Historical pattern matching

The idea

Traders have eyeballed chart patterns for a century — flags, wedges, double bottoms. The weakness was never the premise that price behaviour rhymes; it was that humans can only remember a few dozen examples, tend to remember the ones that worked, and can't attach a base rate to any of them. Pattern matching at machine scale fixes exactly those three failures: the library is large, it is complete (winners and losers), and every match comes with measured outcomes.

From candles to a fingerprint

A window of recent price bars is converted into a fixed-length numerical vector — an embedding. The encoding is scale-free: it captures the shape of the move (normalised returns, the volatility profile, where the window sits relative to its own range) rather than absolute price, so a pattern on a $0.002 memecoin can legitimately match one on a $60,000 crypto major. Windows are encoded per timeframe, because a shape that resolves in fifteen minutes and the same shape on a daily chart are different events with different base rates.

Searching millions of windows in milliseconds

The fingerprint is queried against an approximate nearest-neighbour vector index — the same class of technology that powers modern semantic search and recommendation systems — holding historical windows from many assets and years of data. The search returns the most similar historical moments along with their similarity scores, filtered to the timeframe you are trading so hourly questions are answered with hourly evidence.

From neighbours to a probability

Each matched window is stored with its forward returns over several horizons — what price actually did one bar, four bars and a session later. The engine aggregates the neighbours' outcomes into a distribution, weighting closer matches more heavily and discounting stale ones (a pattern from a structurally different market era earns less trust than one from last quarter). The spread of that distribution — not a fixed formula — is what produces the forecast cone you see on the chart: when history disagrees with itself, the cone is wide and the engine says so.

Honest limits

Pattern matching is evidence, not prophecy. It assumes the future resembles the past, which regime breaks (a delisting, a war, a rate shock) violate by definition. That is why this channel gets one bounded vote in the ensemble rather than the whole say, and why its weight is continuously re-audited against realised outcomes.