Does the Bearish Engulfing Pattern Work? A Weekly Test

What was tested

The 12-week test was specified on 30 September 2026, before any return was read, in the study's internal record (not an external registry): the event, the horizon, the comparison, the four market states, the traditional setup as a secondary analysis, the bootstrap and the rule for reading the result. The design follows Study 003. On 1 October a chart of median price paths on the first 100 stocks was viewed, after the specification; the record logs it, and no part of the test was changed because of it.

Exploratory question, added after the first results were seen: how did the price sit against its 26-week average before and after the pattern, compared with ordinary weeks? Its intervals use the same bootstrap, but because the measure was chosen after looking at the data they are not a preregistered test.

How the bearish engulfing was detected

The pattern is detected with the TA-Lib engulfing function, the same detector the app uses for its candle patterns. The candlestick pattern guide gives the full definition with diagrams.

A week whose close equals its open counts as an up week. Patterns two or more weeks apart are separate events, so their 12-week windows can overlap; the calendar-block bootstrap below accounts for that dependence. A weekly candle covers one calendar week: the open of its first trading session, the highest high, the lowest low and the close of its last session. It is dated on the week's last trading day and is complete at that day's close, the first moment the pattern is known; an unfinished week is never used. Prices are adjusted for splits before the weeks are built, and returns are price returns, without dividends.

The stocks and the data

The sample uses today's S&P 500 companies (one share class per company: 500 companies) rather than the companies that belonged to the index at each historical date. This creates survivorship bias and limits how broadly the result can be generalised; its effect on a comparison between pattern weeks and ordinary weeks in the same stocks is not known in advance. 480 companies were kept. The 20 dropped all had too little history after the warm-up period (at least 104 weekly candles were required): ECHO, EG, EXE, FDXF, FISV, GEHC, GEN, GEV, HONA, KVUE, MRSH, P, PSKY, Q, RDDT, SNDK, SOLV, SW, TKO and VLTO.

The 480 companies do not all have the same historical coverage. 289 have usable weeks from 26 January 2007 and 191 from 24 August 2018; all end on 25 September 2026, the last complete week. Because all 480 contribute from 2018 onwards, the pooled sample gives more weight to the later period.

Before any pattern was detected, each stock's prices were screened. A run of ten identical closes would mark a stalled price feed and drop the stock. A gap of more than 30 days inside a stock's prices starts its history after the gap, because indicators and windows across a hole mix two unrelated stretches of price. A merger special dividend, which a price-only series reads as a collapse, starts KDP's history after the merger. For some companies the price history begins at a rename, so it is shorter than the company's age.

How the comparison was built

Every week is labelled the same way, pattern weeks and ordinary weeks alike:

• Trend. The close above or below its 50-week simple moving average. The app does not compute 200-period averages on weekly candles, so the 50-week average sets the trend label.
• Volatility. NATR, the average true range as a share of price, below ("calmer") or above ("more volatile") that stock's median over 2007–2026.

The two labels give four states. Each pattern week is compared only with ordinary weeks in its own state. "Ordinary weeks" are every eligible week in the same stocks and state, meaning every week with 12 weeks of data after it; the pattern weeks themselves are included, and are about 5% of them (18,209 of 343,643). Including them pulls the comparison slightly towards the pattern weeks, so it can only shrink a difference.

The interval comes from a calendar-block bootstrap (2,000 replications, blocks of 84 calendar days, pattern and ordinary weeks drawn together). It was repeated at 84, 120, 168 and 252 days with three block designs, 12 configurations per state. A state shows a difference when its interval excludes zero in the main run and in every configuration; it is sensitive when it excludes zero in some configurations only. Intervals are not adjusted for testing several states.

The conditioning is coarse. Each label is a two-way category, and no pattern week is paired with individually similar weeks.

The 12-week result after weekly bearish engulfing patterns

Prices were higher 12 weeks later after 60% of bearish engulfing weeks, compared with 60% of ordinary weeks.

By market state

Differences ranged from -0.24pp to +0.46pp across the four states, below zero in three and above zero in one. The interval included zero in all four states, in all 12 bootstrap configurations.

In the traditional setup, above the 50-week average after a 12-week rise (an additional subgroup fixed in advance), the difference was +0.01pp (95% interval -0.46pp to +0.41pp); the interval included zero.

Exploratory analysis: the price against its 26-week average

The distance from the 26-week average is the weekly close divided by its 26-week exponential moving average, minus one. For each state, the charts show the median of that distance at every week from 12 before the pattern to 12 after, for pattern weeks and for ordinary weeks.

• Before. Over the 11 weeks up to the week before candle 1, the price sat higher against its average than ordinary weeks in the same state: by 0.6 to 1.7 points on average, across the four states. The 95% interval excluded zero in all four states. In the week of candle 1, an up week by definition, the gap grew to 1.2 to 3.2 points.
• The pattern week. During the pattern week the price moved below the ordinary-week level, to 0.7 to 2.3 points lower (the interval excluded zero in all four states). Part of the week-0 movement is built into the pattern definition: the second candle's body must cover the first candle's body.
• After. Over the next 12 weeks the difference from ordinary weeks narrowed. Twelve weeks later the interval included zero in one of the four states; above the average and calm sat 0.8 points below (interval −1.2 to −0.4), above the average and volatile sat 0.9 points below (interval −1.2 to −0.4) and below the average and volatile sat 0.7 points below (interval −1.3 to −0.1).

Part of the narrowing is expected mechanically. A 26-week exponential average closes about 60% of any gap with the price within 12 weeks, even if the price does nothing unusual. Measured against that catch-up, the gap closed further in one of four states, about as far in two and less far in one. After the pattern the price mostly moved like an ordinary week, and the average came to meet it.

The 50-week average gives the same picture: 12 weeks after the pattern the median distance was within 1.4 points of ordinary weeks in every state.

Interpretation

Some of the week-0 movement is part of the pattern definition. What the measure adds is the stretch before the pattern: the price sat above its ordinary-week level, with an interval that excluded zero in all four states. The result describes a change in the measured gap; it does not show that the pattern caused the price to return to its average.

Conclusion

Result

The preregistered test found a median 12-week change of +3.04% after bearish engulfing weeks, against +3.01% after ordinary weeks: a difference of +0.02pp (95% interval -0.44pp to +0.53pp); the gap was negative in three states and positive in one, and every interval included zero.

Interpretation

In this sample, bearish engulfing weeks made no clear difference to the 12-week return. The 26-week-average analysis, added after the main result, showed the price above its ordinary-week level before the pattern.

What the study does not show

This result does not establish that bearish engulfing patterns carry no information in other samples, timeframes or definitions, and it does not test a trading strategy. The volatility label uses hindsight, the sample is today's S&P 500 companies with survivorship bias and unequal history, and returns carry no market or factor adjustment.

Limitations and next tests

• Exploratory analysis. The 26-week distance was added once the first figures had been seen. Its intervals use the same bootstrap as the 12-week test, but they are not a preregistered test.
• Hindsight volatility label. NATR is compared with each stock's median over 2007–2026. A trailing median built only from earlier weeks would make it known at the time.
• Survivorship bias. The sample is today's S&P 500 companies, not the index members at each date.
• Unequal historical coverage. 191 of the 480 companies start in 2018, so the pooled results lean towards recent years.
• Coarse conditioning. Each label is a two-way category, and no pattern week is paired with individually similar weeks.
• No market or factor adjustment. Returns are compared within the same stocks and states only.
• Price returns. Dividends are excluded from both pattern and ordinary weeks.
• The prior-move label contains the pattern. The traditional setup's 12-week move runs up to the pattern week and includes both candles. Each stock's first 12 weeks, which have no prior move, fall in the "fell" group; leaving them out moved no difference by more than 0.14pp.
• An event study. No entry, exit, stop, position size or trading cost is simulated.

Next tests: fix the 26-week measure in advance and repeat it on a different period or universe; test whether the gap before the pattern is visible before candle 1 often enough to be measured on its own; and classify volatility from earlier weeks only.

How to reproduce this

The app can reproduce the study design, but not the full pooled result in one run: the published study uses 480 stocks, and the Events panel pools up to ten stocks per run.

In the Events panel, choose weekly bars (they need a Premium plan and a Tiingo key), build an event on the Bearish engulfing candle, set the horizon to 12, and add one Trend and one Volatility precondition per run. On weekly bars the Trend precondition uses the 50-week average. Each run gives one state of the 12-week result, and four runs give all four. The distance from the 26-week average was computed in the study harness.

With these ten stocks, MMM, AOS, ABT, ACN, ADBE, AMD, AES, AFL, A and APD, and a Custom range from 4 January 2010 to 25 September 2026, the app returns the figures below.

A ten-stock rebuild runs on a smaller sample, so its figures differ from the study's.