Do Chart Patterns Actually Predict Returns? What the Data Says

The Eternal Debate

Technical analysis has always been controversial. Academics dismiss it as tea-leaf reading, while practitioners swear by it. The truth, as usual, is somewhere in between — and it depends entirely on what you mean by "predict."

At Chart Library, we have a unique vantage point: a database of over 16 million chart embeddings spanning thousands of stocks. We can systematically measure whether charts that look similar tend to produce similar outcomes.

What We Measured

We used our forward testing system, which runs daily automated predictions. Each day, the system identifies the most interesting chart patterns from the market, finds their historical matches, and records the predicted vs. actual forward returns at 1, 3, 5, and 10-day horizons.

This isn't cherry-picked backtesting — these are real, timestamped predictions made before the market opens, with results tracked automatically.

The Results

Chart patterns do carry information, but not in the way most traders think. Here's what we found:

• Direction accuracy hovers around 50-55% for individual matches — barely above a coin flip. No single pattern reliably predicts direction.
• Average return predictions are more useful. The mean predicted return across matches is a better signal than any individual match, especially at the 5-day horizon.
• Similarity score matters. Matches with >90% similarity have meaningfully higher direction accuracy than weaker matches.
• The biggest edge is in identifying risk. Charts with high-variance match outcomes (wide fan charts) reliably predict high future volatility, even if direction is uncertain.

The Honest Takeaway

Chart patterns are not crystal balls. They're context. They show you the range of historical outcomes for situations that looked similar to now — and that's genuinely valuable for risk management and position sizing.

The traders who get the most from pattern analysis aren't the ones who bet the farm on a single pattern. They're the ones who use it as one input among many: "History says this setup has a wide range of outcomes, so I'll size smaller" or "All 10 matches went up — that's worth paying attention to."

Where We're Headed

Our current system matches on shape similarity and pairs every match with rich metadata — regime, sector, liquidity, earnings proximity — so users and agents can mine for the conditional edge downstream. That downstream filtering is where edge actually lives, not in a single 'predicted return' number.

Near-term research directions: calibrated retrieval (already shipped — every cohort now returns conformal-corrected quantile bands with validated 80% coverage), deeper regime conditioning, and scale-invariant pattern matching so a 5-minute shape can match a 1-hour shape.