Reality Check: Do High-Volume Breakouts Lead Anywhere?

You now have the tools to stop taking claims on trust. This lesson uses them on one of the most widely repeated claims in technical analysis, and the method matters far more than the answer you get.

The claim

A breakout on unusually heavy volume is more significant than a breakout on ordinary volume. The volume confirms the move.

You will find this in books, courses and every trading forum. It is stated as though it were a property of markets rather than a hypothesis about them.

Notice what is wrong with it as written: every load-bearing word is undefined. Breakout above what, over what window? Unusually heavy compared to what? More significant in what measurable sense, over what horizon? As stated, the claim cannot be true or false, because it cannot be tested.

Our job is to replace it with something that can be.

Step 1: state the claim so it can be measured

From a slogan to a testable question

1. 1The slogan"Volume confirms a breakout." Not testable — no definitions, no horizon, no comparison.
2. 2Define every term objectivelyBreakout = close above the highest high of the prior 50 bars. High volume = at least twice the 50-bar average volume computed to yesterday.
3. 3Define the outcomePercentage change of the close over the following 20 bars. One number, computed the same way for every observation.
4. 4Define the comparisonThe same measurement over every eligible bar, condition ignored. Without this the first number is uninterpretable.
5. 5Decide in advance what would count as an answerWrite down, before running it, what difference would be interesting and what would not.

Here is the version we will actually test:

Among liquid shares priced above a floor, does the average 20-bar forward return following a close above the prior 50-bar high on volume at least twice the prior 50-bar average differ materially from the average 20-bar forward return following any eligible bar in the same universe and period?

Every term in that sentence is now a number in a formula. Somebody who disagrees with a choice can change one constant and re-run instead of arguing about words. That is the point.

∴Research note

The definitions are arbitrary but explicit, and those are different things. 50 bars is not the correct lookback and 2× is not the correct volume multiple, because there is no correct value — they are choices. Being explicit means a reader can see exactly what was chosen and change it. Being arbitrary means you must not present a result from these particular numbers as a finding about breakouts in general. Step 7 comes back to this.

Step 2: the two guards that make the measurement honest

Nothing in the condition may use today’s own data to judge today

Fragment — not a complete formula

PriorHigh   = Ref( HHV( High, BreakLookback ), -1 );
PriorAvgVol = Ref( MA( Volume, VolLookback ), -1 );

Both windows are shifted back one bar. Without the shift, HHV( High, 50 ) includes today’s high, so “close above the 50-bar high” becomes nearly impossible to satisfy — today’s high is almost always at least today’s close. And MA( Volume, 50 ) including today means a huge volume day inflates the very benchmark it is being compared against, which weakens the test in the opposite direction.

Neither error announces itself. Both quietly change what you measured.

The outcome uses the future on purpose, and must never be copied

Fragment — not a complete formula

FwdReturn = 100 * ( Ref( Close, Horizon ) / Close - 1 );

Ref() with a positive period references bars in the future — the documentation says so directly, and calls it “looking up the future”. In a formula that assigns Buy or Sell, that is look-ahead bias and the results are worthless.

Here it is the measurement itself. The study asks “what followed?”, not “what should I do?”. Measuring history that has already happened is a legitimate and completely different activity from trading on information you did not have.

⚠Warning

Ref( Close, +20 ) must never appear in a formula that produces signals. The line exists in this study only because the study produces a table of past outcomes, not decisions. Part 30’s look-ahead bias lesson shows what happens when this line escapes into a backtest, and the answer is: a spectacular equity curve that describes a market where you can see next month.

Note also that Ref( Close, 20 ) is Null within 20 bars of the end of the data — correctly, because those bars have no outcome yet. Measurable drops them. Counting them as zero would drag every average towards zero by an amount that depends on how recently you ran the study.

Step 3: choose the universe and the period, and write them down

Both are decisions, and both change the answer.

Universe. Liquid shares — a median turnover floor and a minimum price — because a “breakout” in an instrument that trades £4,000 a day is not a thing anyone could act on, and because penny stocks produce percentage returns that dominate any average they are included in.

Period. As long as your data allows, and long enough to contain more than one kind of market. A study that covers one sustained bull market has one observation of “a market”, not thousands. Part 30’s regime lesson develops this.

Data. Split- and dividend-adjusted daily bars. Unadjusted data manufactures both breakouts and volume spikes on corporate-action bars, which is precisely the event this study is looking for. This one is not a preference — an unadjusted database will produce a stronger-looking result for a purely mechanical reason.

⌀Data quality warning

Before you run this on your own database, run the data-quality checks from Part 2 on the universe you are about to use. A study is a very efficient machine for converting bad data into confident conclusions.

Step 4: the formula

Complete runnable AFL

breakout-forward-study.afl

// breakout-forward-study.afl
// Part 12 - Reality Check: Do High-Volume Breakouts Lead Anywhere?
//
// Measures what actually happened in the twenty bars after an objectively
// defined high-volume breakout, and - on a second run with one line changed -
// what happened in the twenty bars after EVERY eligible bar regardless of the
// condition. The second number is the base rate. Without it the first number
// means nothing at all, because a positive average after breakouts in a market
// that rose over the period is not evidence about breakouts.
//
// THIS FORMULA READS FUTURE BARS ON PURPOSE.
// Ref( Close, +20 ) looks twenty bars ahead. In a formula that produces Buy or
// Sell that would be look-ahead bias and the results would be worthless. Here
// we are measuring history that has already happened, which is a different
// activity: the study asks "what followed?", not "what should I do?". Never
// copy this line into a trading formula.
//
// DEFINITIONS - the whole point of the exercise is that these are arbitrary
// but explicit, so that somebody who disagrees can change one number and
// re-run rather than argue about words:
//   breakout    : today's close above the highest HIGH of the previous 50
//                 bars (the window is shifted back one bar, so today's own
//                 high cannot make the test trivially impossible)
//   high volume : today's volume at least twice the 50-bar average volume
//                 computed up to YESTERDAY, so today's volume does not inflate
//                 its own benchmark
//   eligible    : liquid and priced above a floor, measured to yesterday
//   outcome     : simple percentage change of the close over the next 20 bars
//
// ASSUMPTIONS AND KNOWN WEAKNESSES, stated before the numbers are seen:
//   - Daily bars, split- and dividend-adjusted. Unadjusted data manufactures
//     both breakouts and volume spikes on corporate-action bars.
//   - The database contains the symbols that exist today. Instruments that
//     were delisted are absent, so the measured outcomes are biased upward by
//     survivorship. This cannot be fixed from inside the formula; it can only
//     be reported.
//   - Rows are not independent observations. One symbol can contribute many
//     overlapping 20-bar windows, and symbols move together, so the effective
//     sample size is far smaller than the row count.
//   - No costs, no slippage, no position sizing, no risk model.

// ---- The one line you change between the two runs -----------------------
// False : measure only the bars where the condition was true
// True  : measure every eligible bar, which is the unconditional base rate
MeasureEveryBar = False;

Horizon       = 20;
BreakLookback = 50;
VolLookback   = 50;
VolMultiple   = 2;
MinTurnover   = 2000000;
MinPrice      = 2;

// Everything used to define the condition is shifted back one bar, so the
// condition is decidable from information available before today's close is
// compared against it.
PriorHigh   = Ref( HHV( High, BreakLookback ), -1 );
PriorAvgVol = Ref( MA( Volume, VolLookback ), -1 );
PriorTurn   = Ref( MA( Close * Volume, VolLookback ), -1 );

Eligible = Close > MinPrice AND PriorTurn > MinTurnover AND PriorAvgVol > 0;

Breakout   = Close > PriorHigh;
HighVolume = Volume > VolMultiple * PriorAvgVol;
Event      = Breakout AND HighVolume;

// The outcome. Ref() returns Null within Horizon bars of the end of the data,
// which is correct: those bars have no outcome yet and must not be counted as
// zero. Dropping them is what Measurable does.
FwdReturn  = 100 * ( Ref( Close, Horizon ) / Close - 1 );
Measurable = NOT IsNull( FwdReturn ) AND NOT IsNull( PriorHigh ) AND
             NOT IsNull( PriorAvgVol );

if( MeasureEveryBar )
    Population = Eligible;
else
    Population = Eligible AND Event;

Filter = IsTrue( Population AND Measurable );

AddColumn( Close, "Close", 1.2 );                                    // column 3
AddColumn( Volume / PriorAvgVol, "Vol / 50d avg", 1.2 );             // column 4
AddColumn( 100 * ( Close / PriorHigh - 1 ), "% above 50d high", 1.2 );// column 5
AddColumn( FwdReturn, "Fwd 20-bar %", 1.2 );                         // column 6
AddColumn( FwdReturn > 0, "Up in 20?", 1.0 );                        // column 7
AddColumn( DateTime(), "Bar (ISO)", formatDateTimeISO );             // column 8

// AVERAGE (2), MIN (4), MAX (8), STDEV (32) and COUNT (16) for the two
// outcome columns only. The average of column 7 is the proportion of
// observations that were positive; the count is the sample size; the standard
// deviation of column 6 is what tells you how little the average means.
// All of these rows appear at the TOP of the result list, not the bottom.
AddSummaryRows( 2 + 4 + 8 + 16 + 32, 1.2, 6, 7 );

SetSortColumns( -6 );

Download breakout-forward-study.afl95 lines

One line is the whole experiment:

Fragment — not a complete formula

// False : measure only the bars where the condition was true
// True  : measure every eligible bar, which is the unconditional base rate
MeasureEveryBar = False;

You run the exploration twice, changing that one value, and compare the two summary blocks.

AddSummaryRows( 2 + 4 + 8 + 16 + 32, 1.2, 6, 7 ) requests AVERAGE (2), MIN (4), MAX (8), COUNT (16) and STANDARD DEVIATION (32) for columns 6 and 7 only. Those flags are documented values, and summary rows are added at the top of the result list — so the numbers you came for are the first thing on screen, not something you scroll to find.

Column 7 is FwdReturn > 0, a Boolean. Its average is therefore the proportion of observations that were positive, which is the second number you want.

Step 5: run it and record both results

Run it with MeasureEveryBar = False, record the summary. Change the line to True, run it again, record that summary. Fill in this table — on paper, before you interpret anything:

	Condition true	Every eligible bar
Count (sample size)
Average 20-bar forward return, %
Standard deviation of that return
Proportion positive
Min / Max

=Expected result

Do not expect a dramatic difference, and do not treat a small one as a discovery. What you should definitely see is that both averages are positive if your period covers a market that rose overall, and that the standard deviation in both columns is several times the size of either average. Those two observations are the substance of this lesson.

Step 6: read it against the base rate

This is the step the original claim skips, and it is the reason this lesson exists.

Suppose the conditional average is +1.4%. That sounds like something. Now suppose the unconditional average — every eligible bar, condition ignored — is +1.2%. The condition contributed 0.2 percentage points, in a sample where the standard deviation of individual outcomes is perhaps 12%.

A positive average after breakouts, in a market that rose over the period, is not evidence about breakouts. It is evidence that the market rose. The only number that can speak to the claim is the difference, and the difference has to be judged against how noisy the individual observations are.

Three sanity questions to ask of the difference

How big is it compared with the standard deviation? If the gap between the two averages is a small fraction of the spread of individual outcomes, you cannot distinguish it from noise by looking at it.

How big is it compared with trading costs? A 0.2% edge over 20 bars does not survive commission and spread. An effect too small to trade is still interesting scientifically, but be clear which claim you are making.

How much did the sample actually contain? Which brings us to the honest part.

Step 7: what this exploration cannot establish

Every one of these is a real limitation of the design, not a caveat added for politeness. They are also written into the formula’s own header, which is where limitations belong.

The rows are not independent observations. One symbol contributes many overlapping 20-bar windows — a breakout on Monday and another on Wednesday share 18 bars of outcome. And symbols move together, so a single market-wide surge can produce hundreds of “independent” rows that are really one event. The effective sample size is far smaller than the row count, and no standard statistical formula applied to the row count will tell you the truth about it.

Survivorship bias, and it cannot be fixed here. Your database contains the symbols that exist today. Companies that broke out on heavy volume and then went to zero were delisted and are absent. This biases the measured outcome upward, by an unknown amount, and the only honest response is to state it. Part 30 takes survivorship apart properly.

No costs, no slippage, no sizing, no risk model. This is a measurement of price change, not a simulation of trading. A positive average forward return is not a profitable strategy, and Part 28 is where the difference gets built.

One set of parameters was tested. If you now try 30 bars, and 100 bars, and 1.5×, and 3×, and report the best — you have performed data snooping, and the best of twenty tries on one dataset looks good for reasons that have nothing to do with markets.

Nothing here says anything about the future. The study measured what followed these events in this period in this universe. Whether the same relationship holds in a period you have not seen is a completely separate question that this design cannot address.

?Reality check

The honest conclusion of a study like this is very often “the difference is small relative to the noise, the sample is much less informative than the row count suggests, and I cannot distinguish this from nothing.”

That is a result, not a failure. It is the result that most single-condition studies of popular technical claims produce, and being able to reach it — and say it out loud — is the skill this course is actually trying to teach. The alternative is to keep adjusting the definitions until the number looks better, which is a procedure guaranteed to succeed and guaranteed to mean nothing.

Step 8: what a stronger design would look like

If you want to push further, these are the changes that add real information, roughly in order of value:

1. Split the sample by period. Run the same study over three or four disjoint multi-year spans. A relationship that appears in one span and not the others is telling you about that span.
2. Compare against a matched control rather than everything. Instead of “every eligible bar”, use “every breakout without the volume condition”. That isolates the volume claim, which is what the original slogan is actually about.
3. Look at the distribution, not the average. Export the forward returns and look at the shape. An average pulled up by three enormous outcomes is a different phenomenon from a consistently mild positive shift.
4. Vary the horizon. 5, 10, 20, 60 bars. If an effect exists at 20 but nowhere near it, be suspicious of the 20.
5. Record everything you tried, including what did not work. That log is the only defence against fooling yourself, and it is what the screen journal was for.

▷Exercise

Run the study on your own database and fill in the table in Step 5. Then, before you look at anything else, write one sentence answering: what would I have needed to see for this to change what I do?

If you cannot answer that — if any result would have been interesting — then the study was never going to settle anything, and that is worth knowing before you run the next one.

Recap

A claim that cannot be false cannot be tested, so the first job is always to replace undefined words with numbers. Every element of the condition must be computable from information available before the bar it judges. The outcome may legitimately read future bars, because measuring the past is not the same activity as trading — but that line must never migrate into a signal formula. And the conditional result on its own is meaningless: the base rate is the measurement. What you are left with is usually a small difference sitting inside a large spread, drawn from a sample far less informative than its row count, in a universe missing everything that failed. Saying that clearly is the finding.

Check your understanding

Question 1. Why is HHV( High, 50 ) wrapped in Ref( ..., -1 ) in the breakout condition?

PriorHigh = Ref( HHV( High, BreakLookback ), -1 );
Breakout  = Close > PriorHigh;

• To make the formula run faster
• Because HHV( High, 50 ) includes today's high, so "close above the 50-bar high" would be nearly impossible to satisfy
• Because HHV cannot be used without Ref
• To delay the signal by one bar for realistic execution

Show the answer and why

Answer: Because HHV( High, 50 ) includes today's high, so "close above the 50-bar high" would be nearly impossible to satisfy

Today's high is almost always at least today's close, so an unshifted window makes the test trivially hard to pass. The same shift is applied to the volume average for the opposite reason: a huge volume day would otherwise inflate its own benchmark.

Question 2. The study finds a +1.4% average 20-bar return after the condition. Which additional facts are required before that number means anything? Select all that apply.

• The same measurement over every eligible bar in the same universe and period
• The standard deviation of the individual forward returns
• The number of observations, and how many independent events they actually represent
• The maximum single forward return in the sample

Show the answer and why

Answer: The same measurement over every eligible bar in the same universe and period, The standard deviation of the individual forward returns, The number of observations, and how many independent events they actually represent

Without the base rate the figure may be entirely explained by the market rising. Without the spread you cannot tell a real difference from noise. Without a sense of independence, the row count overstates how much you learned. The maximum is interesting but not required to interpret the average.

Question 3. This formula contains Ref( Close, 20 ). Which statements are true? Select all that apply.

• A positive period references bars in the future, as the official documentation states
• Using it here is legitimate because the study measures outcomes rather than producing signals
• It returns Null within 20 bars of the end of the data, and those rows must be excluded rather than counted as zero
• It would be equally legitimate in a formula that assigns Buy

Show the answer and why

Answer: A positive period references bars in the future, as the official documentation states, Using it here is legitimate because the study measures outcomes rather than producing signals, It returns Null within 20 bars of the end of the data, and those rows must be excluded rather than counted as zero

The distinction is between measuring history and acting on it. In a signal formula the same line is look-ahead bias and invalidates everything downstream. Treating the trailing Nulls as zero would bias every average by an amount depending on when you ran the study.

Question 4. Why is "the rows are not independent observations" the most serious limitation of this design?

• Because AmiBroker cannot compute statistics on dependent data
• Because overlapping 20-bar windows and symbols that move together mean one market event can produce hundreds of rows, so the effective sample is far smaller than the row count
• Because it makes the exploration run slowly
• Because it means the average is computed incorrectly

Show the answer and why

Answer: Because overlapping 20-bar windows and symbols that move together mean one market event can produce hundreds of rows, so the effective sample is far smaller than the row count

The arithmetic of the average is fine. What is wrong is the inference: any confidence you place in the number depends on how many independent things you observed, and the row count systematically overstates that — often by more than an order of magnitude.

Question 5. True or false: if the conditional and unconditional averages turn out to be nearly identical, the study has failed and should be redesigned until it shows something.

• True
• False

Show the answer and why

Answer: False

"No detectable difference" is a result. Redesigning until a result appears is data snooping with extra steps: given enough definitions to try, one will look good on any dataset, and it will not survive contact with data you have not seen.

Check my answersClear