Four things a percentage has to tell you
A number like sixty per cent is not a fact until four things are fixed, and a sheet that leaves any of them out has not measured anything you can check.
What counted as success. Did the price have to close higher the next day? Reach some distance? Not fall by some other distance first? Each of those is a different question and they give different answers on the same boards.
By when. A shape followed by a rise within five sessions and one followed by a rise within fifty are not the same claim. A study that lets the horizon float until the number looks good has measured the searching, not the shape.
On what. Which market, which timeframe, which years. Daily boards in one decade are not weekly boards in another.
Compared with what. This is the one that is almost always missing, and it is the one that decides whether the other three were worth collecting.
The comparison that is almost always missing
Suppose prices in some market close higher on 52 of every 100 days, for no reason anyone needs to name. That 52 is the base rate: what happens anyway, with no pattern, no rule and no reading.
Now a sheet reports that a pattern is followed by a higher close 55% of the time. Set against nothing, 55 looks like an ability. Set against 52, it is three, and three is well inside what a few hundred examples would produce by chance alone.
A percentage with no base rate beside it is not evidence. It is a number that has been allowed to compare itself with zero.
The same trap runs the other way. A pattern reported at 45% sounds useless until you learn the base rate was 38%. Neither number can be read on its own, and neither sheet usually prints both.
A success rate without a base rate is a measurement of the market, dressed as a measurement of the pattern.
What happens when you run the rules on nothing
There is a cheap way to see how much of pattern recognition is the pattern and how much is the recognising. Generate boards that contain no information at all — a random walk, where each day's close is the last one nudged by chance — and then run the book's own grader across them.
Granthara did exactly that: 1,000 invented boards, from a seeded walk so the run can be repeated, graded by the same `scripts/candle-rules.js` that names every candle in this Part.
732 of the 1,000 boards carried at least one name. Not approximate shapes. Names that passed every measured condition in the rules, including 19 hammer shapes, 20 inverted hammers, 131 bullish engulfings, 50 tweezer bottoms, 3 three-white-soldiers and 3 morning stars.
Nothing caused any of them. There was no fear in the random walk, no truck arriving late, no hotel buyer changing her order. The shapes were there because shapes are cheap, and four numbers a day will assemble into recognisable arrangements whether or not anything is happening.
- Boards generated
- 1,000, each from the last close nudged by chance
- Information in them
- none, by construction
- Boards that earned a name
- 732
- Single-candle names
- 535 — 165 long bearish, 144 long bullish, 115 spinning tops
- Multi-candle matches
- 544 — 131 bullish engulfing, 126 bearish engulfing, 50 tweezer bottoms
- Rarer names still present
- 3 morning stars, 1 evening star, 3 three white soldiers, 5 three black crows
- Grader used
- scripts/candle-rules.js, unchanged
Five questions to put to any pattern statistic
Open each one. A sheet that cannot answer it has not earned the number it printed.
None of these five asks whether the pattern is real. They ask whether the sheet has measured anything.
Why this book prints no number of its own
The four-filter card from 04.05.02 has no published success rate, and it is not going to get one. That is a deliberate refusal, so it is worth saying plainly why.
To publish a rate, Granthara would have to fix an outcome, a horizon, a market and a period, measure the base rate beside it, exclude nothing, use only information available at the time, and take out the costs. Doing that honestly is a research project, and the result would be a number about that market in those years — not a property of the card.
Publishing one without doing the work would make the card look like an instrument for predicting prices, which is exactly what 04.05.02 said it is not. A card that exposes missing evidence does not become more trustworthy by acquiring a percentage. It becomes easier to misread.
The card's claim is that it makes a reading inspectable. That claim is checkable by using it. It needs no statistic, and a statistic would not support it.
More on the two biases this page only names (optional)
Survivorship and look-ahead are the two errors that most often manufacture a result out of nothing, and Part 03 teaches both properly. They are worth reading before you judge any published backtest, including a good one.
Deeper study
What survives all of this
Not much, and that is the honest answer. A name is a compact description of what four numbers did. Location makes the description specific. The card makes the reading inspectable and the gaps visible.
None of that is a forecast, and the noise test is the reason to keep saying so: the same names appear, in quantity, in prices that mean nothing. What a reader has after Part 04 is a precise vocabulary and a way of showing their working — which is worth a great deal, and is not the same thing as an edge.
Lakshmi has never claimed otherwise. Her line has not changed since the first evening Rahul watched her write on the board: “My board tells you what happened today. It has never once told me tomorrow's price.”