Analysis

Buy gold when the real yield falls, sell when it rises: the test since 2004

Published November 15, 2026

You read it everywhere: gold rises when real yields fall and retreats when they climb. We turned that into a rule, the simplest one possible — one reading of the US five-year real yield at the start of each month, compared with the reading a year earlier — and applied it from 2004 to 2026, trading costs included, against someone who buys gold once and never sells.

Holding gold when the real yield is down on the year, against staying invested Gold (Ounce), in US dollars, from January 1, 2004 to November 2, 2026 — ×9.5 against ×2.5. 100 200 400 800 1600 2005 2010 2015 2020 2025 ×9.5 invested throughout ×2.5 real-yield rule Rebased to 100 on 1 January 2004. Costs of 0.5% per order. Log scale.
Two value curves since January 2004, rebased to 100: staying invested throughout reaches 953, the real-yield rule 249.

From January 2004 to November 2026, with costs of 0.5% per order, the rule turns $1,000 into $2,494. Someone who bought on the same day and never touched it again has $9,529 — close to four times as much.

From 1 January 2004 to 2 November 2026, in US dollars, costs of 0.5% per order. Same period and same prices for both.
Strategy Total return Annualised Volatility Largest drawdown
Real-yield rule ×2.49 +4.08% a year 12.22% −35.90%
Invested throughout ×9.53 +10.38% a year 15.80% −40.68%

The rule, word for word

The yield we follow is that of five-year US Treasury inflation-protected securities — what lending to the US government pays once inflation is taken out, hence the name real yield. The Federal Reserve Bank of St. Louis has published it every day under the code DFII5 since 2 January 2003.

Once a month, at the first value published that month, compare the yield with the same month a year earlier. Lower than a year ago: hold gold, buying it if you have none. Higher: hold none, selling if you do. Then look at nothing until the following month. The two readings were never equal over the period: the rule always had an opinion.

The first comparison needs a year of history, so the first decision falls in January 2004. On that day the yield stands at 1.28%, against 1.75% a year earlier — it has fallen, the rule buys, at $414.50 an ounce. The last reading is that of 2 November 2026: 2.31%, against 1.31% a year before. In between, 275 readings and 28 orders.

On the other side, someone buys gold on the same day in January 2004, at the same price, and leaves it alone. Both pay 0.5% on every order: once for the holder, twenty-eight times for the rule.

The relationship exists — in the rear-view mirror

The starting idea is no myth. Across the 136 months when the real yield was lower than a year before, gold had gained +20.86% on average over those same twelve months, and was up 119 times. Across the 139 months when the yield was higher, gold had gained only +5.21%, up 84 times. The correlation between the one-year change in the yield and the change in gold over the same year is −0.44: the two do move in opposite directions.

But read the sentence again: “had” gained. When the rule observes that the yield has fallen over the past year, the rise in gold that came with it has already happened. It is in the price you buy at. What the rule must predict to make money is not the year gone by, but the month ahead.

Over the following month, the signal separates nothing

And there, nothing is left. In the month after a “hold” signal, gold gains +0.899% on average. In the month after a “sell” signal, it gains +0.960% — slightly more. Gold rises 78 times out of 136 in the first case, 76 times out of 138 in the second.

The gap is six hundredths of a point, in the wrong direction, against a monthly dispersion of 4 to 5%. Welch’s test on the difference between the two means gives a t of −0.11: as close to zero as a test gets. Over twenty-three years, knowing whether the real yield is lower or higher than a year before says nothing about what gold will do the next month.

Yet over those twenty-three years gold rose in both cases. So the rule does exactly what any rule does when it leaves a rising asset for no good reason: it is absent half the time — invested for only 136 months out of 274 — and misses half the rise.

Three absences that cost nearly everything

The final gap does not come from steady erosion. It comes from three periods when the real yield was rising, when the rule was therefore out, and when gold rose anyway.

From July 2005 to April 2007, the US central bank raises rates and the real yield moves from 1.50% to above 2%. The rule sells at $424.42 and only buys back at $679.54: +60.11% without it.

From November 2008 to September 2009, in the thick of the financial crisis, the real yield jumps to 3.66% — inflation-linked bonds are dumped along with everything else. The rule sells at $760.11, at the very moment gold’s great run begins, and buys back at $996.24: +31.07% missed.

From January 2022 to August 2024, the real yield goes from −1.58% to nearly 2%, the sharpest climb in the whole series. By the relationship, gold should have collapsed. It goes from $1,797 to $2,497: +38.95%, again without the rule.

Together, these three absences multiply the gold price by 2.9. Of the thirteen times the rule left and came back, it bought back cheaper than it had sold only four times, and never by more than 4.30%. On average, it bought back 12.11% higher.

What costs change

Final value of 1 invested on 1 January 2004, by the cost charged on each order. The rule places 28 orders, the permanent holder a single one.
Cost per order Real-yield rule Invested throughout
0% ×2.87 ×9.58
0.1% ×2.79 ×9.57
0.5% ×2.49 ×9.53
1% ×2.17 ×9.48
2% ×1.63 ×9.39

Costs make it worse, but they do not explain it

We used 0.5% per order: more than a listed gold-backed product costs at an online broker, bid-ask spread included, and well below physical gold, where the gap between the price a dealer sells a coin at and the price it buys it back at commonly reaches several per cent.

At that rate, the twenty-eight orders cost the rule 13% of its final value. That is real, but it is not the point: with no costs at all, it ends at ×2.87 against ×9.58. The rule does not lose because it is expensive to run. It loses because it gets out at the wrong time, and costs merely add to the bill. At 2% per order — the order of magnitude for physical gold — ×1.63 is left: twenty-three years to make 63%.

Ten round trips out of fourteen are winners before costs; only six remain after. Four of those ten gains are so thin — +0.49%, +0.65%, +0.76%, +0.78% — that the two commissions are enough to wipe them out.

What if the cash earned interest during the absences?

The objection is fair: when the rule leaves gold, yields are rising, and a rising yield is precisely what pays cash. The calculation above leaves it idle at 0%.

At 2% a year, cash lifts the rule to ×3.13. At 4% — a rate the dollar offered risk-free for only part of the period — to ×3.92. To match the investor who did nothing, it would have taken 12.36% a year, risk-free, in each of the 138 months spent out of gold. No risk-free investment has paid that since 2004.

What the rule buys: a little calm, at a very high price

Absent half the time, the rule takes fewer knocks: annualised volatility of 12.22% against 15.80%, largest drawdown of −35.90% against −40.68%. The difference is slim. Both portfolios suffer the same fall, the one that follows the 2011 peak — the rule was invested at the top and only got out in June 2013, at $1,232, once most of the decline had been taken.

Set against its volatility, the rule’s annualised return gives 0.33, against 0.66 for buy-and-forget. It does not cut risk and return in the same proportion: it removes a quarter of the risk and three fifths of the return.

Decade by decade, the finding never changes direction. From 2004 to 2010: −1.01% for the rule, +168.37% for gold held without interruption. From 2010 to 2020: +36.35% against +41.77%. Since 2020: +84.78% against +150.45%. It came out ahead in none of the three periods.

Where it was right

It was not always wrong, and it would be dishonest to hide that. Its three long holding periods are genuine rises: +23.66% from September 2009 to June 2013, +40.39% from April 2019 to January 2022, +53.82% from March 2025 to April 2026.

And its latest decision is, so far, the right one. It sold on 1 June 2026 at $4,475.20, as the real yield moved back above its level of a year earlier; on 2 November 2026 gold trades at $3,969.60, 11.30% lower. That is its best exit in twenty-three years.

One successful exit does not make a rule. But it is a reminder of what this test measures: not whether the real yield matters for gold — it does — but whether a reading taken at the start of the month is enough to know when to get in and when to get out.

The fourteen round trips, one by one

Gold price in US dollars at the month’s reference point. The last column takes off 0.5% on the purchase and 0.5% on the sale. Since June 2026 the rule has been out of the market.
Bought Price Sold Price Return After costs
Jan 2004 $414.50 Jun 2004 $392.51 −5.31% −6.25%
Aug 2004 $400.19 Oct 2004 $421.28 +5.27% +4.22%
Nov 2004 $438.35 Feb 2005 $423.85 −3.31% −4.27%
May 2005 $422.37 Jul 2005 $424.42 +0.49% −0.52%
Apr 2007 $679.54 Jun 2007 $655.42 −3.55% −4.51%
Oct 2007 $754.21 Nov 2008 $760.11 +0.78% −0.22%
Sep 2009 $996.24 Jun 2013 $1,232.00 +23.66% +22.43%
Apr 2015 $1,179.00 May 2015 $1,188.00 +0.76% −0.24%
Mar 2016 $1,232.00 Apr 2016 $1,240.00 +0.65% −0.35%
May 2016 $1,212.00 Mar 2017 $1,243.00 +2.56% +1.53%
Apr 2019 $1,280.00 Jan 2022 $1,797.00 +40.39% +38.99%
Aug 2024 $2,497.00 Jan 2025 $2,835.00 +13.54% +12.40%
Mar 2025 $3,109.70 Apr 2026 $4,783.20 +53.82% +52.28%
May 2026 $4,629.90 Jun 2026 $4,475.20 −3.34% −4.31%

What this test does not say

It does not say that real yields have no effect on gold. It says that a mechanical rule, built on a single monthly reading and a one-year lookback, gets nothing out of it. The relationship between the two is contemporaneous: it describes what has just happened, not what comes next.

It covers only twenty-three years, because the five-year real-yield series starts in 2003. That is one long period, dominated by two gold bull markets. A regime in which gold fell for years while real yields rose would give a different result — over 2013–2015 the rule did in fact sidestep a decline.

We tested the rule as stated, with a twelve-month lookback, and did not look for the lookback that would have given the best result: finding one on the same history would be the product of the sorting, not of the market. As a check only, six- and twenty-four-month lookbacks also lose very heavily to simply buying.

Finally, everything is in US dollars and no tax is deducted. In most countries every sale of gold is a taxable event: fourteen sales for the rule, none for the holder. The real gap is therefore wider than the one on this page.

Frequently asked questions

Should you buy gold when real yields fall?

Not on the strength of that signal alone. From 2004 to 2026, holding gold only when the US five-year real yield is lower than a year earlier gives ×2.49, with costs of 0.5% per order, against ×9.53 for holding it without interruption.

Should you sell your gold when real yields rise?

Over 2004–2026 it was costly. In the months after a one-year rise in the real yield, gold gains 0.960% on average, against 0.899% after a fall. The rule bought back higher than it had sold nine times out of thirteen; across its thirteen returns, 12.11% higher on average.

Do gold and real yields really move in opposite directions?

Yes, over the same period: the correlation between the one-year change in the real yield and the change in gold over that year is −0.44. But the relationship describes the year gone by. It does not predict the next month: Welch’s t between the two signal states is −0.11.

What is the DFII5 series?

The yield on five-year US Treasury inflation-indexed securities at constant maturity. It is a measure of the real interest rate, published every business day by the Federal Reserve Bank of St. Louis since January 2003.

Do trading costs explain the gap?

No. With no costs at all, the rule ends at ×2.87 against ×9.58. Costs make it worse — ×2.49 at 0.5% per order, ×1.63 at 2% — but most of the gap comes from the rises in gold missed during the 138 months spent out of the market.

Did the rule at least reduce risk?

A little: volatility of 12.22% against 15.80%, largest drawdown of −35.90% against −40.68%. But its return per point of volatility is 0.33, against 0.66 for gold simply held.

Method and sources

Gold prices (GC=F contract), in US dollars, read from Finamax’s public API; the DFII5 yield read from FRED, the data service of the Federal Reserve Bank of St. Louis. Both series stop on 2 November 2026. For the yield, each month is represented by its first published value. For gold, by its monthly reference point, the convention of /api/prices.

That point deserves a note. Before 2026 the database keeps a single gold price per month, dated the 1st, and that price is not a quote from the 1st: set against the contract’s daily closes, it coincides with the month’s average until 2010, and with a value close to its close thereafter. Before 2026 the rule therefore decides on a price it could not have known exactly on that day. We reran the calculation executing every decision one month later, which leaves only the past to read: ×2.22 for the rule against ×9.76. The conclusion does not hinge on this detail.

Both strategies start with the same capital on 1 January 2004 and are valued on the same points. Costs are 0.5% of the amount of each order, purchase and sale alike. Cash earns nothing unless stated otherwise. The position held on the last date is marked at the price, with no exit cost. No tax is deducted.

Annualised return, volatility and largest drawdown are computed by shared/financeMath.js, the module from which the site’s comparator draws its own indicators. The backtest lives in shared/signalBacktest.js, and the report is regenerated with node scripts/backtest-signal.js --rule real-yield --fee 0.005 --end 2026-11-02.

Past performance is no guide to future performance. This page is an analysis of historical data and does not constitute investment advice.