Candlestick patterns against the only benchmark that matters: random dates

Marshall, Young & Rose tested the full menu of bullish and bearish candlestick signals on DJIA stocks with bootstrap methods and found they created no value — returns were statistically indistinguishable from chance. This rebuilds their method on synthetic candles you control: a genuine geometric detector, a real forward-return statistic, and a random-date bootstrap null. The detector is not rigged. Turn the drift-injection slider up and the same test lights up as significant — which is how you can tell the negative result is a measurement and not a rhetorical trick.

The pattern

Body / range threshold0.25

Two thresholds come off this one dial, so nothing is hard-coded: SMALL = body/range ≤ this value (doji, star bodies), LONG = body/range ≥ max(0.55, 1 − this value) (marubozu, and the long candles inside piercing, harami and morning star).

Shadow multiple2.0×

Hammer and shooting star need a shadow at least this many times the real body, opposite shadow no longer than 0.6 bodies. If you think a definition here is unfair, change it and re-run.

Prior-trend context

Reversal patterns are supposed to reverse something. “Require” counts a bullish signal only when the close 1 bar back is below the close 4 bars back; mirrored for bearish.

The test

Holding period H5 bars

Forward return is measured from the signal bar's close to the close H bars later.

Bootstrap resamples B500

Each resample draws the same number of entry dates uniformly at random from the same series, held the same H. That is the null: same trade count, same horizon, no pattern.

Injected post-pattern drift0.00% / bar

The honesty control. At 0 the candles carry no pattern edge at all. Above 0 the generator adds real drift to the H bars after each genuine signal — and the same unchanged test must find it.

Sample#1

Same generator, new random draw. Watch the observed statistic wander inside the null band.

Two-sided bootstrap p-value against random dates

Signals detected

Base rate

signals per 1,000 bars

Observed mean fwd return

Null mean

random dates, same N and H

95% null band

2.5th to 97.5th percentile

Percentile of observed

Distinct pattern / tolerance / horizon combinations you have tested this session1

Null distribution of the mean forward return · observed statistic marked

Random-date resamples Central 95% Observed Null mean

Synthetic candles · detected signals

Windowbars 0–120

Up bar Down bar Detected signal

Why this sim is not a strawman: the detector is real geometry on internally consistent OHLC bars — every bar comes from a simulated intraday path, so highs sit above both open and close, lows below both, and body-to-shadow proportions fall out of the process rather than being assumed. The test has real power, and you can prove it: change nothing else and push injected post-pattern drift above roughly 0.2% per bar. The same detector and the same bootstrap that just said “indistinguishable from chance” will report a significant p-value within a few clicks. A test that finds an edge when one is planted and nothing when none is, is measuring rather than arguing. Watch the counter above too: run enough pattern and tolerance combinations and one clears 5% on luck alone — the multiplicity problem bootstrap methods exist to control.

What the research actually found

Support level: none. Of every concept in the IndicatorEdge research corpus, this is the only one where the honest answer is that there is nothing to defend.

“Nothing robust. This is the cleanest negative result in the indicator literature.”

“Tested across the full menu of bullish and bearish candlestick signals on DJIA stocks (1992–2002) with bootstrap methods, candlestick strategies created no value for investors — returns were statistically indistinguishable from chance (Marshall, Young & Rose 2006).”

Where the patterns came from: “Attributed to 18th-century Japanese rice trading and popularized in the West in the 1990s. The first robust academic test came only in 2006.” Roughly two centuries of practice, one decade of Western marketing, and then a single bootstrap study.

This simulation demonstrates the method — genuine pattern detection plus a random-date bootstrap null — on randomly generated synthetic candles. It cannot and does not replicate the paper's DJIA result: no market data is used here, and a negative result on synthetic bars is not evidence about real ones. The evidence about real markets is the citation below. The sim exists so you can see for yourself what “indistinguishable from chance” looks like when you run the test by hand.

Marshall, Ben R., Young, Martin R., & Rose, Lawrence C. (2006). “Candlestick technical trading strategies: Can they create value for investors?” Journal of Banking & Finance, 30(8), 2303–2323.