Reference · Method
Technical Analysis and What It Supports
Three different kinds of claim travel under one name: descriptions of what happened, rules for behaving consistently, and forecasts. They need completely different evidence, and only the first is true by construction.
Three kinds of claim
Almost every disagreement about technical analysis is two people arguing about different statements. It is worth separating them before anything else, because each needs its own kind of evidence and one of the three needs none at all.
Descriptions. "Volume expanded as the index fell." "Fewer issues advanced than last week." "This range is twice the average of the last twenty sessions." These are arithmetic on published data. They can be wrong only by miscalculation, and they are the entire content of the breadth and volume measures on this site.
Rules. "Exit if the close falls below the twenty-day average." A rule is not a prediction; it is a decision procedure specified in advance. Its value lies in removing a judgement from the moment when the judgement would be made worst, and that value is real whether or not the rule has any forecasting power. Claiming more for it than that is where the trouble starts.
Forecasts. "This pattern means the decline is over." Here the burden is ordinary and heavy: evidence from data the claim was not built on, with the failures counted alongside the successes. Very little of the published material in this field meets it, and the gap between the confidence of the writing and the strength of the evidence is the reason the field has the reputation it does.
What a search looks like when it is presented as a test
The figure below is not market data. It is a synthetic series generated at build time from a seeded random number generator, 260 steps of a random walk, with no structure of any kind in it, nothing to detect and nothing to predict. On that series, thirty-six moving-average periods were tested as a simple long-only rule, and the best of them is shown.
Nothing about that outcome is surprising once stated: the spread between the best and worst of thirty-six settings has to come from somewhere, and on an unpredictable series it comes entirely from which noise each setting happened to sit on. The winner is not a discovery. It is the right-hand tail of a distribution, reported as though it were a finding.
| Outcome | Period | Final value | What it is |
|---|---|---|---|
| Best setting | SMA(22) | 109.4 | The number that would be published, with a chart, as evidence that the rule works. |
| Median setting | SMA(15) | 103.4 | A fairer estimate of what an unseen setting would have done, and still not a forecast. |
| Worst setting | SMA(7) | 88.5 | The same rule, the same series, a different arbitrary number. Never published. |
| Holding the series itself | — | 89.2 | The benchmark the rules have to beat, which is where a comparison should start. |
Notice what the last row does to the story. This particular walk ended below where it started, so the winning rule did not merely make money. It beat holding the series by a comfortable margin, on data containing nothing to beat it with. That is the shape of almost every published backtest result: a rule that avoided some of the declines, on one sample, chosen because it did.
Two further points about the table, because both are routinely skipped. It contains no transaction costs, and a rule that moves in and out of the market accrues them on every switch, enough on a short average to consume the whole difference between the best setting and the benchmark. And the comparison that matters is against the last row rather than against zero: beating nothing is not a result.
What survives, and why it is mostly counting
Set the forecasts aside and a usable core remains. It is smaller than the field claims and more dependable than its critics allow, and it has one property in common: every item in it counts something that was published rather than inferring something that was not.
Breadth. How many issues advanced, how many declined, how many reached new extremes. A breadth reading is a census of the session. Its interpretation is arguable; the count is not, and it answers a question no index level can: whether a move was general or the work of a few large constituents.
Volume. How much trading a price move required. The number comes from the exchange, and comparing it against the instrument’s own recent baseline is the most useful single habit in this reference. It is also where the honest caveats live, which the short-volume study demonstrates on real published data: half of all consolidated volume is short-marked, and it means almost nothing about sentiment.
Range and volatility. Average true range states how far an instrument has been moving in its own units, which turns a stop distance or a target from a round number into something scaled to the instrument. It forecasts nothing and it makes every other number on the chart comparable.
How to test a claim you have just read
Four questions, in this order. They are quick, and they dispose of most of what circulates without any need to run a backtest.
How many variants were tried? If the answer is not stated, assume many. The figure above is what an unstated sweep looks like.
Where did the threshold come from? A number like "above 70" or "a ratio under 0.5" was computed on some sample. Which market, which years, and what happened when it was applied to a different one? Thresholds transfer badly, which is why every page in the library that quotes one says to recompute it as a percentile of your own series.
Were the failures counted? A pattern illustrated by three occasions when it worked is an illustration, not evidence. The question is how many times the same configuration appeared and nothing followed, and that number is almost never in the article.
What would refute it? If no observation could, the claim is not about the market. This is the most efficient of the four questions and it is worth asking of this site as readily as of anything else: each page here states what would change its mind.
Why this site is organised the way it is
The indicator library is grouped by what each measure counts rather than by how it is drawn, because the grouping is the argument: measures that count the same thing tend to agree with one another and to fail together, and knowing which family a new indicator belongs to tells you more than its name does.
Every page follows the same order — what it is calculated from, what it is documented to do, where it misleads — and each one names its sources and its data. The calculators print their intermediate steps for the same reason: a figure you can reproduce is a figure you can dispute productively, and most disputes in this field turn out to be about the inputs rather than the method.
None of that makes the material here correct. It makes it checkable, which is a different and more attainable standard, and the one worth holding any reference to.
Frequently asked questions
Does technical analysis work?
The question is too coarse to answer, which is why it produces a decade of argument. Technical analysis contains three different kinds of claim: descriptions of what a market has done, rules for behaving consistently, and forecasts of what happens next. The first are true by construction, the second have real value that has nothing to do with prediction, and the third are where the evidence is weak. Separating them is the single most useful thing a reader can do, and almost no published material does it.
Which measures on this site do you consider defensible?
The ones that count something rather than predicting something. Breadth measures state how many issues took part in a move, which is a fact about the session. Volume measures state how much trading a price move required, which is another. Average true range states how far the instrument has been moving in its own units. None of those is a forecast, all of them are checkable against the published data, and each one changes how a chart reads.
And the least defensible?
Anything with a fixed threshold borrowed from another market or another decade, and anything whose evidence is a gallery of selected examples. A specific oscillator level said to mark an extreme, a pattern illustrated only by the occasions it preceded a move, a divergence rule with no count of how often the divergence led nowhere: these are not wrong so much as unsupported, and the distinction matters because an unsupported rule can be tested and a wrong one cannot be rescued.
Why does a rule have value if it does not predict?
Because most of what damages a result is behaviour rather than analysis, holding a loss because closing it makes it real, adding to a position out of conviction, changing horizon after the fact. A rule that is specified in advance removes those decisions from the moment they would be made badly. It is worth being clear that this is a claim about discipline and not about the market, and that a rule earning its place this way needs no forecasting power at all.
What does a proper test of a rule require?
Four things, and dropping any one of them invalidates the result: data the rule was not designed on, transaction costs and slippage of a realistic size, a fixed parameter chosen before the test rather than the best one found afterwards, and a count of how many variants were tried. The last is the one nearly always omitted, and it is decisive, testing forty settings and reporting the winner is not a test, it is a search.
What is the multiple-comparisons problem in this context?
That the best of many results is a biased estimate of what any of them will do next. If you test forty moving-average periods on the same series, the top one will look good even when the series is unpredictable, because you have selected the setting whose noise happened to be favourable. The figure on this page shows exactly that on a synthetic series with no structure in it at all, the winning rule produces a rising equity curve on coin flips.
Is volume more informative than price?
Not more informative, but harder to argue with, which is why this reference is organised around it. A volume figure is a count of shares that changed hands, published by the exchange; how you interpret it is contestable but the number is not. That makes volume claims testable in a way that pattern claims usually are not, and it is why the measures here that survive scrutiny are mostly the counting ones.
Do you publish signals or recommendations?
No, and not out of caution: a static reference cannot honestly carry a signal, because a signal is a statement about a specific moment and this page will be here in five years. What is here instead is the arithmetic of each measure, the data it needs, and the failure modes documented for it. That remains true whatever the market does, which is the only kind of claim a reference page should be making.
Where should someone start?
With volume and breadth rather than with indicators, because both are counts and both change how everything else reads. Learn what an advance/decline count includes and excludes, learn to compare a session’s volume against a baseline rather than against an impression, and only then look at the oscillators, most of which turn out to be one of those two ideas with smoothing applied. The indicator library is ordered on that basis.