Research · Dataset

Does buying gold dips actually work? 1968–2026

"Buy the dip" is the most repeated piece of gold advice there is, and the easiest to believe: gold falls, you buy, gold rises. This page tests that instruction on 14,722 daily fixings — and the test matters more than the claim, because on a metal that drifted upward for fifty-eight years, a dip rule looks good for a reason that has nothing to do with timing. The short answer: there is a small, short-lived rebound effect and a large amount of drift, and separating the two changes the conclusion.

Data through 2026-11-10 · updated 2026-11-15

Source and method

Price series
LBMA Gold Price PM benchmark fix, USD per troy ounce — scripts/daily_gold_lbma_pm.csv
Sample
14,722 fixings, 1968-04-02 to 2026-11-10 (58.6 years, 14,721 daily returns)
Gap rule
Fixings separated by more than 10 calendar days are dropped, so a long data gap never appears as a single huge return.
Excess return
Forward return after a dip minus the unconditional mean forward return over the same horizon and the same sample. Every comparison on this page is an excess return for that reason.
Inference
Forward windows overlap, so ordinary t-statistics would overstate significance. Every t-statistic here uses a Newey-West HAC estimator with a Bartlett kernel and 10 lags, and each event study also carries a moving-block bootstrap 95% interval.
Costs
Strategy simulations charge a flat 0.30 USD per troy ounce round trip and never overlap positions.
Download
gold-dip-buying-events.csv (date, fix, daily return, drawdown from the 60-fixing high, event flags, forward returns)
Recompute
scripts/gold_dip_buying_analysis.py, independently rechecked with numpy by scripts/verify_gold_dip_buying.py (267 checks over the published figures)

The headline numbers

Excess after a ≥3% drop, next day +0.33 pp t = 2.68 · 206 events · the one significant result
Excess after a ≥3% drop, 20 days +0.82 pp t = 1.36 · not distinguishable from zero
Buy and hold CAGR 8.33% 108.6× over 58.6 years · max drawdown -70.3%

The unconditional forward return over this sample was 0.039% over one fixing, 0.792% over 20, and 2.419% over 60. That is the number any dip rule has to beat. A rule that returns +1% may simply be collecting the drift.

Step one: is gold even predictable at daily frequency?

If gold behaved like a random walk, then no daily rule could add anything, and "buy the dip" would be indistinguishable from setting a recurring buy order. The evidence says gold looks like a random walk with a positive drift — the least favourable regime for a timing rule.

Variance ratio (q) VR(q) z-statistic Reading
2 0.979 -1.79 consistent with a random walk
5 0.957 -1.05 consistent with a random walk
10 1.014 +0.16 consistent with a random walk
20 1.108 +0.59 consistent with a random walk
30 1.046 +0.17 consistent with a random walk

Under a random walk, VR(q) = 1. At q = 20 the variance ratio is 1.108 with a z-statistic of +0.59 — nowhere near the ±2 threshold, so the null is not rejected. The autocorrelation of daily log returns is -0.0188 at lag 1 (just outside the 95% white-noise band of 0.0162) and -0.0022 at lag 5 (inside it). The one-day reversal is real but tiny; it is not a forecasting signal anyone could trade on.

Step two: what actually follows a drop?

Here is the event study. After a day of at least X% decline, what is the forward return, and how does it compare with the unconditional average over the same horizon? The t-statistics use the Newey-West estimator because consecutive events overlap.

Trigger Horizon Events Mean forward Excess vs unconditional HAC t Significant
-1% day 1d 1827 +0.067% +0.028 pp +0.83 no
-1% day 5d 1826 +0.292% +0.096 pp +0.95 no
-1% day 20d 1823 +0.939% +0.147 pp +0.46 no
-1% day 60d 1813 +3.153% +0.733 pp +0.85 no
-2% day 1d 573 +0.167% +0.127 pp +1.65 no
-2% day 5d 573 +0.245% +0.049 pp +0.27 no
-2% day 20d 572 +0.875% +0.084 pp +0.18 no
-2% day 60d 569 +3.292% +0.872 pp +0.64 no
-3% day 1d 206 +0.371% +0.332 pp +2.68 yes
-3% day 5d 206 +0.889% +0.693 pp +1.96 no
-3% day 20d 206 +1.614% +0.822 pp +1.36 no
-3% day 60d 205 +4.695% +2.276 pp +1.33 no
-4% day 1d 91 +0.507% +0.468 pp +3.05 yes
-4% day 5d 91 +1.098% +0.903 pp +1.84 no
-4% day 20d 91 +0.901% +0.110 pp +0.10 no
-4% day 60d 90 +3.833% +1.414 pp +0.64 no
-5% day 1d 50 +0.948% +0.909 pp +2.62 yes
-5% day 5d 50 +1.299% +1.104 pp +1.28 no
-5% day 20d 50 +0.445% -0.347 pp -0.18 no
-5% day 60d 50 +2.321% -0.099 pp -0.03 no

The pattern is clean. At a one-day horizon, big drops are followed by a small bounce: a ≥3% drop brought a +0.33 pp excess return (t = 2.68, 95% CI 0.01 to 0.63 pp) and a ≥4% drop brought +0.47 pp (t = 3.05). Those are genuine short-horizon mean reversion.

At twenty days it is gone. The ≥3% excess falls to +0.82 pp (t = 1.36) and a ≥5% drop produced a -0.35 pp excess — the wrong sign. The rebound decays into the drift within about a month.

The mirror control: what follows a jump?

A mean-reversion effect has to be symmetric in the sense that matters: if a drop predicts a bounce, an equal-sized jump should predict a pullback. Otherwise the "dip" effect is really just the fact that gold goes up.

Trigger Horizon Down events Down excess Up events Up excess
±3% day 1d 206 +0.332 pp 236 -0.021 pp
±3% day 5d 206 +0.693 pp 236 +0.251 pp
±3% day 20d 206 +0.822 pp 236 +0.862 pp
±3% day 60d 205 +2.276 pp 236 +2.111 pp

After a ≥3% up day the 20-day excess is +0.86 pp with a t-statistic of 0.76 — statistically indistinguishable from zero, and statistically indistinguishable from the down-side excess at the same horizon. The two event sets behave the same way. The dip is not a special setup; it is just a day.

Step three: the dip people actually mean

A single bad day is not what most people mean by "a dip". The everyday definition is a pullback from the recent high, so the next test uses the drawdown from the highest fixing in the trailing 60 fixings.

Drawdown from 60-fixing high Events 5-day excess 20-day excess 60-day excess 20-day HAC t
≤ −-5% 5699 -0.100 pp -0.446 pp -0.509 pp -2.10
≤ −-8% 2924 -0.015 pp -0.076 pp +0.444 pp -0.23
≤ −-10% 2005 +0.099 pp +0.027 pp +0.792 pp +0.06
≤ −-15% 562 +0.758 pp +2.021 pp +3.327 pp +2.08

The ordinary dip does not work. A 5% pullback from the 60-fixing high was followed by a 20-day return -0.45 pp against the unconditional average (t = -2.10) — significantly worse than doing nothing. A 10% pullback was neutral (+0.03 pp, t = 0.06).

Only the deep dip survives. After a 15% drawdown the 5-day excess was +0.76 pp (t = 2.08) and the 20-day excess was +2.02 pp (t = 2.08), with a block-bootstrap interval of 0.64 to 3.42 pp. There were only 559 such events in 59 years — a rule that only works at 15% drawdowns is not the rule most people are following.

Is it stable across decades?

An effect that appears in one decade and reverses in another is not a property of gold. Here is the 20-day excess after a 10% dip, by decade.

Decade Events Mean forward Excess vs unconditional HAC t
1970s 347 +1.676% +0.885 pp +0.87
1980s 724 -0.304% -1.096 pp -1.29
1990s 213 +0.438% -0.354 pp -0.62
2000s 280 +2.814% +2.022 pp +2.13
2010s 231 +1.387% +0.595 pp +0.66
2020s 160 +0.530% -0.262 pp -0.20

The excess is positive in 3 decades and negative in 3. The 2000s carry the effect (+2.02 pp) and the 2020s are slightly negative. That is a sample, not a mechanism.

What happens if you actually trade it?

The rule: buy when gold is at least X% below its 60-fixing high, hold for a fixed number of fixings, flat otherwise, no overlapping positions, 0.30 USD per ounce round trip.

Dip threshold Hold Trades Win rate Final multiple CAGR Max drawdown
≤ -5% 20d 394 52.3% 1.314× +0.466% -70.98%
≤ -8% 20d 221 52.9% 2.264× +1.404% -55.63%
≤ -10% 20d 151 55.6% 1.865× +1.069% -51.84%
≤ -10% 60d 72 56.9% 2.139× +1.306% -51.77%
≤ -15% 20d 57 57.9% 2.119× +1.289% -43.54%
≤ -15% 60d 31 64.5% 3.998× +2.393% -42.79%
Buy and hold — 1 — 108.603× +8.327% -70.26%

Every dip rule underperformed simply owning gold. The 10%-dip, 20-day rule produced 151 trades at a 55.6% win rate for a CAGR of 1.07% against 8.33% for buy and hold. The best CAGR of any dip rule was 2.39% (a 15% dip held 60 days, 31 trades) — roughly 5.9 percentage points of annual return handed back every year.

What the rules do buy is drawdown. The shallowest rule cut the worst peak-to-trough decline from -70.3% to -42.8% — a 27-point improvement. That is a genuine benefit for an account that would otherwise have been liquidated. It is a benefit bought with return, not a source of return.

What this study does and does not show

Supported: a small short-horizon rebound after a sharp single-day drop. A ≥3% drop was followed by a +0.33 pp excess return the next day with a HAC t-statistic of 2.68 and a bootstrap interval that excludes zero. The effect is real and it is small, and it is largely gone after a month.

Not supported: the claim that buying dips is a reliable way to beat holding gold. Over 20 fixings the excess after a ≥3% drop was +0.82 pp (t = 1.36); after an ordinary 5% pullback it was -0.45 pp (t = -2.10) — worse than doing nothing; and the effect reversed sign in 3 of 6 decades. Every dip rule tested underperformed buy and hold by roughly 5.9 percentage points of annual return.

Limits: The LBMA PM fix is a benchmark auction price, not a tradable bid/ask. A strategy that buys at the fixing could not actually transact at that price without paying the spread. The $0.30 per ounce round-trip cost is a flat assumption. Real spreads widen in exactly the conditions that generate the largest dips, so the true cost of catching a dip is understated. The signal is defined on the fixing and executed at the next fixing, so no intrabar fill is assumed. This flatters a rule that would in practice be triggered by an intraday price. Event studies on overlapping forward windows are not independent; every t statistic here uses a Newey-West HAC estimator for that reason. This is a price-path study. It says nothing about position sizing, leverage, margin calls, or whether the drawdowns a dip-catching rule experiences are survivable at real leverage. These are arithmetic results on one historical price path. Nothing on this page is investment advice, a recommendation, or a forecast.

Frequently asked questions

Does buying gold dips actually work?

Partly, and much less than the folklore claims. A >=3% drop was followed by a 0.33 percentage point excess return over the next day, with a Newey-West t-statistic of 2.68 across 206 such days — that is the one clean, statistically significant result in this study. Extend the horizon to 20 days and the excess falls to 0.82 percentage points with a t-statistic of 1.36, which is not distinguishable from zero. The rule does not have a durable edge; it has a short-lived rebound that mostly decays within a month.

Why do you compare against an unconditional average instead of against zero?

Because gold drifted upward over the sample: the unconditional forward return was 0.792% over 20 fixings and 2.419% over 60. On an asset with positive drift, anything you buy tends to go up, so a strategy that returns +1% may simply be collecting the drift rather than predicting anything. Every figure on this page is therefore an excess return: the forward return after a dip minus the unconditional mean forward return over the same horizon and the same sample. That is the only comparison that can distinguish a dip rule from simply owning gold.

Is gold predictable at all at the daily frequency?

Barely, and not in a way that supports a dip rule. The one-day autocorrelation of daily log returns is -0.0188, which sits just outside the 95% white-noise band of 0.0162, and the five-day autocorrelation is -0.0022, inside it. Variance ratio tests are consistent with a random walk: at q=20 the variance ratio is 1.108 with a z-statistic of 0.59, nowhere near the ±2 threshold. Gold at daily frequency looks like a random walk with a positive drift, which is precisely the regime in which "buy the dip" is easiest to believe and hardest to justify.

What is the strongest evidence in favour of the dip rule?

Deep dips. After a 15% drawdown from the trailing 60-fixing high, the 20-fixing forward return was 2.02 percentage points above the unconditional average, with a t-statistic of 2.08 and a block-bootstrap 95% interval of 0.64 to 3.42. The 5-fixing excess was 0.76 points, also significant. But there were only 559 such events in 58 years, and a rule that only works on 15% drawdowns is not the rule that most people are actually following.

What is the strongest evidence against the dip rule?

The ordinary dip. A 5% drawdown from the trailing 60-fixing high was followed by a 20-fixing return 0.45 percentage points BELOW the unconditional average, with a t-statistic of -2.1 — that is, significantly worse than just holding gold. A 5% pullback in gold is common, and buying it was a worse decision than doing nothing. The mirror control makes the same point: after a >=3% up day, the 20-fixing excess was 0.86 percentage points, statistically indistinguishable from zero, so the up-moves were no better than the down-moves as a starting point.

Is the dip effect stable across decades?

No, and this is the reason to distrust it. For 20-fixing excess returns after a 10% dip, the excess was positive in 3 decades and negative in 3: 1970s 0.88 pp, 1980s -1.1 pp, 1990s -0.35 pp, 2000s 2.02 pp, 2010s 0.6 pp, 2020s -0.26 pp. Only the 2000s shows a large positive excess, and the 2020s are slightly negative. An effect that appears in one decade and reverses in another is not a structural property of gold; it is a sample artefact.

What happens if you actually trade it?

It buys the drawdown but badly lags the metal. The rule that buys a 10% dip and holds 20 fixings produced 151 trades over the sample, a 55.6% win rate, a final multiple of 1.86x and a CAGR of 1.07%. Buy and hold over the same period returned 108.6x, a CAGR of 8.33%. The dip rule wins on drawdown — -42.8% at its shallowest against -70.3% for buy and hold — and loses by roughly 5.9 percentage points of annual return every year.

So is "buy the dip" worthless?

It is not worthless as risk management, and it is close to worthless as a return source. The shallowest-drawdown rule cut the worst peak-to-trough decline from -70.3% to -42.8% while still participating in the recovery, which is a real benefit for an account that would otherwise have been forced out. But that benefit was bought with roughly 5.9 percentage points of annual return, and no dip rule tested here beat simply owning gold. The honest summary is: it changes the shape of the return, not its source.

What are the limits of this study?

Four, and they matter. First, the LBMA PM fix is a benchmark auction price, not a tradable bid/ask, so a rule that buys at the fixing could not actually transact at that price without paying the spread. Second, the 0.3 dollar per ounce round-trip cost is a flat assumption, while real spreads widen in exactly the conditions that generate the largest dips, so the true cost of catching a dip is understated here. Third, the signal is set on the fixing and executed at the next fixing, which flatters a rule that in practice would be triggered intraday. Fourth, the event studies use overlapping forward windows, so every t-statistic here uses a Newey-West HAC estimator rather than an ordinary t-test. And none of this speaks to leverage: a rule that sits out for weeks is a rule that behaves very differently when a broker's margin requirement is 5% of notional.

Sources

  • Gold: LBMA Gold Price PM benchmark fix, USD per troy ounce — scripts/daily_gold_lbma_pm.csv (1968-04-02 to 2026-11-10).
  • Analysis script: scripts/gold_dip_buying_analysis.py. Independent recomputation with numpy: scripts/verify_gold_dip_buying.py (267 checks over the published figures).
  • For the drawdown side of the same question, see gold tail risk and stop-loss evidence; for the structural-risk framing, see risk parity on leverage; for the calendar side, see gold seasonality.
Educational content only. Not investment advice. Trading forex and CFDs carries substantial risk of loss.