PATTERN REALITY CHECK · INDICATORS · TESTED 8 OCT 2026

MACD divergence: does it work after costs?

Price makes a new extreme, MACD does not. Courses say the move is running out of steam.

No detectable edge after costs

Verdict

No detectable edge after costs

Applied mechanically to 29 coins, this rule showed no detectable edge after costs: the average is within the range of luck, and it does not beat entering at random in a way that holds up.

Average per trade-4.0%-0.040Rof the amount risked, after costs and funding. 95% range: -14.1% to +6.2%.

Trades tested1,40029 coins2020-06-11 to 2026-09-19, 4-hour candles.

Trades won42%needs 44% to break evenWins averaged +1.042R, losses -0.822R.

Against random entries14%of random runs beatenRandom entries averaged -0.9%, because costs hit every trade. Beating them is not enough: the average must also be clearly above zero.

Percentages are % of the amount risked per trade. New to R? Read this first.

What courses teach

The idea

The same idea as RSI divergence, with the MACD line as the oscillator.

The rules we froze

These numbers are the usual textbook values. They were written before we saw a single result and were not changed afterwards.

PartRule, fixed before the test

Signal

Two confirmed swing lows (or highs) 8 to 60 candles apart. Price makes a lower low (higher high) while the MACD line (as % of price) makes a higher low (lower high).

Entry

Market order at the close that breaks the high (low) between the two swings, within 30 candles.

Stop

0.25 ATR beyond the second swing. At least 1 ATR from entry.

Target

Twice the distance to the stop (2R).

Time limit

Closed after 60 candles (10 days) if neither stop nor target was hit.

Costs

0.14% round trip plus real funding, charged on every trade.

What it looks like

Examples, good and bad

Two real trades picked by a fixed rule, not for being pretty, and six drawn at random from the held-back period (the most recent two years). 1 of the 6 won. Examples show the rule, they are not results: the results are the numbers further down.

MACD divergence example that reached the target

MACD divergence example that was stopped out

Results

Trade after trade

All 1,400 trades added up, with the three periods shaded. The right axis shows euros if you risked €100 on each trade. The dashed line is what random entries would have done.

By period

The same average, split by period. The black bars are the 95% range: if they cross zero, we cannot tell the result from luck.

PeriodTradesAverage per tradeTrades won95% range

Build

649

-2.4% -0.024R

42.2%

-16.9% to +12.7%

Validation

539

-7.6% -0.076R

41.9%

-22.9% to +9.7%

Held-back

212

-0.1% -0.001R

41.0%

-24.9% to +24.3%

By version

VersionTradesAverage per tradeTrades won

Long

752

-6.4% -0.064R

38.2%

Short

648

-1.3% -0.013R

46.3%

How the trades ended

The honest yardstick

Compared with entering at random

A pattern should beat luck, not zero. We ran 1,000 simulations that enter at random times on the same coins and periods, in the same direction, with the same stop distance, target and time limit.

MACD divergence compared with random entries

In plain words

Why the numbers look like this

Random entries lose too. Entering at random times with the same stops and targets averaged -0.9% per trade (-0.009R) because of fees, funding and spread. The pattern averaged -4.0% and did better than 14% of 1,000 random runs. That is the honest comparison: not zero, but what luck plus costs produce.

Before costs: -2.3%. After costs: -4.0%. Fees, funding and slippage took 1.8% of the amount risked from every trade, on average. Even with zero costs the average is close to zero.

Win rate versus break-even. It won 41.9% of trades. With average wins of +1.042R and losses of -0.822R, it needed 44.1% to break even. The win rate is below what it needs.

Opposite side, same stops. Taking the opposite direction on every signal, with the same stop distance, averaged +4.1%. If the pattern carried real information, the pattern direction should beat its mirror by a clear margin. Here the gap is 8.1 percentage points.

Consistency across time. The average was positive in 0 of 3 periods. A real edge should not depend on which period you look at.

Sensitivity. Across 32 variations of the rule, 0 were positive and 32 negative. The best variation (-0.4%, 1,400 trades) is not a result: with 32 variations, a few green cells appear by chance. We look for a broad green region.

Stress tests

Does it survive changes?

Changing the pattern shape

Each cell is the same test with a different setting. The outlined cell is the rule we froze. Green is positive, red negative. We do not pick the best cell: we look for a broad region.

Changing the exits

Same idea for the target and the stop.

Changing the costs

CostsAverage per trade

No costs, funding only

-2.5% -0.025R

Half the costs

-3.3% -0.033R

Our costs (0.14% + funding)

-4.0% -0.040R

Double costs

-5.6% -0.056R

Triple costs

-7.1% -0.071R

Another timeframe

On daily candles the same rule gave +3.5% over 248 trades.

Coin by coin

11 of 29 coins had a positive average. With a few dozen trades per coin, some are always positive by chance.

Limits

What this does not say

It does not say macd divergence never works. It says this version of the rule, applied to every case, shows no detectable edge after costs in this data.
It is a simulation. Real fills, delays and emotions make results worse.
The market may change. Results from the past are not a forecast.
We publish tests of popular patterns, we do not publish or comment on strategies we use ourselves.

Method: how we test · Errors: corrections · Education only, not investment advice.

Questions

Questions about macd divergence

Does macd divergence work in crypto?

In our test of 1,400 trades on 29 crypto perpetuals, with real costs and funding, the average result was -0.040R per trade (-4.0% of the amount risked). Verdict: no detectable edge after costs. This applies to the textbook rule defined on this page, not to every way of using it.

Why do I see charts where macd divergence worked?

Because any rule produces winning examples. A chart that shows a winner says nothing about how often the same rule fails. The test above applies the rule to every case, with costs, without picking.

What would make this result change?

A different timeframe or market, different exits, or filters that add information the pattern does not contain. We tested a second timeframe and 32 variations of the rule; the numbers are on this page.

How should I read the percentages?

They are percentages of the amount you risk per trade, not of your account or of the coin price. -4.0% means -4.0% of the amount risked, on average. See what is R.