Research · Dataset

Gold and the yield curve inversion

"The yield curve is inverted, so a recession is coming, so gold will rally" is one of the most repeated chains of reasoning in gold commentary. It had never been measured on this desk, despite the curve being the single most watched macro signal in the world. This page runs the chain end to end over 605 months: first, does the inversion actually predict a recession; second, does gold actually do something different afterwards; third, would trading it have made money.

Data through 2026-10 · Gold source updated November 03, 2026 · Spread from fred.stlouisfed.org/series/T10Y2Y

Source and method

Gold and silver
World Bank Pink Sheet, monthly average gold price, column "Gold", US$ per troy ounce; World Bank Pink Sheet, monthly average silver price, column "Silver"
Yield spread
FRED series T10Y2Y: 10-Year Treasury Constant Maturity Minus 2-Year Treasury Constant Maturity, daily, percentage points. Averaged over each calendar month for the monthly series; the raw daily series is used for the episode table.
Second spread
FRED series T10Y3M: 10-Year Treasury Constant Maturity Minus 3-Month Treasury Bill Secondary Market Rate, daily, percentage points, averaged over each calendar month.
Recession
FRED series USREC: US Recession Indicators for the Period from the Peak of the Business Cycle through the Trough, monthly, from the NBER
Controls
FRED series DTWEXAFEGS: Nominal Broad U.S. Dollar Index, daily, averaged over the month; FRED series DFII10: 10-Year Treasury Inflation-Indexed Security, Constant Maturity, daily, averaged per month
Coverage
605 months, 1976-06 to 2026-10
Download
gold-yield-curve-inversion-monthly.csv (month, gold and silver prices, the 10y-2y and 10y-3m spreads, the NBER recession indicator, the 10-year real yield, the dollar index, and the monthly, 12-month and forward 12-month gold returns used throughout)
Source links
T10Y2Y · T10Y3M · USREC · DFII10 · DTWEXAFEGS
Reproduction
scripts/gold_yield_curve_inversion_analysis.py; independently recomputed from the raw sources by scripts/verify_gold_yield_curve_inversion.py (121 checks, 0 failures)

The 10y-2y spread starts in 1976-06, which is what limits this page to a 50-year sample. Gold prices come from 1960, but the signal does not. An episode is a contiguous negative stretch in the daily series, with negative runs separated by fewer than 21 calendar days merged into one episode, because a single positive day does not end an inversion in practice.

The headline: the curve predicts recessions, and gold is harder to read

Recession within 12m after inversion 50.5% against 15.0% without, 99 inverted months
Gold, year after an inverted month +23.8% against +6.9% otherwise
Difference +16.89 pp 95% CI -4.20 to 41.00, crosses zero
Trading it, annualised +1.85% against +8.62% for holding gold

Read those four numbers together and the story writes itself. The inversion signal is genuinely good at what it is famous for: a recession followed within 12 months in 50.5% of inverted months against 15.0% of months without one, a gap of +35.5 percentage points whose 95% interval (8.7 to 61.3) does not cross zero. So the mechanism is present.

Gold's response to the same signal is weaker and less certain. The average forward year after an inverted month returned +23.8% against +6.9% otherwise, a gap of +16.89 percentage points. The block-bootstrap 95% interval runs from -4.20 to 41.00 percentage points, and it crosses zero. The direction favours the claim; the evidence does not establish it.

And the third number is the one that matters for anyone planning to act on it. Trading the signal — hold gold for a month when the curve is inverted, hold cash otherwise — produced +1.85% a year against +8.62% for simply holding gold over the same 592 months. The rule was out of the market 83.3% of the time, and it missed the rallies that mattered most.

First, does the curve predict a recession?

This is a control, not a gold result, and it comes first because a gold finding is meaningless if the signal behind it does not do what it is famous for. The test uses the 10y-2y spread averaged over each calendar month and the NBER recession indicator, and asks whether a recession month appears anywhere in the following 12 months.

Condition at month t Months Recession within 12 months
Curve inverted (10y-2y below zero) 99 50.5%
Curve not inverted 494 15.0%
Difference 593 +35.5 pp

The difference is +35.5 percentage points, with a block-bootstrap 95% interval from 8.7 to 61.3 percentage points. This one does not cross zero, so the signal carries real information about recessions. That is the standard result in the macro literature, and it is reproduced here from raw FRED data rather than quoted.

USREC is the NBER indicator: 1 in a recession month, 0 otherwise. The forward window counts any recession month in the 12 months after the signal month.

Every inversion episode in the sample

Averages hide the mechanism. The table below is the actual record of what the curve and gold did during each of the 16 episodes, from the first in 1978 to the most recent one ending 2024-09-05.

Episode Months Deepest spread Recession? Gold during Gold 12m after
1978-08-18 to 1980-05-01 22 -2.41 pp yes +149.5% -6.6%
1980-09-12 to 1981-11-05 15 -1.70 pp yes -38.7% +0.5%
1981-12-28 to 1982-07-16 8 -0.71 pp yes -17.3% +18.9%
1988-12-13 to 1989-06-29 7 -0.45 pp no -12.2% -4.3%
1989-08-11 to 1989-10-11 3 -0.20 pp no +0.5% +3.8%
1989-11-03 to 1989-11-06 1 -0.05 pp no +0.0% -3.0%
1990-03-08 to 1990-03-29 1 -0.14 pp no +0.0% -7.6%
1998-05-26 to 1998-07-27 3 -0.07 pp no -2.0% -12.6%
2000-02-02 to 2000-12-28 11 -0.52 pp no -9.7% +1.8%
2005-12-27 to 2005-12-30 1 -0.02 pp no +0.0% +23.5%
2006-01-31 to 2006-03-29 3 -0.16 pp no +1.3% +17.6%
2006-06-08 to 2007-03-20 10 -0.19 pp no +9.9% +47.8%
2007-05-03 to 2007-06-05 2 -0.06 pp no -1.6% +35.5%
2019-08-27 to 2019-08-29 1 -0.04 pp no +0.0% +31.3%
2022-04-01 to 2022-04-04 1 -0.05 pp no +0.0% +3.3%
2022-07-06 to 2024-09-05 27 -1.08 pp no +48.4% +42.7%

Gold rose during 31.3% of the episodes, averaging +8.00% during the episode itself and +12.02% in the 12 months after it ended. Only 3 of the 16 episodes overlapped an NBER recession, which is worth pausing on: the gold move during most inversions had nothing to do with an actual recession.

The dispersion is the real finding. The best episode (1978-08-18 to 1980-05-01) saw gold rise +149.5% while the curve was inverted, and the worst (1980-09-12 to 1981-11-05) saw gold fall 38.7%. An average of +8.00% sitting between a +149.5% best case and a 38.7% worst case is not something you can position on.

9 of the 16 episodes lasted three months or less, and several sit within a tenth of a percentage point of zero, so whether a given month counts as inverted is partly a rounding convention. That is a genuine limit on the precision of any statement built on the monthly series.

Does the gold effect hold across decades?

It does not, and this is the strongest single argument against trading the signal.

Decade Months Gold after inversion Gold otherwise Difference
1980s 120 -6.80% -1.36% -5.44 pp
2000s 120 +13.84% +17.74% -3.90 pp
2020s 69 +26.74% +14.05% +12.69 pp

In the 1980s the gap was -5.44 percentage points — inverted months were worse than average, with gold returning -6.80% after them against -1.36% otherwise, and only 32.3% of inverted months producing a positive year.

In the 2000s the gap was -3.90 percentage points, again the wrong way for the claim, though the absolute returns were better: 75.0% of inverted months produced a positive year. Only in the 2020s does the gap turn positive, at +12.69 percentage points, with 100.0% of inverted months positive. That decade contains the 2022–2024 inversion, the longest in the sample, which ran through the single strongest gold advance on record.

A rule whose sign depends on the decade is not a rule. The headline +16.89 percentage point gap is a blend of a negative result across forty years and a positive one across the last five.

Is it just the dollar and real rates in disguise?

Most of the gold move in any window is explained by the dollar and real rates, so the honest test is whether the inversion flag carries information those two drivers do not already contain.

Relationship, forward 12-month window Value
Dollar change vs gold return -0.576
Real 10-year yield change vs gold return -0.493
Inversion flag vs gold return, controlling for the dollar +0.241
Inversion flag vs gold return, controlling for dollar and real yield +0.323
Silver, forward 12m after inversion +26.06%
Silver, forward 12m otherwise +8.85%

Over the 238-month control panel ( to ), the dollar's 12-month change correlates -0.576 with the forward gold return and the real yield's change -0.493. Both are large, and both are the reason a recession of any kind tends to lift gold.

But the inversion flag is not redundant with them. Controlling for the dollar leaves a partial correlation of +0.241 between the inversion signal and the forward gold return, and controlling for the dollar and the real yield together leaves +0.323. So there is information in the inversion that the other two drivers do not contain — which is the more interesting finding here than the headline average. It is also a weak signal in absolute terms, and it is a partial correlation over a sample that overlaps heavily.

Silver behaves the same way as gold, returning +26.06% after an inverted month against +8.85% otherwise. A gold-specific mechanism would be expected to show up more strongly in silver, not less.

The partial correlation asks whether the inversion dummy still carries information about the forward gold return after the linear effect of the dollar change and the change in the 10-year real yield are removed. This is a decomposition of overlapping stories, not an identification strategy: the Fed cuts rates in response to the same conditions that invert the curve, so the controls are not independent of the variable being tested.

Does the curve lead gold?

A signal that works should be visible before the price moves. This scan correlates the spread against the forward gold return from 12 months before to 12 months after the signal month.

Strongest absolute correlation 0.151 at a lag of -1 months
Correlation at lag zero -0.150 contemporaneous
12-month change correlation +0.050 95% CI -0.113 to 0.221
Level correlation +0.044 R² 0.002, meaningless by construction

The peak of the scan sits at -1 months with a correlation of 0.151, against -0.150 at lag zero. There is no clean lead: the strongest absolute value is at the edge of the window, which is what a weak contemporaneous relationship looks like rather than a leading indicator.

The level correlation of +0.044 with an R² of 0.002 is included only to show why it should never be quoted: the spread oscillates around zero while gold trends upward, so a level regression between them is close to meaningless. On 12-month changes the correlation is +0.050, with a 95% interval from -0.113 to 0.221 — wide enough to contain no relationship at all.

What would trading it have produced?

Each month, if the 10y-2y spread averaged over that month is negative, hold gold for the next month; otherwise hold cash. The signal uses only information available at the end of the signal month. No cost is charged.

Measure 10y-2y rule 10y-3m rule Buy and hold gold
Annualised return +1.85% +1.82% +8.62%
Maximum drawdown -66.70% -13.99% -62.07%
Time in the market 16.7% 11.0% 100%
Months tested 592 525 592

Both rules lost to holding gold. The 10y-2y version returned +1.85% a year against +8.62%, and it did not even reduce the drawdown: -66.7% versus -62.1%.

The 10y-3m version is the more interesting failure, and the more useful result on the page. The 3-month spread inverted far more often — 58 months against 99 for the 2-year spread — so the rule sat in cash 89.0% of the time. That halved the maximum drawdown, from -47.9% to -14.0%, while giving up most of the return: +1.82% against +6.43%. If you already hold gold and wanted to reduce the worst of it, that is the trade this data supports. It is not a return-enhancement signal, and no cost is charged in either case.

The signal is real, the mechanism is present, and the trading result is bad. Those three facts are not in conflict: the curve predicts recessions well and gold responds to recessions only sometimes, for reasons that mostly run through the dollar and real rates rather than through the inversion itself.

Where the curve stands now

10y-2y spread 0.39 pp 2026-10, 33.5th percentile since 1976-06
10y-3m spread 0.90 pp not inverted
Months inverted, 10y-2y 99 of 605 months
Months inverted, 10y-3m 58 of 605 months

Neither spread is inverted at the end of the sample. The 10y-3m spread is the one the New York Fed uses for its recession probability model, and it inverted far more often. The 10y-2y spread is the one most market commentary quotes. The 10y-2y spread at 0.39 percentage points sits at the 33.5th percentile of its own history, so it is well inside the range that has been associated with recessions without ever crossing into it.

Frequently asked questions

Does an inverted yield curve actually predict a recession?

Yes, and this page tests it first because a gold result only means something if the mechanism behind it is present. Using the 10y-2y spread averaged over each month and the NBER recession indicator, a recession started within the following 12 months in 50.5% of inverted months versus 15% of months without an inversion, a difference of 35.5 percentage points. The block-bootstrap 95% interval runs from 8.7 to 61.3 percentage points and does not cross zero. The signal does what it is famous for.

Does gold actually go up after an inversion?

On average it does, but the sample cannot rule out noise. Taking the 10y-2y spread averaged over month t as the signal and measuring the gold return over the following 12 months, gold returned 23.78% after an inverted month (99 observations) against 6.89% otherwise (493 observations), a gap of 16.89 percentage points. The block-bootstrap 95% interval is -4.2 to 41 percentage points, and it crosses zero. Because the 12-month windows overlap heavily, taking every twelfth signal month to get independent windows still gives 18.45% versus 7.73%, a gap of 10.72 percentage points, but on only 9 inverted observations.

How many separate inversion episodes are there, and what happened to gold in each?

There are 16 episodes in the sample, defined as contiguous negative stretches in the daily series with gaps of less than three weeks merged. Gold rose during 31.3% of them on average 8%, and in the 12 months after an episode ended gold averaged 12.02%. The dispersion is the real story: the best episode ran from 1978-08-18 to 1980-05-01 with gold up 149.5% during it, and the worst, from 1980-09-12 to 1981-11-05, had gold down 38.7%.

Is the effect stable across decades?

No, and this is the strongest argument against trading it. In the 1980s the 12-month gold return after an inverted month averaged -6.8% against -1.36% in non-inverted months, so inversion was worse than average by 5.44 percentage points. In the 2000s the gap was -3.9 percentage points, again the wrong way for the claim. Only in the 2020s does the gap turn positive, at 12.69 percentage points, with inverted months up 100% of the time. A rule whose sign flips by decade is not a rule.

Is the gold move just the dollar and real rates in disguise?

Partly, but not entirely. Over the 238-month control panel the correlation between the 12-month change in the broad dollar index and the forward gold return is -0.576, and between the change in the 10-year real yield and the forward gold return it is -0.493. Controlling for the dollar leaves a partial correlation of 0.241 between the inversion signal and the forward gold return; controlling for the dollar and the real yield together leaves 0.323. Silver behaves the same way, returning 26.06% after an inverted month against 8.85% otherwise. So the inversion flag carries information that the dollar and real rates do not already contain, even though most of the gold move is explained by those two drivers.

Does the yield curve lead gold, or is it just moving with it?

It does not lead it. Scanning the spread against the forward gold return from 12 months before to 12 months after, the strongest absolute correlation is 0.151 at a lag of -1 months, and the correlation at lag zero is -0.15. The peak sits at the edge of the scan rather than in the middle, which is what a weak contemporaneous relationship looks like, not a leading indicator.

What happened when you actually traded it?

It lost money. The rule holds gold for the next month whenever the monthly-average 10y-2y spread is negative, and holds cash otherwise. Over 592 months it was in the market 16.7% of the time and produced a 1.85% annualised return with a -66.7% maximum drawdown. Simply holding gold over the same window returned 8.62% a year with a -62.1% drawdown. The 10y-3m version of the same rule was in the market 11% of the time and returned 1.82% a year against 6.43% for holding gold, with a smaller -14% drawdown versus -47.9%. No cost is charged in either case.

Where does the curve stand right now?

The 10y-2y spread averaged 0.39 percentage points in 2026-10, which is the 33.5th percentile of the 605-month history, and it is not inverted. The 10y-3m spread averaged 0.9 percentage points and is also not inverted. Across the whole sample the 10y-2y spread was negative in 99 months and the 10y-3m spread in 58 months.

What is the correlation between the spread and gold on levels?

It is 0.044 on levels with an R-squared of 0.002, and it is close to meaningless, because the spread oscillates around zero while gold trends upward. On 12-month changes, which is the version that could carry a mechanism, the correlation is 0.05 with a block-bootstrap 95% interval of -0.113 to 0.221, an interval wide enough to contain no relationship. The page reports the level number only to show why it should not be quoted.

Why not use the 10y-3m spread, which is what the New York Fed uses?

Both are tested here. The 10y-3m spread inverted far more often, in 58 months against 99 for the 10y-2y, which makes it the harsher test of the same idea because it spends more time out of the market. Trading it produced a 1.82% annualised return with a -14% drawdown, against 6.43% and -47.9% for holding gold. The interesting part of that comparison is the drawdown: holding gold only during 10y-3m inversions did cut the maximum drawdown roughly in half, at the cost of most of the return.

So does the yield curve predict gold or not?

Neither answer is clean, and the page says so rather than picking the flattering one. The inversion genuinely predicts recessions: 50.5% versus 15%. And gold did in fact do better after inverted months, 23.78% against 6.89%. But the confidence interval on that gold gap crosses zero, the sign of the gap flips in two of the three decades, and trading it underperformed simply holding gold by a wide margin. The curve is a useful recession indicator and a poor gold trading signal.

What are the limits of this study?

Four. First, the sample is capped at 605 months because the 10y-2y spread starts in 1976-06; gold prices go back to 1960 but the signal does not. Second, the monthly series averages the daily spread over the calendar month, so a month that dips negative in the middle and recovers still counts as inverted, and sixteen months sit within one percentage point of zero where that call is arbitrary. Third, the forward windows overlap, so the naive sample size overstates the evidence; the independent-window figure rests on only 9 inverted observations. Fourth, gold here is a monthly average auction price rather than an executable quote, and no spread, financing or storage cost is charged in the trading tests.

Educational content only. Not investment advice. Trading forex and CFDs carries substantial risk of loss.