Four Linkages, One Anchor

The easiest thing to get wrong: GLD holds not a single futures contract. It is a physical gold trust (a grantor trust) holding allocated bars in HSBC’s London vault, with NAV struck off the LBMA afternoon fix. So what GLD directly anchors to is the London spot market; its link to COMEX futures is indirect, held together entirely by arbitrage.

The linkage has two layers. The first is creation / redemption, which pins GLD to spot. Authorized Participants (APs — essentially the big market makers and investment banks) can swap physical bars with the trust for new shares in 100,000-share baskets, or redeem shares back into bars. When GLD trades above NAV, an AP buys physical gold, delivers it to the trust, creates shares, and sells them in the secondary market for the arb; at a discount it runs the reverse. The trust itself is entirely passive — when money flows in, it’s the AP, not GLD, that actually “buys the gold.”

The second layer is EFP and cost-of-carry arbitrage, which pins London spot to COMEX futures. An EFP (Exchange for Physical) swaps a futures position for a physical one; its quote is the futures-spot basis itself. That basis is bound by cash-and-carry — futures price roughly equals spot plus financing, storage, and insurance cost, and if it strays too far, arbitrage steps in. In practice, when a market maker provides GLD liquidity on NYSE Arca, the handiest hedge is COMEX futures (deepest liquidity, margin trading); it then uses an EFP to swap the futures leg into London spot and convert it into allocated bars to complete creation. So buying and selling pressure in GLD often hits the futures pit first, then travels back to the physical market via EFP.

On the direction of price discovery, most empirical work finds COMEX futures lead (especially in US hours), with spot and GLD following — but transmission runs both ways; the three prices are nailed by arbitrage into “one price.”

Today’s LBMA gold price is, mechanically, no longer the old-style fixing — since March 2015 it has been run by ICE Benchmark Administration (IBA) as an electronic auction regulated under UK benchmark rules, twice daily: 10:30 London time (the AM price) and 3:00 pm (the PM price).

At its core is a round-based auction: IBA’s algorithm sets an opening price referencing the prevailing market, then runs 30-second rounds in which a dozen-odd direct participants submit their buy and sell volumes. At the end of each round the totals are compared — if the imbalance falls within a 10,000 oz tolerance the auction ends, that round’s price becomes the fix, and participants absorb the small residual pro rata; if it exceeds tolerance the price moves toward the imbalance (up if buys exceed sells, down if the reverse) and a new round begins, until the market clears.

A few extras: trades settle in London unallocated gold; the fix is quoted in dollars and simultaneously published in sterling and euro; and since 2017 IBA has offered a central-clearing option so participants no longer depend entirely on bilateral credit lines, lowering the barrier to entry.

Picking up the previous section: GLD’s NAV uses exactly this PM print, which makes the 3 pm auction the single most heavily-weighted moment in the global gold-pricing chain — central-bank accounting, mining long-term contracts, and ISDA derivatives settlement mostly reference it too. The renminbi counterpart is the Shanghai Gold Benchmark Price (since 2016), a similar centralized auction but quoted in yuan per gram with physical delivery.

First, two details straight: the relationship is a tight inverse correlation (strong dollar, weak gold), and its “tightness” is far less stable than it looks. This inverse relationship isn’t a single cause — it’s at least three layers of mechanism stacked together.

The most mechanical layer · the numeraire effect. Gold is priced globally in dollars, and the DXY is essentially “the price of the dollar against a basket of currencies.” Gold behaves like a currency of its own (its ticker is literally XAU): gold/USD = gold/EUR × EUR/USD, a triangular relation locked down by arbitrage. So when the dollar depreciates against the euro or yen, the dollar price of gold must mechanically rise even if gold priced in other currencies doesn’t move a hair. The euro is nearly 60% of the DXY, so much of the time “DXY driving gold” is really EUR/USD driving it.

A deeper layer · the common driver. Both assets are pulled around by the US real interest rate. Gold is a zero-coupon asset, so the cost of holding it is the real yield; the dollar is priced by rate differentials and capital flows. When the real rate rises, the dollar gets dearer and the cost of holding gold rises — the same hand pushes the two assets in opposite directions. That’s why in the instant a CPI, payrolls, or FOMC lands, the DXY and gold futures jump almost as mirror images: they aren’t causing each other, the same headline is pricing both at once.

Then add one fast and one slow channel. The slow one is physical demand: a weaker dollar means higher purchasing power for buyers in China, India, and Europe, so jewelry and investment-bar demand recovers (and is suppressed in reverse) — on a quarterly timescale. The fast one is microstructure: on COMEX, CTAs and macro funds explicitly trade “dollar-gold” as a pair, and market-making algorithms hedge cross-asset, so the moment EUR/USD moves, systematic money hits gold futures within milliseconds. So a good part of the “tight synchrony” you see at tick level is self-reinforcing, executed by the algorithms.

But this correlation is regime-dependent and can’t be treated as a constant. The rolling correlation of daily returns swings roughly between −0.3 and −0.6 over the long run, and occasionally even flips positive:

Unpacked, this transmission chain is a millisecond relay running from a Washington server to matching engines worldwide.

The physical layer. The 8:30 data (CPI, payrolls) is released on the BLS site and the FOMC statement is published in Washington at 2 pm; machine-readable news services like Bloomberg and Reuters parse “actual vs expected” into a structured surprise in micro- to milliseconds. The signal is then blasted over microwave links to Aurora, Chicago (home of the CME’s gold and Treasury matching engines) and New Jersey (ICE, equities); city-to-city one-way is about 4 milliseconds, well faster than fiber — a pure physics race.

The pricing layer. The surprise is first “anchored” in the rates complex — SOFR futures and 2-year Treasury futures rewrite the policy path in milliseconds, because this news is fundamentally about rates; event algorithms then sweep the euro, gold futures and other legs at market, according to pre-calibrated sensitivities (how many basis points, how many ticks, per 0.1 pp of core CPI surprise). Ahead of the release market makers have already pulled their quotes to avoid risk, spreads widen, the order book is paper-thin — so the first wave of price isn’t “traded” out so much as gapped and swept out: price first, volume after. The mirrored DXY move you see is effect, not cause: the DXY is just a weighted computation of its component pairs; EUR/USD moves first and the index is recomputed.

Once the new price is set on COMEX, the arbitrage web spreads it across the network within seconds: the EFP basis drags futures onto London spot, spot then drags the XAU crosses in each currency via the triangular calc; GLD is aligned to NAV by market makers’ hedging (back to the chain in Section 1); and if it catches the SHFE night session on Beijing time, Shanghai gold reprices within seconds, or gaps to catch up at the open if outside trading hours. Within seconds to a minute, every number called “the gold price” worldwide points back to the same anchor.

So “mirrored jumps in the same second,” taken to the bottom, is three things: the same server releasing the data, the same set of algorithms placing orders on two matching engines at once, and the arbitrage web spreading out within seconds. What’s actually more interesting is what comes after — only in the first minutes to hours do human macro funds set the tone, and the first reaction is often reversed, because algorithms only read the headline number; the details (CPI components, FOMC statement wording, the press-conference Q&A) don’t hit the price until people and slower models have digested them. To see it directly, just pull the tick data from a CPI release second: spreads snap wide, price gaps, then volume explodes — the whole baton-pass is right there in the ticks.