Chart-Pattern Engine, Validated by Backtest

The situation

The first pattern detector labelled more than half of all charts with something. Members found that noisy, and I did not trust it. Before shipping pattern-based features to paying customers I wanted two answers: can we detect patterns cleanly, and do they actually predict anything?

What I built

A single engine with 16 detectors sharing five gates: a volatility-scaled zigzag for structure, a prior-trend requirement, a depth check, volume confirmation, and a margin-plus-recency test. One engine serves the chart analyzer, the scanner's pattern filter, and the setup board, so labels cannot drift between surfaces.

Then the honest part. I wrote a point-in-time replay harness that walks 271 symbols across eight years of daily bars, evaluating each detector only on data available at that moment, and compares forward returns against a matched base rate. 347,000 evaluations. The result: every pattern landed within about three points of the base rate. No edge.

The outcome

The rewrite cut false signals sharply: bars firing fell from 12.99% to 4.85%, labelled charts from 58% to 27%, and about 73% of charts now correctly show no pattern. The product claims were narrowed to what the data supports. Patterns describe geometry; they do not forecast. The ranking score built on top is labelled as a ranking, never a probability.

Why it matters

Most fintech content asserts. This platform measured, and published the measurement. Data discipline over marketing is the whole brand, and it started with being willing to find out the feature did not do what people wanted it to.

Stack