Key takeaways
- Gold set a new record on 2.78% of its 14,654 London fixes since 1968, against 27.0% of months for the S&P 500 since April 1968.
- After a gold record month, real returns ran 24.2% over one year but 4.4% over three and 2.6% over ten, against baselines of 4.5% and 2.9%.
- A 5% gold sleeve rebalanced each December bought gold at an average 34.8% below its record and sold at 17.1% below, across 55 years.
- Erb and Harvey's gold-to-CPI ratio predicted about -6% a year from March 2012; gold actually returned 1.57% a year nominal to March 2022.
- Gold ETF holdings peaked at a record 4,176 tonnes on 27 February 2026, then March brought US$12bn of outflows, the largest month on record.
Picture two people reading the same headline on 29 January 2026. Gold has set a record. The London afternoon fix is $5,405.00 an ounce, the highest price in a benchmark series that runs back to 1 April 1968. What would you have done next?
The first of them already holds a 5% gold sleeve and does nothing, because her rule doesn't let her touch it until December. The second buys, because the line on the chart is going up and waiting feels like the expensive choice.
By 4 August 2026 the fix was $4,084.20, or 24.4% below the record. Put money in at the top and you were down roughly a quarter inside six months. Hold a fixed weight and you were holding roughly what you always held.
So did the second one miss it, or catch it? That's the question a record price puts in front of you, and it's really two questions wearing one coat. The first is what fixed share of your portfolio belongs in gold at all. The second is whether to put money in after a run. Most of the published research answers the first question. It says surprisingly little about the second, and the two aren't the same decision at all.
The allocation question already has a piece of its own on this site, covering the 5-10% gold weight and its 45-year real drawdown. This is the other half: what the record itself tells you, if anything, and what a fixed weight does that a purchase at the top doesn't.
Buying gold at a record high is rarer than buying shares at one
Start with how often it even happens, because that decides how much weight a record deserves in your head. The London Bullion Market Association has published 14,654 afternoon gold fixes since April 1968. Exactly 408 of them set a new all-time high. That's 2.78% of trading days. How you take the position matters as much as when: the three ways to invest in gold track different things and are taxed differently.
Month-end data makes the comparison with equities cleaner. Take the 666 months from April 1968 to September 2023, which is where Robert Shiller's public S&P 500 series ends. Gold's month-end fix set a record in 66 of those months, or 9.9%. The S&P 500 price index set one in 180 of them, or 27.0%.
Widen the definition and the gap widens too. The S&P 500 sat within 5% of its running record in 50.5% of those months. Gold managed that in 19.2%. Within 10% of a record, it's 59.6% for shares against 29.1% for gold.
Now think of the waiting in between. After the fix of $850.00 on 21 January 1980, gold didn't set another record until 3 January 2008. That's 10,209 days, or a shade under 28 years. How long would you sit with something that did that to you?
The strongest objection: a rising asset lives near its high
Here's the counter-argument, and it's a good one. Anything that compounds upward spends most of its life at or close to a record, so "record high" carries no information. Waiting for a non-record entry point in a rising market means waiting more or less forever, and the wait costs you the return you were trying to buy cheaply.
For equities that objection is simply correct. Half of all months since 1968 put the S&P 500 within 5% of a record. Someone who refused to buy shares near a high spent half their investing life on the sidelines, and the forward returns show no penalty for buying anyway. We'll come to those in a moment.
But the objection is much weaker for gold. Gold spent four fifths of those months more than 5% below its own record, and it once went 28 years without setting one. Gold isn't a compounding asset. It has no earnings, pays no coupon, and does nothing except sit there. Its price chart is a series of long plateaus punctuated by short violent repricings, and 192 of its 408 record days fell before 1981. A gold record genuinely is an unusual state. Whether unusual means predictive is the next question.
What actually followed buying gold at a record high
If you want to check any of this yourself, here's exactly what it's built from. Same 666 months, same method, forward returns after inflation. Gold is a price return, because gold has no income. The S&P 500 is a total return including dividends. The two columns aren't comparable to each other in level, but each is comparable to its own baseline.
| Held from | Gold: after a record month | Gold: any month | S&P 500: after a record month | S&P 500: any month |
|---|---|---|---|---|
| 1 year | 24.2% | 5.5% | 7.3% | 7.2% |
| 3 years | 4.4% | 4.5% | 8.3% | 6.8% |
| 5 years | 3.4% | 3.9% | 7.5% | 6.6% |
| 10 years | 2.6% | 2.9% | 6.2% | 6.8% |
Annualised real returns, April 1968 to September 2023, computed from LBMA month-end fixes and Shiller's series, deflated by US consumer prices.
Read the gold columns first. Buying at a record was followed by a spectacular twelve months: 24.2% a year in real terms, against 5.5% for a randomly chosen month. That's momentum, and it's large. It's also the number your second reader was implicitly betting on when she bought the headline.
Then it vanishes. At three years the record months returned 4.4% against a baseline of 4.5%. At five years, 3.4% against 3.9%. At ten, 2.6% against 2.9%. Those differences are noise. On this evidence a gold record is neither a warning nor a green light beyond about a year.
The equity columns tell the same story with less drama. Records were followed by 7.3% over one year against 7.2% for any month, and by 6.2% over ten years against 6.8%. Nothing there would justify sitting out.
One episode is doing most of the work
So that settles it, doesn't it? Not quite. Those 66 record months aren't 66 independent observations. They cluster into a handful of episodes, and the overlapping ten-year windows recycle the same data many times over.
Of gold's 408 record fixing days, 25 fell in the 1960s, 160 in the 1970s, seven in the 1980s, 41 in the 2000s, 59 in the 2010s and 116 since 2020. Restrict the test to real records, adjusted for inflation, and only 32 months out of 698 qualify since April 1968 — of which 18 are from the 1970s and 11 from 2024 onwards.
So when you look at gold's three-year and five-year figures, you're mostly looking at the aftermath of 1980 blended with the aftermath of 1974. Two episodes. The 2024-2026 records don't have three years of history behind them yet, so they contribute nothing to those rows. Anyone claiming a firm law here is over-reading a very small sample, and that includes me.
The best case against buying gold at a record high, and how it fared
There is one serious academic argument that a high gold price predicts poor returns. Claude Erb and Campbell Harvey made it in "The Golden Dilemma", published as NBER working paper 18706 in January 2013 and in the Financial Analysts Journal later that year. Their measure is the ratio of the gold price to the US consumer price index.
Since gold futures began trading in January 1975 that ratio has averaged about 3.2. It bottomed at 1.46 in March 2001 and peaked at 8.73 in January 1980. When they wrote, in March 2012, it stood at 7.3. Their arithmetic was blunt: if the ratio reverted to 3.2 over the following decade, gold would return roughly -6% a year.
Suppose you had acted on it. You'd have sold out, or refused to buy, in March 2012, when gold's month-end fix was $1,662.50. Ten years later, in March 2022, that same ounce fixed at $1,942.15. You'd have sidestepped +1.57% a year in nominal terms and -0.70% a year after US inflation — mildly disappointing, and nothing like the -6% you were promised. The ratio itself barely moved, from 7.25 to 6.76. The reversion the model implied simply didn't arrive over the horizon the paper suggested.
Which leaves the metric in an awkward place, because on 4 August 2026 it reads about 12.2. That's not just above the long-run average of 3.2. It's about 40% above the January 1980 peak that has defined "expensive gold" for two generations, and the ratio touched 16.0 in February 2026. So is that the loudest valuation warning in the series' history, or a signal that failed its one out-of-sample test and is now failing louder? Erb and Harvey were honest enough to frame it as a dilemma rather than a forecast, and that framing has aged better than the number.
What a strategic gold weight does instead
Back to the first of your two people — the one whose rule wouldn't let her act until December. Here's what that rule actually did, tested over the same 55 years. Take a portfolio of 5% gold and 95% S&P 500 total return, rebalanced every December, from April 1968 to September 2023.
The rebalanced version compounded at 10.29% a year. The same portfolio left alone compounded at 10.13%. Sixteen basis points. As a return story it's thin, and Vanguard's own work points the same way: its December 2024 research puts the benefit of threshold-based rebalancing over monthly rebalancing at 15-22 basis points a year during accumulation. The site's comparison of annual rebalancing against 5% threshold bands goes through that trade-off in more detail.
So why bother at all? Because the return was never the point. What matters is when the rule traded. Over those 55 Decembers the rebalance bought gold 33 times and sold it 22 times. When it bought, gold was on average 34.8% below its own record. When it sold, gold was 17.1% below — and six of those sales came with gold within 5% of a record.
That's the whole thing in one line. A fixed weight is an instruction to buy an asset after it's fallen a third and to trim it after it's run. Nobody has to forecast anything. The rule takes the entry-timing decision off your hands by making it automatic and contrarian, which is roughly the opposite of what a purchase at a record does.
None of which means the weight sits still. In that backtest the December gold weight ranged from 2.79% in 1975 to 10.58% in 1974, so a 5% target spent plenty of time nowhere near 5%.
The current cycle is no gentler, and it's worth walking through slowly, because it's where a lot of people are standing right now. Say your sleeve began 2025 exactly on target at 5%, held as a single ounce. Gold rose 67.4% during 2025, from $2,609.10 on 30 December 2024 to $4,367.80 on 30 December 2025. So the ounce you started the year holding at $2,609.10 was worth $4,367.80 by the end of it. If everything else in your portfolio had been flat, your 5% sleeve would have finished the year at 8.10% without you touching it.
Which is why, when the headlines said record in January 2026, the disciplined half of this story was about to be a seller of gold rather than a buyer. Not because she had a view. Because the drift had made her one.
Where the money actually went
Flows suggest most people did the opposite. World Gold Council data shows global gold ETF holdings peaking at a record 4,176 tonnes on 27 February 2026, within a month of the record price. Then March brought US$12bn of global outflows, the largest monthly outflow the series has recorded, with North America alone accounting for US$13bn of selling. By the end of May holdings were 4,121 tonnes and US$604bn of assets. Reading a large institution's holding as a signal is the error at the centre of institutional crypto allocation coverage.
Peak holdings at the peak price, record redemptions on the way down. If you've ever sold something after it fell and felt relief rather than regret, you already know how that happens from the inside. It's the pattern documented in the gap between fund returns and what investors actually earn, showing up in gold in real time. It also sits oddly beside the physical demand picture, where gold demand reached a record 5,002 tonnes in 2025 on very different buyers with very different motives.
What would change the conclusion
Several things, and some of them are likely.
The sample is the biggest weakness. Sixty-six record months across roughly four episodes isn't enough to establish anything firm about three-year or five-year outcomes, and overlapping windows make the standard errors worse than the counts suggest. If gold's 2024-2026 records are followed by a 1980-style decade, those rows will look very different by 2036. If they're followed by another leg up, they'll look different in the other direction.
Everything here is in US dollars. If you hold in sterling or euros, you own a currency bet alongside the metal, and your record dates aren't the same dates as these. The equity comparison also stops in September 2023, because that's where the public Shiller file ends, so the last three years of the S&P are absent from that side of the table.
The regime argument would break the valuation case entirely. If central bank buying has permanently shifted demand, then Erb and Harvey's ratio may never revert, and a reading of 12.2 means nothing more than that the old average described an old world. Plenty of people argue that. They may be right, and there's no way to test it except by waiting.
And the fixed-weight result depends on the rest of your portfolio. Paired with 95% equities, a 5% gold sleeve rebalances into gold after equity strength. Paired with bonds, or with a different target weight, the trade dates and the buy-side drawdowns shift. The 34.8% average discount at purchase isn't a constant of nature. It's what one specific rule produced against one specific portfolio over one specific stretch of history.
So where does that leave the two people at the top? Somewhere narrower than either of them would like. Beyond a year, gold's record highs haven't predicted much in either direction, so the fear of buying at the top isn't well supported by gold's own history. What is well supported is that a fixed weight and a purchase at a record are different decisions with different risk profiles, and only one of them needs you to be right about what happens next.