Fuzzy-scored chart patterns as testable rules

Imprecise matching instead of a binary template

Chart patterns that look alike still fail exact definitional matches, so recognition can be framed as an imprecise matching task rather than a binary template check.

Approximate verbal slope language can describe a desired shape so a scan can report when a time series resembles that shape. That scan is a fuzzy-pattern-indicator, which scores how closely a recent price or indicator window matches a verbal multi-segment shape, typically on a zero-to-one scale.

Segment codes and the membership score

A fuzzy-pattern-indicator may be applied to a close series or to another series such as RSI or MACD, and it reports a membership-score from zero to one for the current window.

Segment parameters use integer segment-code labels in which two means rises sharply, one means rises, zero means remains steady, minus one means drops, and minus two means drops sharply. Each segment-code encodes slope language for one stretch of the pattern.

Long-side and short-side seed paths

One long-side example scanned a two-segment close path of a rise followed by a sharp drop over a 10-day window, using a five-point change to define sharpness.

One short-side example scanned a two-segment close path of a sharp rise followed by a steady stretch over a 10-day window with the same five-point sharpness scale.

Thresholds and overlapping rule firings

A mechanical-threshold-rule is a fixed comparison that turns a membership-score into an entry, exit, or abstention decision. A mechanical long rule in the example fired when the long-side score exceeded 0.5, with a parallel threshold rule on the short-side score.

When an observation sits in overlapping membership ranges, multiple rules can fire at partial strength. Defuzzification collapses those overlapping partial firings into one numeric control or signal value. In this workflow a center-of-mass step converted the strengths into that single numeric output.

Search over encodings and thresholds

A genetic-pattern-search can start from a seed pattern encoding, generate mated and mutated variants of segment-codes, window length, and sharpness, and retain children that pass a fitness test.

In the reported optimization the original long and short pattern encodings were left unchanged while the long threshold moved from 0.5 to 0.3 and the short threshold moved to 0.07.