever published row it behaved most like. That shipped as a deliberate interim with its cost written into the code, and it is gone.

A matrix measured pair by pair is not guaranteed to describe a possible world: pairs measured over different windows can disagree with each other in ways no real set of returns could. So every eigenvalue is lifted to 0.01 and the matrix rebuilt, and the last column above is how far that moved the worst single cell. The gate runs before anything consumes the matrix, so an optimiser never sees one that has not passed it.

Checking the restatement against their own matrices

J.P. Morgan publish a full correlation matrix in every edition, which makes those editions a test of the currency machinery and not only an input to it. Take the dollar matrix, restate a pair of assets into sterling by the identity above, and compare the answer with the figure printed in their sterling book. Do that for every pair, in every currency.

This is a check on the restatement, not the route the app prices by. Nothing in the pricing path runs it: the correlations above are measured from returns. What it establishes is that the arithmetic which converts a series into a reader’s currency recovers a figure J.P. Morgan arrived at independently, which is the part that would otherwise be taken on trust.

Across fixed income and equities the restatement reproduces their published correlations to a mean absolute error of 0.0110 in Canadian dollars and 0.0312 in rand. A correlation runs from minus one to one, so an error of a hundredth is the third decimal place of a number most houses publish to two. For fixed income and equities the identity is not an approximation of their treatment. It recovers their treatment from the outside, which is why the conversion it validates can be trusted to carry a series into another currency before that series is measured.

One property makes this a real test rather than a circular one. The restatement needs the exchange rate’s own volatility, and every figure in the table above is the same whichever value is used for it, because the term that carries it cancels out of a correlation. The test suite checks that rather than taking it on trust. The next section is about that value, which is the one number here that has to be chosen.

Rows that share a published line

J.P. Morgan publish one “Euro Government Bonds” line, and a euro reader may reasonably want the German part and the Italian part separately. Every row in this tool owns a published line, so for a long time those rows could not exist: they each wanted a line another row already had.

A cut is the answer. It shares its parent’s published line, moves the return by a spread and the volatility by a ratio that are both measured, and is exempt from the market-portfolio weight because the parent already carries the capitalisation. 17 of the 34 base-local rows are cuts today.

Two properties are easy to get wrong and are pinned by tests. The volatility travels as a ratio rather than a level, because a base-local line appears in more than one edition and a fixed number would assert one currency’s figure in all of them. And a cut of a cut compounds from the published line rather than from its immediate parent, so the German short band carries the German spread plus its own.

Two windows, and both are right

The correlations above are measured over 2006-07 to 2025-06, which is J.P. Morgan’s own risk window. A backtest in this tool can run to 1926-07. Those are different numbers on purpose and a reader who notices will reasonably wonder which is the mistake.

Neither is. They answer different questions. A forward-looking correlation is an estimate of how assets will move together over the next decade, and J.P. Morgan choose a window recent enough to describe today’s market structure. A backtest asks what a holding actually did, and throwing away eighty years to match a forecasting convention would answer a worse question. Every surface says which window it used, in its own words, rather than leaving the reader to reconcile two dates.

The one number we choose

Everything above is read from a published book or worked out from one. One number is chosen, and a reader is owed the reasoning.

The risk identity needs the exchange rate’s own volatility, and J.P. Morgan’s tables cannot pin it down. Every candidate value reproduces every volatility they publish, exactly, because the correlation with the currency is solved out of that same equation and absorbs whatever goes in. A correlation between two assets comes out the same either way. A volatility restated from another forecaster’s dollar figure does not, and that is where the choice is felt.

One house shows exactly where. AQR’s dollar volatilities sit within 14.220 of a percentage point of J.P. Morgan’s on every name both publish, so restating AQR is the same arithmetic as restating J.P. Morgan’s own row and the answer does not move whatever value is chosen. BlackRock differs from them by up to 12.1 points on a single line, and BlackRock’s restated figures move with the choice. The gap between one house’s volatility and another’s is the whole of what this number touches.

The tables do restrict a range, in two steps. A correlation cannot be larger than one, which rules out most values. Adding the point that no foreign asset rises when the dollar rises against your currency rules out more.

The chosen value is the one at which J.P. Morgan’s implied correlations best match correlations measured from history for the same assets, so it uses their book and the market’s own record together rather than either alone. The realised column is a cross-check and never the input: these are forward-looking assumptions, and a house is entitled to expect more currency risk than the past delivered.

A value drifting outside that region is caught rather than absorbed. The rand shipped for a time at 12.0%, which is inside what the tables arithmetically allow and below the 12.9% floor the sign rule sets, and at that value World Government Bonds implied a positive correlation with the rand: a foreign bond rising as the rand fell. The code now refuses to price a currency whose chosen value implies that, and says which asset gave it away.

Where the method fails

Let the alternatives into the same test and it breaks. The mean error rises from 0.0110 to 0.0772 in Canadian dollars, and the worst pair in the sterling book is Emerging Markets Equity against Direct Lending, published at -0.18 where the identity predicts 0.42, positive where their own book is negative.

The same names are responsible in every currency: Commercial Mortgage Loans, Direct Lending, European Core Real Estate, Global Core Infrastructure, Global Core Transport, Global Timberland, Private Equity, U.S. Core Real Estate, Venture Capital. All of them are valued by appraisal rather than by a market price every day. Their published volatilities are smoothed by that process, and each edition appears to re-estimate them rather than restate them, so there is nothing for the identity to restate.

Measuring a holding against a currency

How much a holding moves with an exchange rate is measurable directly, from the holding’s own dollar returns and the month-end rate, over the last 240 months. The rates are taken at month end rather than averaged across the month, because an average hides the months a currency actually moved.

A South African government bond is the clearest case. Priced in dollars it looks like one of the most volatile things a portfolio can hold, because its dollar price carries the whole rand exchange rate on top of the bond. For a South African, who is paid in rand and spends in rand, it carries no currency risk at all.

Nothing in that table is anyone’s forecast. The identity takes 20.0% of dollar volatility down to 7.4% in rand, which is what a South African holding those bonds lived through, and it needs a correlation close to minus one to get there. The measurement, made without reference to that answer, returns -0.947.

Where the local history comes from

A backtest in sterling needs gilts, a backtest in rand needs South African government bonds, and index funds for those markets are young. 15 local bond lines are built in two segments joined at the month their fund began trading: before the seam, total returns reconstructed from OECD long-term government bond yields at a duration fitted to the fund’s realised volatility, and after it, the fund’s own monthly total returns.

Local equity lines are read from a listed fund throughout: UK All Cap from 0P0000WN7B.L on the London Stock Exchange, fund NAV quote, 2014-03 onward; UK Small Cap from CUKS.L on the London Stock Exchange, 2010-11 onward; UK IG Corporate Bonds from SLXX.L on the London Stock Exchange, 2008-03 onward; Euro IG Corporate Bonds from IEAC.L on the London Stock Exchange, euro-denominated, 2009-05 onward; Euro High Yield Bonds from IHYG.L on the London Stock Exchange, euro-denominated, 2010-11 onward; Euro Area Small Cap from CSEMUS.SW on the SIX Swiss Exchange, euro-denominated, 2009-08 onward; European Large Cap from EXSA.DE on the Deutsche Boerse Xetra, 2008-03 onward; European Small Cap from SMC.PA on the Euronext Paris, 2008-03 onward. Each is checked against the volatility J.P. Morgan publish for the same line in the edition that carries it, and the generator refuses to write a series that misses by more than its stated tolerance.

The cuts an edition adds beside its main lines, UK small cap, German mid cap, Canadian small cap, South African mid and small cap and euro area small cap, carry the figures their edition prints. They are not marked to market yet, for the same reason as the domestic bond lines, and their weight in the market portfolio is a slice of their own country’s share rather than of the dollar line they take their shape from.

Cash and inflation in your own money

Both are read from your edition rather than converted from the dollar one. Cash is a different instrument in every market and UK inflation is a measurement of a different economy, so neither is a second quote of an American number.

Every Sharpe ratio, every risk parity solve and every real return on screen uses the row for the currency you are working in. Rand cash pays 7.0% against US cash at 3.1%, so a rand portfolio measured against a US cash rate would show a Sharpe ratio no South African investor could earn, and a rand retirement plan escalated at US inflation would understate what the shopping costs by 2.4 points a year.

The assumptions behind these figures are set out in return forecasts, and the bond model in fixed income. The full method index is at methodology.