Economic Compass

The chain from a raw series to a plotted z-score: raw level, year-over-year change, a z-score against an EMA-weighted baseline of the series’ own past (ten-year half-life, at least eight years of history required before a series is used), then a composite in which each economy is weighted by its own GDP and each economy’s indicators count equally among themselves — so a country that happens to publish four series doesn’t outvote one that publishes three. Nothing is weighted by how many numbers a statistics office chooses to release.

Those GDP weights step at each era boundary the compass already publishes, rather than applying one present-day snapshot across the whole history back to 1955; within an era, a region’s weight holds constant. The measure is real GDP in constant 2015 USD, from the World Bank (CC BY 4.0), so a currency move alone can’t change a region’s influence without a matching change in real output. The document steps weights at every era boundary it publishes; the plot, though, merges boundaries closer than 24 months apart into a single visible mark, so one mark on the page can span more than one weight step underneath it. Stepping the weights tells most in the deep past, where one modern snapshot sits furthest from the economy as it actually was: carrying today’s weights back across the whole history would move the published quadrant label on about 5% of monthly readings, almost all of them before 1999. From the 2000s onward the two approaches agree about the label almost everywhere, so this is a correction to how the compass reads its own history rather than to how it reads the present.

That means the compass measures the economies it can actually see, weighted by their size, and never speaks for one it can’t name. Together they add up to a bit under half of world output on growth, and roughly two-thirds on prices — phrased loosely on purpose, because the exact share moves whenever the weights step and a typed number here would drift out of date.

The absences are the part worth knowing. China publishes no industrial production, retail trade or unemployment this method can use, so it is missing from growth entirely; India and Indonesia are absent from both. An earlier version handed Asia a flat 40% on both axes so that Japan and Korea could stand in for those missing economies, which quietly gave two countries worth 5% of world output more than seven times their own weight. Measuring less, and saying so, turned out to be the more honest instrument.

Both are the US whole-market earnings yield minus a real 10-year rate — the compensation for holding equities instead of an inflation-protected bond. The two cards differ only in which real rate. The euro area is absent from both, unlike on the growth and inflation axes where it carries real weight: its national-accounts history begins only in 1999, and a ten-year trailing mean needs a decade of history before its first reading — too short a run since 1999 to rank honestly against the US’s longer one.

The six measures below the gauges run on a different clock. They move over years, so they are updated quarterly, and a single reading is never news.

The reserve share is the one worth explaining. Central banks report their currency holdings converted into US dollars at each quarter’s exchange rates, so when the euro strengthens and not one reserve manager moves a cent, the dollar’s reported share falls anyway. Read straight, that looks like the world stepping away from the dollar when nothing has happened at all. This page holds exchange rates fixed at their 1999 level, so the number moves only when holdings actually move. The IMF publishes the method for doing that but not the resulting series, so this is our own calculation.

Measured over the whole window since 1999, the adjustment turns out not to change the story: the adjusted decline is slightly larger than the raw one, so exchange rates are not what moved the dollar’s share. In the middle of the window it matters a good deal more, by about five percentage points around 2008, which is precisely when the dollar was weak and the reallocation story was loudest.

The gold measure is the other one worth a closer look, for the same reason as the reserve share above. Central banks report their gold holdings in ounces, which this page converts to tonnes for readability, never to dollars. A dollar figure for gold rises whenever the gold price rises, even if not one central bank buys another ounce, and would count a price rally as buying. Ounces, like the reserve share’s fixed exchange rates, keep the number moving only when holdings actually move.

Gold has a second wrinkle. China first reported its gold to the IMF in 2015, arriving already at 1,658 tonnes. A plain four-quarter difference across that date would have counted the filing itself as a purchase of nearly 1,700 tonnes, when the real change that quarter was about 30. This page differences holdings only across the countries reporting at both ends of the window, so a country’s first appearance is never mistaken for buying.

Gold has a third wrinkle. The IMF’s file occasionally carries a figure that cannot be true: Angola’s holdings arrive at some 30,000 times the gold it actually has, and Brazil’s break by a factor of a thousand partway through 2026. One impossible reading would dwarf every real one in the basket, so both countries are screened out by name and the published change is the change across everyone else.

And several of the largest recent buyers report little or nothing to the IMF, so what is published here is the visible part of official gold buying, not the whole of it. Reported net purchases came to 252.4 tonnes in 2022, 363.2 in 2023 and 246.9 in 2024, against a World Gold Council estimate of total official buying near 1,000 tonnes a year across the same period, a lower bound running at roughly a quarter to a bit over a third of the total.

The gold readings also arrive late on purpose. A quarter is held back until enough of its reporters have filed, and central banks file at their own pace, so the newest figure on the card is the last completed quarter rather than the one under way, a couple of months behind the calendar. Publishing sooner would mean publishing a quarter with half its reporters still missing, which reads as selling that never happened.

Each measure carries a criterion that was written down before the data ran, dated, and expressed as a position in that measure’s own recorded range rather than as a fixed level. A level chosen in advance can turn out to be unreachable, and then it is not a cautious threshold, it is an absent one. A position in the range cannot be.

One thing this section does not do: it does not test how long a measure has been where it is. The criteria describe where a reading sits now, not how long it has sat there.

Every reading is point-in-time from the latest published vintage; macro series are revised, so a historical point differs from what was knowable on the day it was first published. The composite draws on Ray Dalio’s public writing on economic machine cycles; this page is Constantin’s own implementation, not affiliated with or endorsed by Dalio or Bridgewater. A series missing from the newest month is carried forward from its own last reading for up to three months before it drops out, so the composite stays full instead of shrinking to whichever handful of series happened to report first — the honest cost is that the newest point on the plot is a better match for last month’s settled reading than for this month’s.