Shooting Star and Inverted Hammer: One Bar, Two Names, Two Stories

Two names for a bar that only has one shape

Most candle patterns are at least arguably distinct objects. An engulfing bar and a doji look different from each other, and whatever you think of what they mean, you can tell them apart with your eyes shut. The upper-shadow family is not like that. A shooting star and an inverted hammer are pixel for pixel the same candle, and every published description of them agrees on this. The difference is not in the drawing, it is in the caption.

The lower-shadow half of this family is covered in full in our guide to the hammer candlestick, which owns the hammer and its identical twin the hanging man, along with the anatomy conventions, the confirmation discipline, the volume reading and the position-sizing arithmetic that a long shadow forces. This page is the mirror image and it is deliberately a different kind of page: not a description of what the upper-shadow bar records, but a measurement of whether the two names it can receive describe anything different. The wider taxonomy that both pages sit inside is set out in the guide to candlestick charts.

Here is why this particular pair is worth the trouble of measuring. In almost every other pattern argument, shape and context are tangled together. If a breakout works better in a trend, you cannot easily tell whether the trend improved the breakout or the trend was doing all the work by itself. The upper-shadow family gives you a rare clean experiment, because the shape is held exactly constant by construction. Two piles of bars, identical in every measurable respect, sorted only by what preceded them. If the two piles behave the same, the naming convention is describing nothing at all. If they behave differently, the difference has to be the context, because there is nothing else left.

That is a real question with a real answer, and it is answerable without any argument about whether candles work. It also has a second, less comfortable form. Suppose the two piles do behave the same. There is still the question of whether the bar itself does anything, pooled across every context, compared with an ordinary bar picked at random nearby. Most published discussion skips straight past that question, because it is busy distinguishing the two names. Both questions are answered below, and one of them produced a small result that complicates the tidy story.

Writing down the one definition

There is only one bar, so there is only one definition to code. That is worth saying plainly, because the literature presents two patterns and the instinct is to write two detectors. Two detectors would be two copies of the same five clauses with a different label at the end, and writing them separately is exactly the sort of thing that hides the point.

The verbal definition is short: a small real body at the lower end of the range, an upper shadow at least twice the height of that body, and little or no lower shadow. Every clause in that sentence stops the code from compiling until a number is supplied, and none of the numbers are in the sentence. How small is small? Small relative to what? What counts as the lower end of the range, and measured how, from the body's top or its bottom? How little is little? And what stops a bar so tiny that its shadow is noise from qualifying on a technicality?

Five thresholds answer those questions, and all five are choices.

Two things emerged from writing that table that were not visible from the prose version. The first is that the body-position clause is doing far more work than the famous one. On its own, the requirement that the top of the body sit inside the lower forty percent of the range admits 7.8% of all bars, while the twice-the-body shadow rule admits 19.2%. The clause everybody quotes is the loose one.

The second is worse, and it only shows up when you sweep. Setting the shadow multiple to 1.0 and setting it to 1.5 return exactly the same population, 25,842 detections both times. The reason is arithmetic rather than coincidence: if the top of the body sits inside the lower forty percent of the range, then the upper shadow is at least sixty percent of the range and the body is at most forty percent, so the shadow is already at least one and a half bodies. The most-quoted clause in the definition is inert below 1.5, because a clause nobody states has already enforced it. That is the sort of hidden coupling you only find by being forced to write the thing down.

Applied to 1,500,000 daily bars, which is 500 independent generated instruments of 3,000 bars each and about 6,000 instrument-years, the definition produced 28,764 qualifying bars. Discarding overlaps, so that no two measured outcomes share a holding period on the same instrument, leaves 23,436 independent detections. That is 3.9 per instrument per year, or roughly one every three months on any given chart. The typical detection had an upper shadow worth 3.5 bodies, a body top sitting 0.29 of the way up the range, and a lower shadow worth 0.12 of the upper one. The bars in the far tail were far more emphatic than the convention requires: a tenth of them had shadows longer than 9.4 bodies.

Two thirds of them do not get a name at all

Now the classification. The convention says a shooting star appears after an advance and an inverted hammer after a decline, which is a sentence with no numbers in it. Coding it requires deciding how much of a move counts, over how many bars, and measured against what.

The test used here reads only the closes, and it stops on the bar before the candidate, so that the candidate's own high and close can never contribute to the classification that names it. Take the change in the close over the previous ten bars, divide it by the recent daily volatility scaled to a ten-bar horizon, and you get a standardised prior move. A reading at or above 1.00 is an advance and the bar becomes a shooting star; a reading at or below the negative of that is a decline and the identical bar becomes an inverted hammer. Anything in between is neither.

The immediate consequence is one the convention never mentions. Of 23,436 detections, 4,252 followed an advance and 3,921 followed a decline. The remaining 15,263, almost two thirds of the population, followed neither. They came after a market that was going sideways, or drifting mildly, or doing something that a ten-bar test cannot call in either direction. No name in the convention applies to them, and yet they are the same bar and they are the bulk of the population.

In practice nobody leaves them unnamed. A chart with an upper-shadow bar in the middle of a range gets called a shooting star if the reader is bearish and an inverted hammer if the reader is bullish, and neither reading has any basis in a rule that was written down in advance. That is not a criticism of anybody's eyesight. It is the predictable result of a classification scheme whose only input is a trend test that was never specified.

The threshold has a second consequence, which is that it is unstable. Hold the 23,436 detections completely fixed, change nothing about any bar, and simply move the trend threshold across the three settings either side of the one used here. Only 5,780 of the 8,173 bars that received a name at the middle setting kept that name across all three. The rest either swapped from one name to the other or fell out of the naming scheme entirely. The market did nothing. A number that no textbook prints moved slightly.

Building a comparison that is not circular

A pattern statistic on its own means nothing. If a shooting star is followed by a fall sixty percent of the time, the question is what fraction of ordinary moments on the same data are followed by a fall, because the answer might also be sixty percent. That much is standard. What is not standard, and what turns out to decide the answer on a page like this one, is where the comparison entries are drawn from.

The natural choice is to draw random entries from the bars surrounding each detection, so the market conditions are comparable. For a test conditioned on prior trend, that choice is circular, and the circularity is not subtle once it is pointed at. The bars immediately before a shooting star are the advance that made it a shooting star. Comparing the bar against random entries taken from inside that advance is comparing it against the thing being used to classify it.

So every control entry on this page is drawn from bars strictly after the detection, between twenty-two and two hundred and fifty bars later, far enough ahead that neither the control's own trend test nor its holding period can reach back and touch the event or the ten bars being measured. Each control is also matched on the same prior context, takes a position in the same direction, holds for the same ten bars, and is required not to satisfy the shape definition itself. Two hundred controls are drawn for each detection, and the comparison is paired, so each detection is measured against its own controls rather than against a single pooled average.

That repair is necessary and it is not sufficient. A forward-only control is still not neutral, because a control drawn from a different stretch of tape is not matched on whatever slow-moving condition the detection happens to be sitting inside. The way to find out how big that problem is, rather than to argue about it, is to run the identical machinery on bars that are not the pattern. Take the same number of bars in each context class, require them to fail the shape definition, and push them through the same control apparatus. Those placebo bars carry nothing by construction. Anything the machinery reports for them is the machinery.

It reports a great deal. Under the forward-only control the shooting stars came out 0.60 percentage points ahead of their controls, and the placebo bars came out 0.32 points ahead of theirs. Most of the apparent excess was never about the bar. Every headline number on this page is therefore the difference of two differences: the pattern's excess over its control, minus the placebo's excess over its own.

The progression in that figure is the reason the repair was worth making. Draw the controls only from bars before the detection and the shooting star's raw excess is 0.53 points. Draw them from two hundred and fifty bars either side and it falls to 0.21. Widen to a thousand bars either side and it is 0.22. Draw them only from after and it is 0.60. Four defensible-sounding choices, a range of nearly three to one, and no way to tell from the number itself which one you are reading. Once the placebo arm is netted out, the gap between the two names sits between 0.11 points below zero and 0.12 above it under all four, and never reaches one standard error.

The central experiment: do the two names differ?

With the comparison built, the question can finally be put properly. Take the 4,252 bars that qualified after an advance and the 3,921 that qualified after a decline. Both groups are measured the same way, as a long position taken at the next bar's open and held ten bars, so the two sit on one scale and can be subtracted. The bearish reading of a shooting star is the exact negative of its long figure, so reporting it separately would be reporting the same number twice with the sign changed, and the page says so rather than doing it.

The shooting stars finished 0.27 percentage points ahead of their matched placebo comparison over ten bars. The inverted hammers finished 0.15 ahead of theirs. The gap between them is 0.12 percentage points, with a standard error of 0.16 and a ninety-five percent interval running from 0.20 below zero to 0.44 above. That is not a small effect that failed to reach significance. It is an interval tight enough to rule out any difference larger than about four tenths of a percentage point in either direction, on more than eight thousand named instances.

Read in the direction the names imply, which means selling after the shooting star and buying after the inverted hammer, the combined exercise returned −0.07 percentage points, with an interval from 0.23 below to 0.09 above. It is on the wrong side of zero and it is indistinguishable from zero, which are two different statements and both of them are true.

One detail in that pair of numbers should be flagged rather than buried, because it is the opposite of what the reader is expecting. Both arms lean the same way, upward, and the shooting-star arm is far enough above zero on its own to clear two standard errors. A bar that is supposed to be the bearish half of the family is the one with the mildly positive drift after it. What that is, and why it should not be treated as a finding, is the subject of the next section.

A null result is worth nothing unless the instrument can be shown to detect something. So the identical detector, the identical control machinery, the identical placebo arm and the identical thresholds were run over a second tape, generated the same way but with one real effect deliberately inserted. The effect is stated purely as two distances, how far the session high stood above the open and above the close in average true ranges, with the sign taken from the twenty-bar change in the close. It never mentions a real body, a lower shadow, or any ratio between the parts of a bar, which are the only things the detector measures. It fires on roughly one bar in six across the whole tape, far broader than the definition, so the detector has to find it by intersection rather than by being handed it.

It found it. On that tape the two names separated by 2.42 percentage points, more than thirteen standard errors, and the folklore reading returned 1.19 points at a similar margin. The instrument works. On the tape where there was nothing to find, it found nothing.

The row worth pausing on is the third one. The bars that received no name at all behaved essentially the same as the two named groups, 0.09 points against 0.27 and 0.15. Whatever the population of upper-shadow bars is doing on this tape, it is doing it uniformly, and the trend that assigns the name is not sorting it into anything.

Does the bar carry anything at all?

The second question is the one the naming argument distracts from. Forget both labels. Pool every detection, named or not, and ask whether an upper-shadow bar is followed by anything different from what follows an ordinary bar.

It is, slightly, and the honest thing is to publish it rather than to round it to zero. Pooled across all 23,436 detections the bar was followed by 0.13 percentage points more than the matched placebo comparison over ten bars, with an interval from 0.04 to 0.22 and a t statistic of 2.94. On the strength of the number alone that is a result.

Four things about it are worth stating before anybody does anything with it. It is 0.13 of a percentage point over ten bars, which is roughly the cost of the round trip that would be needed to collect it. It points upward, which is the wrong direction for the shooting star's bearish reading and the right direction only for the half of the population called an inverted hammer. It is present in all three context classes including the one with no name, so it is not a property of either label. And this page ran several dozen comparisons, so a single t statistic near three, in a set that size, is not a finding anyone should defend.

The obvious explanation for a small positive residual is arithmetic rather than behavioural. The definition selects bars with long shadows, long shadows come with high volatility, volatility is persistent, and the expected simple return over a fixed window rises with variance even when the underlying drift is exactly zero. That is a real effect and it needs no market in it. It was checked rather than assumed. Measured in log returns, where the variance term disappears, the pooled figure is 0.12 points against 0.13 in simple returns, and the realised volatility of the ten bars following a detection is only 1.05 times that of a placebo bar, which accounts for about 0.010 of a point. The convexity explanation is not the explanation. Something small and real is there in the generated data, it is smaller than the friction required to reach it, and this page cannot tell you what it is.

The classical repair is confirmation, and the standard is a close beyond the bar rather than a touch, which for the bearish reading means the next bar closing below the shooting star's low. That was tested too, with entry moved to the open after the confirming bar. Only 1,580 of 4,252 shooting stars were confirmed on that definition, about 37%, and the confirmed subset finished 0.18 points from its placebo comparison with a t statistic of 1.02. For inverted hammers the bullish equivalent, a close above the bar's high, occurred just 465 times out of 3,921 and produced 0.03 standard errors of nothing. Waiting for the close remains good discipline for the reason the parent page gives, which is that it fixes an entry and therefore fixes what is being risked. It did not turn the bar into information here.

How much of this is the definition rather than the data

Every number above sits on five shape thresholds and one trend threshold. A finding that survives only at one setting is a finding about that setting, so the honest report is what happens when they move.

The shadow multiple is the clause the brief for this exercise singled out, and it behaves in two distinct ways. Below 1.5 it does nothing whatsoever, for the arithmetic reason given earlier. Above 2.0 it starts cutting hard: 23,436 detections at 2.0, 15,611 at 3.0, 11,065 at 4.0. So the population is a tolerance setting, exactly as the triangle exercise found for its own classification, and anybody quoting a frequency for this bar without publishing their multiple is quoting their own settings.

What happens to the result as the multiple tightens is the uncomfortable part. At 1.0, 1.5, 2.0 and 2.5 the gap between the two names is indistinguishable from zero. At 3.0 it reaches 0.47 points at 2.39 standard errors, and at 4.0 it reaches 0.74 at 3.07. The same thing happens when the body-position clause is tightened: at 0.30 rather than 0.40 the gap is 0.63 points at 3.03 standard errors. It survives being measured in log returns, so it is not the variance artefact either, coming in at 0.74 points at the strictest shadow setting.

Before anyone reaches for that, note its direction. The gap is positive, which means the bars called shooting stars did better over the following ten bars than the bars called inverted hammers. That is the reverse of both names. And it appears on a tape built with no directional memory in it, where by construction no market mechanism exists to produce it. A statistic that shows up at the strict end of a threshold sweep, on data where nothing is happening, pointing the wrong way, in a set of several dozen comparisons, is a description of the definition and of multiple testing. It is not a discovery about markets. It is, however, precisely the kind of number that would be reported as one if it had turned up on real charts at the setting somebody happened to pick.

The trend threshold moves the classification even more freely. At the loosest setting tested, 64% of detections receive one of the two names; at the strictest, 17% do. The shooting star population runs from 7,735 down to 2,047 without a single bar changing shape. Across the middle three settings, 5,780 of 8,173 named bars keep their name and the rest do not. The other thresholds tell the same story in miniature: the lower-shadow cap moves the population between 10,536 and 32,372, the size floor between 6,672 and 33,077, and the holding period from five bars to twenty leaves the gap between the names within 0.35 points of zero throughout.

What survives all of it is the absence of a difference at any setting a reasonable person would have chosen in advance, and the presence of one only where the population has been cut to a third and the direction has flipped against both names. The population of upper-shadow bars is extremely sensitive to the definition. The finding that the two names do not sort that population is not.

What this supports and what it does not

A simulation can establish some things and cannot establish others, and being precise about which is which is the difference between a study and an opinion with numbers attached to it.

The largest limitation is the obvious one. Generated data has no earnings, no policy announcements, no index rebalancing and no order book, so it cannot tell you what happens when a real rally runs into real overhead supply. What it can do, and what real data cannot, is provide a case where the true answer is known in advance. When a statistic reproduces on a tape with nothing in it, that statistic is not evidence, and separating the numbers that fall into that category from the ones that do not is the whole contribution here.

A second limitation is that this tests one definition and one trend rule. A different set of six numbers is a different detector, and the sensitivity study above is precisely the reason to expect a different detector to find a different population. Anybody who wants to argue that a better definition would produce a different answer is making a testable claim, and the way to settle it is to write that definition down as code and run it against a control that was not drawn from the move being tested.

What to do with an upper shadow instead

None of this makes a long upper shadow useless to look at. It makes the two names useless, which is a much smaller loss than it sounds, and it relocates the part that was worth having.

Read the fact, not the label. The bar records one thing that is not in dispute: price traded well above where the session settled. Somebody was paying up higher and stopped. That is a fact about a single period of trading, it is checkable against the four prices, and it needs no name attached to it. The moment you decide whether to call it a star or a hammer, you have added an unstated trend threshold to a fact that did not need one.

Treat the trend test as the claim it is. If the reason a bar looks bearish to you is that it followed a rally, then the rally is your evidence and the bar is decoration. Say it that way and the position becomes arguable, because someone can ask how much of a rally, over how many bars, and whether that has ever been measured. Say it the other way and the trend hides inside a candle name where nobody has to defend it.

Write your definition before you go looking. This is the habit that transfers to everything else on a chart. Being forced to say how long a shadow must be, and how far up the range a body may sit, and how much prior move earns a name, is how you discover how much of your pattern recognition was a decision rather than an observation. Doing it once for this bar is worth more than reading ten descriptions of it, and the same exercise on a breakout or a support level produces the same uncomfortable clarity.

Know where your comparison group came from. The single largest number measured on this page was not produced by the market. It was produced by moving the control window, and it moved a headline figure by a factor of nearly three. Any pattern statistic you are shown, anywhere, was measured against something. Asking what that something was, and whether it was drawn from the same stretch of price action that defined the pattern, will dispose of a great deal of published material very quickly.

Keep the doji distinction, and keep it for the right reason. Push the body of this bar down to nothing and it becomes a gravestone doji, which our guide to the doji candlestick covers as part of a family with its own limiting cases. The reason to keep the two apart is not that they mean different things. It is that a definition which quietly absorbs its neighbours cannot be tested, because you can no longer say what it excluded.

What is left after all of that is smaller than the usual telling and considerably more durable. A single bar is a fact about one session. Turning facts into decisions requires a rule written down in advance, a comparison group that was not drawn from the evidence, and a willingness to publish the runs that went nowhere. If that way of working is more interesting to you than the taxonomy it replaces, it is the method we teach.

Every figure comes from one deterministic simulation, seeded so that it reproduces identically on each run. Two tapes were generated, each of 500 independent instruments of 3,000 daily bars, 1,500,000 bars in total. Volatility clusters and drift regimes switch on both, so quiet and violent stretches occur naturally and no bar shape is ever inserted by hand; the three drift states are symmetric and equally likely, so the unconditional drift is zero. Upper and lower shadows are drawn from the same distribution, so the generator has no preference for one end of a bar over the other. The first tape contains no mechanism that could make any bar shape predictive. The second adds one effect, stated as two distances measured in average true ranges, how far the session high stood above the open and above the close, with the sign taken from the twenty-bar change in the close. It refers to no real body, no lower shadow and no ratio between the parts of a bar, and it fires on about one bar in six, so the detector inherits it only by intersection. Its size is deliberately large, because a control that cannot be seen is not a control.

Detection runs on the open, high, low and close of the generated bars using the six thresholds in the definition table. The prior-context test reads closes only and stops on the bar before the candidate. Positions are opened at the next bar's open, so no result uses information that was not available at the time, and outcomes are measured over ten bars as simple returns, which is what a position actually makes. Overlapping detections are discarded so that no two measured outcomes share a holding period on one instrument. Control entries are drawn from the same instrument, from bars between twenty-two and two hundred and fifty bars after the detection, matched on the prior-context label, required to fail the shape definition, and entered and held the same way, at two hundred draws per detection with the comparison paired. A placebo arm of the same size, matched on context and required to fail the shape definition, is pushed through the identical machinery, and every headline figure is the pattern's excess minus the placebo's. Returns are shown before costs; an illustrative allowance for a full round trip would apply equally to both arms and would not change the difference between them, though it would sit above most of the differences measured.

All results are illustrative and simulated. They are not a track record, they are not a forecast, and they are not an indication of what any pattern would produce in a live account or on any Indian security or index. The purpose of the exercise is to establish which properties of this bar are consequences of its definition and of the way it is measured, which is a question about definitions and measurement rather than about any particular market.

• The wider family of chart patterns →
• A pattern that builds its own confirmation →
• The eight ways a backtest can lie →
• Triangles: one mechanism, three names →