Winning vectors —
cluster-first chart pattern intelligence.

The simplest way to use a chart-similarity engine is to compute a composite score over the cohort’s forward-return distribution and rank candidates. Higher score = stronger candidate.

On a 5-year, 256-week backtest over a 44-symbol mega-cap + sector-ETF universe, a score-only strategy returned roughly 34% over five years versus SPY’s 103%. The cohort statistics were correct; the problem was that a high cohort score on its own doesn’t tell you whether this particular anchor looks like the cohort members that won, or like the ones that lost. It only tells you the average outcome of the cohort.

A different question: across all 4M+ chart patterns we have, are there regions of embedding space where the typical forward return is consistently and meaningfully different from the universe baseline — and stable across years?

Cluster the V5 embedding space into K=300 mini-batch k-means clusters fit on the train period. For each cluster, compute the 5-day forward-return distribution separately for the train period and for the held-out test period, plus the per-year median over the train period (consistency check). Score each cluster by:

quality = sign(median) × |median| × |win_rate − 0.5| × consistency × log10(n)
  with sign-agreement penalty when median direction
  disagrees with win-rate direction (fat-tail-only clusters
  get attenuated).

Filter clusters where one symbol dominates (max_symbol_share > 30%) or sample is too small (n < 200 train members). The top 20 by signed quality are the winning vectors. The bottom 20 are the losing vectors: regions whose historical 5-day forward-return distribution centered roughly -1% to -2% below baseline.

Each top cluster is characterized against the population baseline. For every Layer 2 feature, compute the cluster mean and the z-score of that mean against the global distribution. Features with |z| > 0.5 are the cluster’s signature.

Loser cluster #164 (one of the top-ranked negative-distribution patterns, −1.30% median train, −1.69% median test):

days_since_earnings:           cluster 280  vs  pop 48     (z = +3.23)
days_since_ath:                cluster 1554 vs  pop 351    (z = +2.62)
pct_off_ath:                   cluster -76% vs  pop -21%   (z = -2.49)
broke_50d_low:                 cluster 24%  vs  pop 6%     (z = +0.78)
cross_sectional_momentum_rank: cluster 0.28 vs  pop 0.50   (z = -0.77)
vix:                           cluster 24.9 vs  pop 19.4   (z = +0.66)
rs_in_sector_rank:             cluster 0.33 vs  pop 0.49   (z = -0.58)
broke_20d_range_low:           cluster 28%  vs  pop 10%    (z = +0.58)

The pattern is concrete and trader-readable: stocks far below their all-time high, with stale earnings catalysts, in elevated VIX regimes, sitting in the bottom quartile of sector relative strength and overall momentum, that just broke a 50-day or 20-day range low. That description doesn’t come from a list we wrote — it falls out of the data.

The matching winner clusters look the opposite. Cluster #185 (TQQQ-led, +1.16% / 60.8% train, OOS replicates) carries elevated VIX with elevated credit spreads and tight macro — a stress-reversal regime where high-beta tech historically rallied. Cluster #1 (industrial-led, +0.86% / 58.5% train, +0.62% / 61.0% test — improves OOS) carries a steepening yield curve with rising broad-market 60-day return.

The macro features alone (VIX, yield curve, credit spreads) didn’t produce strong-enough signatures to beat passive on a risk-adjusted basis. Adding 13 daily-bar-derived features made the signatures concrete enough to actually trade. They’re cheap: pure window functions over the existing daily price+volume table.

• Volume trajectory: today’s volume z-score, 5-day volume acceleration, dollar-volume z — captures “volume is ramping” not just volume snapshot.
• Breakout flags: broke 50d high, broke 50d low, broke 20d range high, broke 20d range low, broke all-time high. Trailing windows exclude today so “broke 50d high” means today beat the prior 50 bars.
• Volatility events: gap z-score, range expansion z, range compression z — the coiled-spring signal.
• Cross-sectional ranks: relative strength rank within sector ETF, cross-sectional momentum rank across the entire universe on the same date.

These features show up across nearly every top winning and losing cluster signature. The ranks in particular surfaced the clearest loser pattern: anchors in the bottom quartile of both their sector and the universe, that just broke a 20-day range low, are the clearest negative-distribution region in the catalog.

On a 5-year, 256-week, 44-symbol universe backtest with all ablations:

Strategy           Total return   Sharpe   Hit rate   Max drawdown
SPY (passive)            +103.0%   +0.97     63.3%        -22.4%
S5 (cohort score + sticky lessons + 5d hold)
                          +75.9%   +0.74     55.5%        -22.5%
S6 (rich brief + dynamic stops)
                          +24.7%   +0.43     48.8%        -27.8%
S8 (winning-vector match + cohort-derived stops)
                          +45.2%   +0.90     28.1%         -6.6%

Total return of the winning-vector strategy lags SPY in absolute terms because it stays in cash on roughly half of trading dates when no anchor matches a known winning cluster. But it cuts max drawdown by 70% and matches passive long-only on Sharpe. Calmar ratio (annualized return divided by max drawdown) is 1.16 versus SPY’s 0.68 in this sample — a drawdown-control result, not a leverage recommendation.

The strategy isn’t supposed to beat SPY in absolute terms on a +103% bull-market universe of mega-cap names. The point is narrower: in this research backtest, cluster-matched, signature-aligned selection showed risk-adjusted separation — evidence the mined regions carry signal, not a validated trading strategy to run live.

Three tools, available both as REST API endpoints and as MCP server tools. The MCP server is published as chartlibrary-mcp on PyPI.

match_winning_vector

POST /api/v1/match_winning_vector
{"symbol": "AAPL", "date": "2024-06-21", "timeframe": "1d"}

Returns the nearest cluster, classifies it as positive- / negative- / unranked-distribution, scores how well the anchor’s current Layer 2 features align with the cluster’s signature, and returns the region’s historical forward-return distribution with cohort-percentile bands (p25/p75).

list_winning_vectors

GET /api/v1/list_winning_vectors?top_n=20&kind=both

Catalog of mined patterns: top 20 winners, top 20 losers, each with train+test medians, win rates, p25/p75, top symbol, sample size.

get_cluster_signature

GET /api/v1/get_cluster_signature/164

Full rule card for a single cluster: numeric features with |z| > 0.5 vs population, train/test outcome distribution, categorical mode (sector / regime / etc.), top symbols.

• Bull-market window. The 2021–2025 backtest period was a +103% SPY bull. A defensive long-only strategy in that regime is geometrically constrained from beating passive long-only on absolute return. Risk-adjusted metrics (Sharpe, Calmar) are a fairer comparison.
• Survivorship in the universe. The 44-symbol test universe is mega-cap names + sector ETFs that existed throughout the period. Strategy-vs-strategy comparison on the same universe is still valid, but the absolute-return numbers wouldn’t survive a wider universe with more failed companies.
• Single-seed determinism. All LLM-driven strategy variants were run with temperature=0 to remove Haiku non-determinism. Bootstrapping over multiple random universes / seeds would tighten the confidence intervals.
• Daily granularity. Walk-forward simulation uses daily OHLC. Same-bar collisions resolve as “stop fires first” (conservative). True intraday data would refine stop-vs-target attribution but is unlikely to flip the sign of the result.

Is the winning-vector catalog static or refreshed?

Refreshed. The catalog is regenerated when the underlying mining run is rerun. The API loads the most recent winning_vectors_<run_id>.json + matching centroids artifact lazily on first call after restart, so updating the artifact + restarting the API picks up new mining results without code changes.

Why K=300 clusters?

Empirical balance. K=30 was too coarse to find tight signatures; K=500+ pushes per-cluster sample sizes below the threshold where forward-return statistics are stable across years. K=300 leaves ~14K members per cluster on average, which keeps yearly-median variance low without smearing distinct regimes together.

What's the next planned extension?

Surface the symmetric negative-distribution regions as a first-class catalog entry. Cluster #164 has a -1.30% median 5-day return that's stable out-of-sample; exposing those regions explicitly is the next step. Beyond that: richer features (volume profile from minute bars, options flow, earnings-call sentiment), broader universes, and a thesis-invalidation check that flags when a live anchor's signature drifts away from the region it matched.