Chart patterns and candlestick setups
Does MACD divergence work?
Rejected Pre-registered and tested, July 2026
The loose version changed the outcome by less than one hundredth of a percent. The strict version widened both tails.
What the pattern claims
MACD is a momentum indicator built from two moving averages. Bullish divergence is the case where price makes a lower low but MACD makes a higher low — the stock is falling, but falling with less force each time. The claim is that this reveals hidden strength before the price turns.
How we tested it
Tested two ways, because how tightly you define divergence turns out to matter. The loose definition used either the MACD line or its histogram with simple pivots — 7,773 occurrences. The strict definition required distinct, properly-formed troughs on both the price and the histogram — 3,999 occurrences. S&P 500 companies, buying at the next morning's open.
Every study here was pre-registered: the exact rule, the pass and fail thresholds and the data window were written down and cryptographically fingerprinted before the test was run. That makes it impossible to move the goalposts after seeing the answer.
Results assume you buy at the next morning's open, not at the closing price that triggered the signal, and they include companies that were later delisted or went bust. Returns are measured against what the rest of the market did on the same days, so a rule that made money only because it fired on days everything rose scores zero here.
What happened
On the loose definition, the divergence condition did nothing at all. The signal and a control group with no divergence requirement whatsoever matched each other to within one hundredth of a percentage point. The underlying trigger was itself a loser, and adding divergence to it changed neither the average nor the typical case.
The strict definition looks better at first glance: average returns turn positive, up 0.4% at 20 days and 0.9% at 60. But the median case fell below the no-divergence control and fewer than 48% of trades were profitable. Both eras we tested were negative after 2024.
What it means
That gap between the average and the median is the whole finding, and it generalises well beyond MACD.
When tightening a filter lifts the average while pushing the typical case down, the filter is not finding an edge. It is selecting for variance — picking trades with fatter tails in both directions. A handful of large winners drag the average up while most trades quietly get worse. If you traded it, you would experience the median, not the average.
This is the most common way a chart pattern appears to work. Whenever a stricter definition improves the mean and worsens the median against its own control, stop there.
A rejection is not "this never works"
It means: tested on this universe, over this period, against a bar written down in advance, it did not clear the bar. Where a test lacked the power to decide either way we record it as undecided rather than rejected. The full research ledger has all 192 experiments and what the failures have in common.
Other patterns we tested
- Does the bull flag pattern work? The pole is the only live part of the pattern. Every flag condition made the result worse.
- Does breakout and retest work? The setups the rule throws away outperformed the ones it keeps.
- Does buying the dip after a spike work? The first pattern in this family to lose money significantly rather than merely do nothing.
- Does a big RSI drop mark a buying opportunity? The family's only statistically significant result, and it vanished completely once we accounted for when it fired.
- Do three shrinking red candles signal a bottom? The only chart pattern that beat all its own control groups — and it still could not clear the error bars.
- Does a tight base after a sharp drop work? Removing the base conditions improved the result. The base was the problem.