Chart patterns repeat. But they don't repeat in a vacuum.

The Problem With Shape-Only Matching

For six months we matched charts by shape alone. Type a ticker, get 10 similar-looking historical charts, see what happened next. It was useful — but it missed something obvious.

Imagine NVDA consolidating at its 200-day MA. A pure shape match might pull up a consolidation from February 2020 (right before COVID), November 2021 (top of the bubble), and October 2022 (bear market bottom). Same shape, wildly different market conditions, wildly different outcomes.

The 12-Dimension Context Vector

Every pattern in our database has a 12-dim context vector attached to it, capturing the market regime when the pattern formed. The features are research-backed, not made up:

Research citations: Bollerslev, Tauchen & Zhou (2009) on variance risk premium. Daniel & Moskowitz (2016) on momentum crashes in panic regimes. Gilchrist & Zakrajsek (2012) on credit spreads as macro predictors. The VIX term structure work is Johnson (2017). Our feature weights lean on their effect sizes.

Two-Stage Search

1. 1
   Shape similarity (L2 distance on 384-dim embedding)

   Every chart is compressed into a 384-dimensional vector capturing its shape. We use pgvector to find the 50 closest historical patterns by Euclidean distance in under 10ms.

2. 2
   Context re-ranking (weighted distance on 12-dim regime vector)

   Of those 50 shape-similar patterns, we re-rank to surface the 10 that also occurred in the most similar market regime. The final list is what you see — matches by shape AND conditions.

What You See On The Website

• ▸
  Regime badge in the navbar — always shows today's regime label (Bull+Calm, Bear+Volatile, etc.) with SPY/QQQ change and realized vol.
• ▸
  Context fingerprint on every match — a 12-cell colored strip showing per-dimension similarity between your query and that match. Green = similar regime, red = different regime.
• ▸
  Regime filter toggle — on any search, switch to "Regime-matched only" and the forward return stats recalculate using only matches from similar conditions.
• ▸
  Market Dashboard — a full view of current regime indicators (VIX, sectors, breadth, crowding) at /regime.
• ▸
  Accuracy by regime — our track record broken down by market conditions at /track-record. Honest reporting — some regimes we're more accurate than others.

For Developers

The context_weight parameter on our /api/v1/analyze endpoint controls the re-ranking blend:

• context_weight=0 — shape-only (default, backward compatible)
• context_weight=0.05 — recommended, shape dominates with regime adjustment
• context_weight=0.20 — strongly context-weighted, useful for regime-sensitive strategies
• context_weight=1.0 — context dominates (experimental)

Honest Caveats