I Ran 16,438 Chart Pattern Predictions. Here's What Actually Works.
The Honest Result First
Chart Library runs nightly automated forward tests: every night we take the top pattern matches across US equities, record our predicted 5-day return, then backfill the actual return when it's known. Over 16,438 predictions we now have the ground truth.
The headline: overall direction accuracy is 50.6%. Essentially coin-flip.
This is the honest number. If anyone tells you chart pattern matching alone is a trading edge, they're either selling you something or haven't tracked their own predictions. The signal-to-noise ratio in shape-only similarity is near zero once you control for market direction.
But the picture changes significantly when you condition on market regime. Below are three findings from the same dataset that tell a different story.
Finding #1: Bearish signals crush bullish signals (in this regime)
The dataset covers March-April 2026, which happened to be a pronounced drawdown period (SPY averaged -2.12% per 5-day window). In that environment, bearish pattern signals massively outperformed bullish ones:
- Bullish signals (7+ of 10 matches up): 34.7% win rate over 5 days
- Bearish signals (3 or fewer of 10 matches up): 71.2% win rate
- Spread: 36.5 percentage points
Note:Important caveat: this isn't 'bearish patterns are better.' It's 'predictions that agree with the underlying market direction are better.' In a bull market, bullish signals would dominate. The lesson is to respect the macro.
Finding #2: Stock-level volatility conditions bearish signal quality
Bucketing forward tests by the query stock's 20-day realized volatility percentile produced the clearest single-dimension effect:
- Low-vol stocks (0-40th percentile): 68.7% bearish win rate
- Medium-low (40-67th): 70.1%
- Medium-high (67-87th): 72.7%
- High-vol (87-100th): 73.4%
A 4.7pp spread between Q1 and Q4 on bearish signals. The effect scales monotonically — the more volatile the stock, the better our bearish signals worked in a drawdown. This is consistent with academic work on momentum crashes (Daniel & Moskowitz 2016) and volatility regimes.
Finding #3: VIX regime flips bullish signal quality
Here's where it gets interesting. Bucketing by VIX level produced a regime reversal on bullish signals:
- Low-VIX quartile (88-91st percentile in our data): 31.6% bullish win rate
- Mid-low VIX: 28.5%
- Mid-high VIX: 49.1%
- Highest VIX (96-100th pctl): 48.4%
A 20pp spread in bullish win rates across VIX regimes. At VIX extremes, bullish signals work much better than at VIX moderates. The counterintuitive reading: when fear peaks and contrarian setups form, bullish patterns actually follow through.
What This Means For Your Trading
Three takeaways:
- Pattern signals by themselves are noise. Any product claiming 70%+ pattern accuracy without regime conditioning is either cherry-picking or fooling you.
- Regime-conditioned signals have real edge. Our bearish signals on low-vol stocks in a risk-off regime: 85.2% win rate. That's a ~35pp edge over baseline.
- Bullish signals need different conditioning than bearish signals. The same regime variable (VIX) works in opposite directions depending on which side you're taking.
What We're Still Working On
Our 16,438 samples cover a narrow time window (March-April 2026) and one regime (bearish, elevated VIX). The findings above are honest for that period but don't automatically generalize. Our forward test pipeline adds ~520 samples per trading day, so we'll have regime-diverse data in 6-12 months.
We're also running a full historical context-vector backfill right now. Once complete, we can rerun the analysis across the full 2016-2026 corpus (~300K+ matched historical samples) rather than just forward tests. That's the definitive version.
Want to query this data yourself? The /api/v1/accuracy/by-regime endpoint returns accuracy bucketed by any of the 12 context dimensions. See chartlibrary.io/developers.
Ready to try Chart Library?
Anchor any ticker + date — see what history says about your setup, with cohort statistics, feature attribution, and AI narrative.
Try it freeLearn the methodology
Chart Library is built on four canonical concepts. Read the pillars to understand what backs the numbers in this post:
Related Articles
Why we stopped backtesting our intelligence layer (and what we found instead)
Backtests are the right tool for trading strategies. They're the wrong tool for AI reasoning infrastructure. We tested Chart Library against itself: two identical Claude agents, one with our tools, one without. A blind LLM judge scored their reasoning across 50 out-of-sample scenarios. The agent with Chart Library won 50-0. Every reasoning dimension lifted. Paired t-statistic above 10 on every dimension. Here's how we got here.
We Mined 4M Chart Patterns. Here's the Cluster-First Paradigm That Came Out.
Traditional chart-pattern intelligence is anchor-first: take a (symbol, date), find its analogs. We tried the opposite — mine the V5 embedding space offline for clusters where forward returns were consistently positive (or negative) across train and test. Top 20 winning clusters and 20 losing clusters with full feature signatures, exposed as MCP tools. Backtested S8 strategy: Sharpe 0.90, max drawdown 6.6% (vs SPY 22.4%).
One Anchor Said -3.6%. 100 Anchors Said -0.5%. The Perils of Single-Anchor Decompositions.
Decomposing a cohort of 500 historical chart patterns for NVDA produced a striking slice: anchors formed inside an earnings window underperformed by -3.6 percentage points. We ran the same decomposition across 100 different anchors. The real population effect is -0.5pp, and half of it is an event-proximity artifact that also shows up on dividend dates. Here's the audit.