Research · 2026-Q1 snapshot

The Reserve Map

One central bank publishes what it owns, position by position. It does so because an American regulator makes it, and it is the only one.

Every quarter the Swiss National Bank files a Form 13F with the US Securities and Exchange Commission listing 2,301 equity positions by issuer, share count and market value. It files because it owns US-listed stock and the rule that catches American asset managers does not exempt foreign monetary authorities. Nothing about reserve management prompted the disclosure. A filing requirement written for somebody else did.

The holdings themselves are unremarkable, and that is the point twice over. A central bank's disclosed equity book turns out to look like the American market it is a slice of, ranked by size — which is what you would expect of a portfolio held for diversification rather than for influence. The remarkable part is that anyone outside the institution can see it at all.

Everything else here arrives at one remove. Gold is a tonnage an authority reports to a statistical agency, months late and in its own time. Treasury holdings are assembled from custodians, so the country attached to a figure is where the security sits rather than who owns it. Currency composition is published as a world total and nothing finer, because participation is voluntary and country-level detail is confidential by design. Of the institutions tracked here, 2 disclose at the level of individual holdings — and one of those two is a sovereign wealth fund, not a central bank balance sheet — while 1 publishes nothing an automated reader can reach.

That narrows the question this page can honestly ask. Not what central banks hold — nobody outside them knows — but what they can be seen holding, on whose word, and how much of the movement in that picture is the picture moving rather than the holdings.

The map takes each reported figure at face value and asks only where it came from. A solid mark is an authority describing its own position; a dashed one is a custodian describing it for them, and the mark is sized by the holding rather than the country. Belgium and Luxembourg carry more recorded Treasuries than Canada or France — shade the countries in and that disappears into two slivers, which is the reason for circles. Underneath sits everything no map can place at all.

Ranked, the same figures make the custodial problem impossible to miss and easy to misread. Belgium, Luxembourg, Ireland, the Cayman Islands and Switzerland sit among the largest recorded holders of US government debt because a security is recorded where it is kept. 30.1% of the recorded total sits in centres of that kind, marked here, and nothing on this page attributes a movement in them to a decision by the country named.

The valuation gap

The dollar's published share of allocated reserves is 57.1% at 2026-Q1, down from the start of the series. That number mixes two different events. Authorities may hold fewer dollars, and the dollar may have moved against everything else it is measured beside. Holding exchange rates at the base quarter separates them: on that basis the share is 56.6%, a difference of -0.57 points. Sources [8], [7].

The correction is a construction and the base quarter is a choice. Measured against 5 different base quarters the gap ranges from -0.52 points to -4.96 points. The sign holds — every base tested puts the corrected share below the published one — but the magnitude moves by an order of magnitude, so no single figure is quoted here as the correction.

Two limits belong with the number rather than after it. 8.2% of the denominator cannot be revalued at all: the "other currencies" residual has no single exchange rate, and a currency has no base rate before it entered the reporting set. Both are carried at their reported dollar value. And the correction removes exchange-rate valuation only — a euro bond portfolio also moves in euro terms when euro yields move, and the published data cannot separate that. The result is closer to a quantity measure than the published share is. It is not one.

Gold in tonnes

The section above spends its length on a valuation problem. This one does not have that problem. Official gold is reported as a volume, and a tonne carries no price, so the quantity question that the currency data can only approach is answered here directly.

Across 85 reporting authorities the reported basis stands at 30,448 tonnes. In 2025, the last complete year, 331 tonnes were bought and 75 tonnes sold, for 257 tonnes net. The chart plots the net, which is the quantity the debate is about; the gross figures stay in the text beside it, because one line called central bank demand hides whoever is on the other side of it and the difference between the two readings is not something a reader should have to measure by eye. Source [10].

Read across, the series divides cleanly in two. Every year from 2000 to 2008 is net selling; every year since 2009 is net buying. That turn is visible without any estimate being added to it, which is the whole argument for staying on the reported basis.

This is smaller than the figure usually quoted, and the difference is the point. Compiler estimates of central-bank gold buying run well above the reported basis, because they add purchases that no authority has reported. Neither number is wrong. One counts what was disclosed and the other estimates what was not, and the distance between them is a measurement of the disclosure gap rather than a disagreement to be reconciled. No estimated figure enters any total, share or rank on this page.

Two reporters are excluded and named rather than quietly dropped: AGO and BRA report volumes above the largest known official holding, so their reporting unit is not fine troy ounces. Rescaling them on a guess would be worse than excluding them. Reporting lags differ, so observations from 2026-06-01 are incomplete and will revise upward; the faded final year on the chart is that window.

How far down each institution publishes

This is the ceiling on everything above. Seven of the 13 institutions tracked here publish a balance sheet split into components and stop; three publish a single total; one publishes nothing an automated reader can retrieve. Whether a series is thin because the institution keeps it that way is a separate question from whether this build collects it, so the two are kept apart — the ladder is the first, and the register below it is the second.

Every series follows, with its publisher's own identifier, its unit and scale, its accounting basis, its registered sources and any documented break. 9 collectors currently return rows and 10 are visibly skipped — counted per adapter, not per institution, because several institutions have more than one.

Method

Adapters emit only what a publisher published. Shares, differences and revaluations are computed after collection, in one place, so they can be tested. Units are asserted before any arithmetic, every row carries a source id, an accounting basis and a retrieval vintage, and weekly and monthly observations are never forward-filled onto one another.

The dollar share is computed from reported levels, not read from a published share. That keeps the denominator on the page and makes the parse checkable: across all 109 quarters the largest disagreement between the published allocated total and the sum of its currency components is 4.8e-13 percentage points, which is floating point rather than a difference.

The constant-exchange-rate share. Each non-dollar holding is converted back into its own currency at that quarter's rate and revalued at the base quarter's rate; the dollar leg is unchanged by construction, which is also the test — at a base equal to the quarter itself the two shares are identical. Rates are end-of-period from one provider for the whole history, matching the convention the compiler uses.

Gold accumulation. Period-over-period changes in reported tonnage, summed across reporters. A reporter with no observation in a year is excluded from that year rather than counted as unchanged, and the excluded count travels with the figure.

Limitations

• COFER is an aggregate and participation is voluntary. It contains no country-level currency composition and none can be inferred from it, which is why one node on the map is a crowd.
• The constant-exchange-rate correction removes exchange-rate valuation only. Bond-price valuation remains inside each currency sleeve and cannot be separated from published data.
• 8.2% of the denominator cannot be revalued: the other-currencies residual has no single rate, and a currency has no base rate before it entered the reporting set.
• The size of the valuation gap depends on the base quarter and varies from -0.52 points to -4.96 points. Only its sign is stable across the bases tested.
• Gold tonnage is a disclosure series. A step in it is a reporting event, not a purchase date, and observations from 2026-06-01 are incomplete.
• AGO and BRA are excluded from every gold figure because their reported units are not fine troy ounces. Estimated compiler figures enter no computed value.
• TIC records custodial location at market value. 30.1% of the recorded total sits in named financial centres, and no movement here is attributed to an owner's decision.
• Balance sheets stay in local currency and are indexed because accounting bases differ. No global total is computed.
• The Swiss National Bank filing covers US-listed equities only: a slice of one portfolio, not that institution's assets.
• 10 adapters are skipped in the current manifest, each with a stated reason, and the coverage grid shows which.
• The largest reserve holders are the least visible across currency composition, gold reporting timeliness and security-level portfolios at the same time. That is the biggest limitation on this page.

Source register

The register contains 23 probed sources. “Corrected” means the research lead and the live publisher did not agree, and the note records the difference.