MEASURED · 18 SEPTEMBER 2026

Does the golden cross work?

Gold roughly tripled over the last twelve years. That is the problem with every back-test of this signal: buy on almost anything and hold, and the average result is positive. So the only question worth asking is whether the golden cross beats simply being in the market. On 9 daily crosses since 2014, it did not.

The claim

When the 50-day moving average crosses above the 200-day, the market is said to have turned bullish. Crossing below is a death cross and is read as bearish. Both get reported as news events, which few other technical signals manage.

It is an appealing idea because it is objective — there is no judgement in whether two lines crossed. That also makes it unusually easy to test properly.

Why most tests of it are meaningless

Search for a golden cross back-test and you will find average returns after the signal. They are almost always positive. So is the average return after any randomly chosen day in a market that tripled.

Everything below is therefore reported as an edge: the return after the signal minus the return from a randomly chosen bar over the same horizon. An edge of zero means the signal told you nothing that being in the market did not already tell you.

Before this ran on gold, it was checked on data with a known answer. A fake signal placed in a strongly rising market returned +1.85% on average — a number that would headline well — and the method correctly rejected it, because the base rate over the same period was 1.99%. A test that cannot catch that would not be worth running.

The famous one: 50/200 daily

This is the signal the financial press reports. It has occurred 9 times on gold since 2014.

The golden cross edge is negative at all three horizons. Not significantly so — with nine events almost nothing could be — but the famous bullish signal did not merely fail to help, it slightly underperformed doing nothing.

Nine events across twelve years is also worth sitting with. Whatever you believe about this signal, the daily chart cannot supply enough of them to settle it. That is the same problem as gold seasonality, where twelve observations per month produced a convincing-looking pattern that dissolved under a proper test.

The result that would have made a better headline

The death cross produced the only two p-values below 0.05 in the entire study: one week and one month afterwards, gold underperformed the base rate by 1.38 and 2.89 percentage points. “Death cross works, golden cross does not” is a genuinely interesting headline and I am not going to write it, for three reasons.

• It is one result, not two. Both come from the same 10 events, and the one-week window sits inside the one-month window. Counting them separately double-counts the same evidence.
• Two hits from 36 tests is chance. At a 5% threshold, 1.8 false positives are expected across 36 tests. We got 2. Correcting properly for that many tests requires p below 0.0014; these were 0.032 and 0.026.
• It partly reports the past. Price had already fallen 0.8% on average in the three weeks before each death cross. A moving average crossover is a lagging construction; some of what looks like prediction is a decline already under way.

Every other setting

If nine events is too few, the fix is more events — faster timeframes and a quicker 20/50 pair. This is no longer the signal CNBC reports, but it is the same mechanism with enough data to test.

Read it top to bottom. The settings with the fewest events show the largest numbers in both directions; the settings with hundreds of events show edges of a few hundredths of a percent. On the hourly 20/50 pair — 636 golden crosses, the largest sample here — the one-week edge is +0.01 percentage points.

Effects that shrink as data accumulates are not weak signals. They are noise being measured more precisely.

And it changes its mind

One practical detail the averages hide. Between 2021-12-28 and 2022-01-21 — a span of 24 days — daily gold produced a golden cross, a death cross and another golden cross. Anyone trading the signal mechanically would have gone long, reversed, and reversed again inside a month, paying the spread each time. A signal with no edge is not free; it costs you every time it fires.

What this does and does not say

It says that on gold, a moving average crossover did not predict the direction of the next week, month or quarter better than being in the market already did. That is the specific claim, and across 36 tests it did not hold.

It does not say moving averages are useless. Describing trend, filtering out trades against it, or defining where a position stops being right are different jobs from prediction, and a tool can do those well while having no forecasting power.

We can say that with some confidence because we ran the experiment on ourselves. Our own system shipped with several moving-average entry strategies, and walk-forward testing across seven months of 2026 removed every one of them: EMA-PULLBACK lost $391 and was negative in five of the seven months, EMA-TREND returned -$61.14 at a 33% win rate over 15 trades, EMA-BOUNCE -$40.06 at 20% over 10. They are disabled in the shipped configuration. That is not a claim about your trading — it is a note that when we tested the same family of signals with money attached, we reached the same conclusion this page reaches with statistics, and acted on it by cutting the strategies rather than defending them.

Limits

One instrument, one broker, 2014-01-14 to 2026-09-18, in a period gold spent mostly rising. A signal that works only in falling markets would be hidden here, and the death cross result is a faint hint in that direction — though it does not survive correction. Simple moving averages were used; exponential ones cross at slightly different moments, but the whole point of the multi-setting table is that the result does not hinge on the exact definition.

Forward windows overlap, which inflates significance in a naive test. The permutation control draws random bars with the same overlap structure, so the comparison stays fair. The full numbers for all six settings are published with our other studies, including the base rates that make each figure checkable, and the same reasoning applies to Fibonacci retracements.

Common questions