Level 1 · Chart ReaderLessonPart 01 · page 6 of 625 min

25Minutes

1Sources

StandardRequires

Probability, Not Prediction

Almost everyone arrives at technical analysis asking “what is going to happen next?”. It is the natural question and it is unanswerable, which is why the material that promises to answer it is usually the material worth least. This lesson replaces it with a question that can be answered, and then sets out the six-step loop this course uses to answer it. That loop is the spine of everything that follows: every lab, project, reality check and capstone component in the remaining parts is one pass around it.

By the end you should be able to explain why a base rate has to be measured before a conditional result means anything, describe what an indicator can and cannot contain, list the specific limitations that constrain all price-based analysis, and run the loop yourself on an idea of your own.

From “what will happen” to “what usually happened”

Replace the unanswerable question with this one:

In situations resembling this one, what happened afterwards — how often, by how much, with how much variation, and across how many observations?

Every word earns its place. Situations resembling this one requires you to define the situation precisely enough that a computer can find every past instance. What happened afterwards requires a horizon: the next bar, the next ten bars, until an exit rule fires. How often is a frequency, not a certainty. By how much prevents you from celebrating a high hit rate attached to trivial gains and occasional catastrophes. How much variation keeps the spread of outcomes in view, because an average alone hides everything that matters about risk. And how many observations is the difference between a finding and an anecdote.

Answering that question gives you a distribution: a description of the range of things that followed, with their relative frequencies. It never gives you the next outcome. A rule with a favourable distribution can lose on its next twelve trades without anything having gone wrong, and this is not a caveat bolted onto the end — it is the central fact you have to be able to live with.

Base rates come first

Here is the mistake that invalidates more amateur analysis than any other. Someone measures what followed a pattern, finds that price rose 56 per cent of the time over the next ten days, and concludes that the pattern has some merit.

The missing number is the base rate: what followed all bars in the same universe over the same period, pattern or no pattern. Suppose that over that period and that universe, the ten-day forward return was positive after 54 per cent of all bars — which is entirely plausible for a share index during a rising decade. The pattern’s 56 per cent is then two points above a benchmark that required no analysis, no rule and no trading at all. Two points may or may not survive a check on sample size, and two points of hit rate is a very thin margin to set against the round-trip costs from lesson 3.

The base rate is the thing your rule has to beat. Without it, a conditional number is uninterpretable in either direction:

A high figure can be entirely inherited from the market’s own drift.
A modest figure can be genuinely interesting if the base rate was lower still.
A figure that looks poor in isolation can be useful if the rule was active only during periods when everything else was worse.

There is a second, subtler version of the same error, and it appears in almost every pattern-trading article you will ever read: reporting how often the pattern “worked” without defining what counts as working, over what horizon, and compared with what. A pattern that “works 70 per cent of the time” has told you nothing until you know both the horizon and the 70 per cent’s benchmark.

And note what the base rate does not settle: the size of the outcomes. A rule that is right 56 per cent of the time with small gains and rare large losses can be considerably worse than one right 45 per cent of the time with the reverse profile. Frequency and magnitude have to be examined together, which is why Part 29 devotes a whole lesson to why win rate on its own tells you very little.

Why “the indicator predicts …” is the wrong frame

Every indicator in this course, and every indicator that has ever been written, is a function of past data. A 50-day moving average is a weighted sum of fifty closing prices you already had. An RSI is an arithmetic rearrangement of recent gains and losses. A Bollinger band is a mean and a dispersion measure computed from the same series that produced the mean.

No transformation can add information that was not in its input. If an indicator appears to know something about the future, one of three things is true: the information was already in the price history and the indicator has made it easier to see; the apparent knowledge is an artefact of how it was selected and tested; or the calculation is reading data it should not have access to, which is look-ahead bias and is the subject of Part 30’s opening lesson.

What indicators genuinely do is worth stating positively, because the course spends several parts building them:

They compress. Fifty numbers become one, in a way you can reason about.
They define states objectively. “Above its 200-day average” is unambiguous, computable, and identical for every reader.
They make instruments comparable. A bounded oscillator puts a 4.00 share and a 400.00 share on the same scale.
They expose relationships — between price and volume, between an instrument and its sector, between today’s range and the typical range.

So the well-formed question is never “what does this indicator forecast?”. It is: does conditioning on this state change the distribution of what follows, on data that was not used to choose the state, after costs? The clause about data not used to choose the state is doing an enormous amount of work, and Parts 31 and 32 are largely about honouring it.

What technical analysis cannot do

An honest list. None of these is an argument for abandoning the field; together they are the list of things a serious practitioner has to account for.

It cannot see what is not in the record. Undisclosed fraud, an unannounced deal, a pending court ruling, a change to regulation. Price may react when these become known; the history could not contain them beforehand.
Relationships are not stationary. The population of participants, the market structure, the cost of trading and the regulatory environment all change. A relationship measured over 2005 to 2015 is evidence about that period, and only an assumption about the next one.
The signal is small relative to the noise. Day-to-day variation dwarfs any effect you are likely to find, so evidence accumulates slowly and requires many independent observations before it means anything.
Costs apply to every trade. Spread, commission, impact and slippage are charged whether the idea was right or wrong, and they scale with turnover. Many effects that are real in the raw data are not large enough to pay for the trading required to capture them.
Capacity is limited by liquidity. An effect concentrated in the smallest, thinnest instruments may be untradable at any meaningful size, for the reasons lesson 3 set out.
The data is imperfect. Splits, dividends, delistings, bad prints, missing bars and survivorship all distort results in ways that usually flatter them. Part 2 is entirely about this.
Search finds patterns in noise. The characteristic failure of systematic method, from the previous lesson, is also a constraint on the field: the more rules you try against one dataset, the better the best of them looks whether or not anything is there. It is arithmetic rather than misfortune, which is why care taken inside any single test does not protect you from it.
History does not establish the future. The most you can obtain is a well-measured historical distribution plus a judgement about whether the conditions that produced it still hold. That judgement is yours and it is not derivable from the data.

One more, less often stated: a regularity that becomes widely known and is easy to exploit tends to attract capital, and capital tends to reduce it. This is a plausible mechanism rather than a law, and pleasingly it is testable — you can measure whether a documented effect weakened after it was published. Part 30’s material on regime dependence gives you the tools.

The research mindset

The attitude that makes the method work can be put in seven lines.

You are trying to break your own idea, not to build a case for it. If you find yourself looking for one more filter that turns a poor result into a good one, you have switched roles without noticing.
Specify before you look. Write down the rules, the universe, the period, the costs and the criterion for success, and only then run the test.
Hold data back, and count how often you have used it. An out-of-sample set consulted five times is no longer out of sample.
Count your attempts. Twenty variations tried means the best of twenty results, which is a very different thing from one result.
Prefer fewer parameters. Every parameter is another dimension in which history can be fitted.
Treat “no detectable effect” as a genuine result. It is the most common outcome of honest research and it is enormously cheaper than trading a bad idea.
Write the assumptions down. Timing, costs, universe, data source, what you did with missing values. A result without its assumptions is not reproducible, including by you.

The loop this course returns to

Everything above assembles into six steps. This diagram will reappear throughout the course; it is worth learning now.

The research loop

HypothesisA claim specific enough to be wrong
RulesObjective enough for a computer
TestA measurement designed in advance
EvidenceWhat actually happened, with its spread
RiskWhat being wrong would have cost
DecisionAdopt, revise, discard — and record it

Every lab, project and reality check in this course is one pass around this loop.

Hypothesis

A claim specific enough that some observation could contradict it. At minimum it names the universe, the condition, the outcome measured and the horizon. Compare these two:

Weak: “Breakouts from consolidation work well.”
Usable: “Among shares with average daily turnover above a stated threshold, the twenty-day forward return following a close above the highest high of the previous fifty bars has a higher mean than the unconditional twenty-day forward return over the same universe and period.”

The second can turn out to be false, which is the only property that matters at this stage. Write down, at the same time, what result would make you abandon the idea. Deciding that afterwards is not a decision.

Rules

The hypothesis expressed so that two people implementing it independently would generate identical trades. That requires more than an entry condition: the exit, the holding period, the universe and its filters, the timing convention — which bar the signal is computed on and which bar it can be acted on — the treatment of missing data, and what happens when two signals conflict or when capital runs out.

If any part still requires interpretation, it is not yet a rule. Part 27 is devoted to this translation, because it is where most ideas quietly change into something else.

Test

A measurement you designed before running it. It specifies the period, the base rate to compare against, the cost assumptions, and which data is being held back. It also specifies, in advance, what counts as support and what counts as refutation.

The temptation at this step is to run the test first and decide what it means afterwards. That converts a measurement into a search, and a search always finds something.

Evidence

What actually happened — and never a single number. You need the sample size, the dispersion of outcomes, the worst cases, and how the result varies across sub-periods and across the universe. A result driven by one extraordinary year, or by five symbols out of five hundred, is not evidence for a general claim; it is evidence about that year or those symbols, which is a different and much smaller finding.

Parts 29 and 30 give you the specific metrics and the specific ways this step is fooled.

Risk

What it would have cost to be wrong, expressed in terms you could actually have lived with: the depth and duration of the worst drawdown, the longest run of consecutive losses, how much of the account was exposed at once, how correlated the positions were, and what a realistic worst case looks like rather than the worst case that happened to occur.

Risk sits before Decision deliberately. An idea with acceptable evidence and unacceptable risk is not a candidate, and discovering that after adoption is expensive. Part 34 builds this out properly.

Decision

Four legitimate outcomes: adopt, revise, discard, or park for want of enough data. Record which one you chose and why, with the date, so that in a year you can tell what you knew at the time.

“Revise” sends you back to the start, and it carries a cost that is easy to overlook: each revision uses up some of the independence of your data, because the next test is being run on data you have now seen. Count the laps. Three revisions of one idea is research; thirty is an exercise in fitting history, however sincerely each one was motivated.

What changes from here

The rest of this course is the loop, applied at increasing scale with better instruments.

Part 2 makes the data trustworthy enough to test anything on. Part 3 puts AmiBroker in front of you, and Parts 4 to 7 build the observational skills and the vocabulary — chart reading, structure, indicators, patterns — each one careful to separate what an idea claims from whether it holds. Parts 8 to 16 give you AFL, the language in which rules become precise enough for step two. Parts 12, 27 and 28 turn rules into scans and backtests, which is step three. Parts 29 to 34 are step four and step five: reading evidence honestly, recognising the ways it deceives, and measuring what being wrong would have cost. Part 35 assembles the whole thing into a repeatable process with the human decision left where it belongs.

The reality checks scattered through the course are the loop applied to specific popular claims: does support hold, does RSI above 70 mean anything, do high-volume breakouts lead anywhere, is the golden cross worth having. The course does not promise you a favourable answer to any of them. It promises that you will be able to work out your own, and to say honestly how much confidence the answer deserves.

Recap

You have swapped an unanswerable question for a measurable one. You know that a conditional result is meaningless without its base rate, and that frequency without magnitude is equally meaningless. You know why an indicator, being a function of past data, cannot contain information its inputs lacked — and what it usefully does instead. You have an honest list of the constraints that bind all price-based analysis. And you have the six-step loop, which from here on is simply how the course works.

That completes the foundations. Part 2 turns to the data itself: what is inside a bar, what a bar throws away, and the specific ways market data gets quietly corrupted before it reaches your analysis.

Sources for this lesson

1 verified · checked 2026-08-31

01AmiBroker's Technical Analysis Guide — Introduction§ Introductionamibroker.com/guide/ta.html2026-08-31

Every technical claim on this page was checked against the official AmiBroker documentation on the date shown. Where the course disagrees with folklore, the source is how you can tell which one to trust.