Strategy autopsy
The golden cross: we tested it on 337 stocks
“When the 50-day moving average crosses above the 200-day, a new uptrend has begun. Buy.”
- 661golden crosses walked forward
- 23%of stocks where timing beat owning
- 37%of stocks where it beat random dates
- 130trading days, median hold
If a trading signal has ever made the evening news, it is this one. The 50-day moving average crosses above the 200-day, and the financial press announces that a new bull market has technically begun.
We tested it on 661 real golden crosses across 337 stocks, from August 2023 to July 2026.
It lost to buying and holding. Then, when we gave it every chance to defend itself, it lost to picking entry dates at random.
The claim
When the 50-day moving average of a stock crosses above its 200-day moving average, a new uptrend has begun. Buy, and stay in until the averages cross back the other way — the “death cross”.
We test the rule, never a person.
How we tested it
| Universe | 337 liquid US stocks with enough history for a 200-day average |
| Period | August 2023 to July 2026 |
| Signal | SMA50 crosses above SMA200 (yesterday at or below, today above) |
| Entry | The next session, at the close of the first minute — a finished bar, no peeking |
| Exit | The death cross, the next session’s first minute (capped at 250 trading days) |
| Costs | A round trip is charged on every crossing. Buy-and-hold pays it once |
The comparison is the one that matters for how this claim is actually used. Nobody screens 337 stocks for golden crosses; people apply it to something they already follow. So for every stock, we asked: would you have done better following the rule, or simply owning the stock over the same window?
It lost to doing nothing
| Golden cross rule | Buy and hold | |
|---|---|---|
| Mean return | +38.1% | +88.1% |
| Median return | +10.0% | +37.0% |
| Time invested | 40.0% | 100% |
Timing beat owning on 23% of stocks. On the other 77%, following the most famous signal in technical analysis left you with less money than doing nothing at all.
Look closely at the last row, because it contains the whole story. The rule was invested 40% of the time and kept 43% of the return. Those two numbers are almost the same. The rule did not select good periods and avoid bad ones — it collected, near enough exactly, whatever a 40% exposure to the market was worth.
The objection this deserves
There is an obvious defence, and it is a fair one:
“2023 to 2026 was a bull market. Any rule that sits in cash most of the time loses to being fully invested. That says nothing about the signal.”
Correct. Comparing a 40%-invested rule against a 100%-invested benchmark in a rising market is not a fair fight, and if that were our only evidence we would have proved nothing.
So we ran the fair fight. For every stock we generated trades with the same number of trades and the same holding lengths the rule produced — but starting on random dates (20 seeded draws each). Same stocks, same period, same market, same time in the market. The only difference is whether the golden cross chose the entry days.
| Mean return | Median | |
|---|---|---|
| Golden cross | +38.1% | +10.0% |
| Random dates, identical exposure | +42.4% | +17.5% |
| Buy and hold | +88.1% | +37.0% |
The rule lost to the coin flip. It beat random entry dates on only 37% of stocks.
That is the finding. Once exposure is held constant, the bull-market objection disappears — and so does the signal. The golden cross is not a signal that tells you when to be in the market. It is a dial that leaves you 40% invested, and it sets that dial slightly worse than chance.
But it protects you in the crashes, doesn’t it?
This is the other standard defence, and unlike the first one, it is partly true. We measured the worst peak-to-trough fall of each path:
| Worst drawdown | |
|---|---|
| Golden cross rule | −27.5% |
| Buy and hold | −39.8% |
The rule drew down less on 84% of stocks — consistent, not a fluke. It really does hurt less.
But look at the price of that comfort:
- Drawdown avoided: 12.2 percentage points
- Return given up to avoid it: 50.0 percentage points
You paid four points of return for one point of comfort. And the comfort does not come from insight — it comes from absence. You cannot fall as far when you are not there. The random-date entries had the same exposure, and they earned more. Being out of the market protects you whether or not the thing that took you out was clever.
What about running it as a screener?
For completeness we also tested the version where you scan all 337 stocks and take every golden cross. It fails for a different reason: the rule holds for a median of 130 trading days, so a three-position account could take only 19 of the 661 signals — 3%. Net of full costs it lost money. But this is a footnote; the screener is not how anyone actually uses this signal.
Does more money save it?
For a rule you run on one instrument, account size changes only one thing: the costs. So we re-priced every stock’s result at three position sizes — roughly $6,000, $16,000 and $40,000 per position. A bigger order pays for its own footprint in the market, which we charged with a standard square-root impact model (a model, not a measurement — the only unmeasured number in this table); everything else is the same measured cost stack as above.
| Position size | Cost per round trip | The rule | Random entries | Buy and hold | Stocks where timing won |
|---|---|---|---|---|---|
| ~$6,000 | 25.2 bp | +38.0% | +42.3% | +88.1% | 23.1% |
| ~$16,000 | 26.3 bp | +38.0% | +42.3% | +88.1% | 23.1% |
| ~$40,000 | 28.1 bp | +38.0% | +42.3% | +88.0% | 23.1% |
The verdict is size-immune, and the table shows why it has to be. Costs move by three basis points across the whole ladder; the gap to buy-and-hold is fifty percentage points. The golden cross does not fail because trading it is expensive. It fails because the signal carries no timing information — it lost to random entry dates at every size — and no amount of money fixes that.
The verdict
DIED.
Cause of death: it is an exposure dial, not a signal. It earns its time in market and no more, and when you strip away the exposure difference by comparing against random dates with identical time invested, it comes out slightly behind chance.
That is a different way to die than our first autopsy, where the signal was real but arrived in bursts too crowded for an account to hold. Here the signal is not crowded at all. It simply is not there.
What we can and cannot say
Being straight about the limits of our own test:
- We tested individual stocks, not indices. The golden cross is most often quoted on the S&P 500. Index behaviour is not stock behaviour, and we do not have index data in this study. We cannot tell you the golden cross fails on the S&P 500. We can tell you it failed on 337 large US stocks.
- We tested one market regime, a strong rising market of about three years. A trend rule gets its worst hearing in exactly this kind of market. The random-exposure control is what makes the result survive that objection — but a longer history including a sustained bear market would test it harder, and we would like to run that when we have the data.
- The drawdown reduction is real. If your genuine goal is a smoother ride rather than more money, this rule delivers one. Just know you are buying it at four points of return for each point of comfort, and that a smaller permanent position would have bought it more cheaply.
What to take from this
- Compare against the boring alternative, then against random. Beating buy-and-hold is not enough on its own, and losing to it is not proof of failure either. The question a signal must answer is whether it beats chance at the same exposure.
- “Time invested” is a number you should always demand. If a rule earns roughly what its exposure is worth, it is not timing anything, however sophisticated it looks.
- Reducing drawdown is easy. Hold less. The hard part is reducing drawdown without paying for it four times over.
We will keep testing the rules people repeat. Some of them will survive.