A pattern scanner is a machine for finding candidates. It is also, unavoidably, a machine for finding false positives: shapes that match the rule but mean nothing in the market.
Every scanner trades off catching real setups against generating noise. Understanding that trade-off is what separates using the tool from being used by it.
A false positive is a shape that matches the scanner's rule but carries no real meaning. Looser tolerance finds more setups and more noise; tighter tolerance finds fewer, cleaner matches but can miss valid ones. The skill is measuring the false-positive rate on a defined sample and tuning the rule against that evidence, not against a few charts you liked.
What a False Positive Actually Is
A scanner encodes a geometric rule, such as a minimum shoulder symmetry or a maximum boundary slope. A match means the geometry was satisfied. It does not mean the shape was followed by a move, a change in trend, or anything else the rule does not measure.
A false positive is exactly that: the geometry matched, and nothing happened. With a loose rule, most flags can be false positives, which is why counting matches tells you about the rule, not about the market.
Every Scanner Makes a Trade-Off
Loosen the tolerance and the scanner flags more shapes. Some will be real setups the tight rule missed, but many will be noise. Tighten it and the flags get cleaner, while valid setups that do not quite fit the geometry stop appearing.
There is no setting that finds every real setup and no noise. The only question is whether the ratio is useful for your style, your timeframe, and your tolerance for digging through false matches.
| Setting | Finds more | Costs you |
|---|---|---|
| Very loose | Every shape vaguely like the pattern | Mostly noise; too many flags to review |
| Balanced | A reasonable set of candidates | Some real setups missed, some noise kept |
| Very tight | Only textbook matches | Valid setups that do not quite fit are skipped |
Why Flag Count Tells You Almost Nothing
More flags do not mean a better scanner; they mean a looser rule. A scanner that flags forty setups a day is not finding forty opportunities, it is finding forty shapes, most of which mean nothing.
The number that matters is the false-positive rate: how many flags, on a defined sample, were followed by nothing. That is measurable, and it is the number you tune against.
| Scanner setting | Total flags in the sample period | Followed by the predicted move | False positives | False-positive rate |
|---|---|---|---|---|
| Loose tolerance (default) | 42 | 14 | 28 | 66.7% |
| Tight tolerance (higher geometry threshold) | 18 | 9 | 9 | 50.0% |
Tightening the settings cut the false-positive rate from 66.7% to 50.0%, but it also cut the number of flags that actually worked from 14 down to 9 — the cleaner rule found about 36% fewer real setups.
In this sample, the loose setting flagged 42 setups over the test period, and only 14 were followed by the predicted move — a 66.7% false-positive rate, meaning two out of every three flags were geometry with nothing behind them. Tightening the rule brought the false-positive rate down to 50%, but the number of real opportunities also fell, from 14 to 9. Counting total flags tells you almost nothing; the false-positive rate in a table like this is the number worth tracking.
Every flag, loose or tight, is a starting point. The scanner confirms the geometry. Whether the shape means anything is decided by your audit and your market context, not by the match itself.
How to Filter Without Overfitting
Pick a defined sample of history, run the scanner, and record how often a flag was followed by nothing. Then adjust the rule against that evidence, one parameter at a time, and re-test on a different stretch of history.
Filtering by false-positive rate is different from tuning until a few charts you liked look good. The first is evidence-based; the second is curve-fitting.
Frequently Asked Questions
Is a scanner with more flags better?
No. More flags usually mean a looser rule and more false positives. What matters is the false-positive rate on a defined sample.
Can I set the scanner to zero false positives?
Only by making the rule so tight it stops finding real setups too. Some noise is the price of catching anything at all.
How do I measure false positives?
Run the scanner on a fixed stretch of history and count how many flags were followed by no meaningful move or trend change. That ratio is your tuning target.
