Pattern Watch accuracy

How to read a hit rate below 50%

A hit rate on its own is uninterpretable — what matters is expectancy: hit rate × average reward, minus miss rate × risk. Pattern calls target roughly 2:1 reward-to-risk, so a 38% hit rate at 2:1 earns about +0.14R per call (0.38 × 2 − 0.62 × 1), while a 55% hit rate at 0.5:1 loses money. The expectancy figure above does this math against every graded call, using each call’s own logged target and stop. We publish it either way — positive or negative — because the point of Pattern Watch is to teach you to demand this number from any system anyone tries to sell you. Most never publish it. Now you know why.

How this is measured

1. Every time a user invokes Pattern Watch on a chart, the AI's response is logged to pattern_watch_predictions: pattern name, confidence, breakout price, stop price, target price, and the bar timestamp.
2. A background job runs every 15 minutes. It walks predictions that are at least an hour old, fetches the real subsequent bars from Twelve Data, and checks which level the price hit first — target, stop, or neither.
3. Target hit: price crossed the AI's target before crossing the stop. Stop hit: price crossed the stop before the target. If a single bar crossed both, we conservatively score it as stop hit. Neither: 24 hours passed without either level being touched.
4. Calls where the AI returned confidence="none" or didn't supply both a target and stop are classified as "no call" and excluded from the hit-rate calculation — they're not predictions to grade.
5. Hit rate = target hit ÷ (target hit + stop hit + neither). We don't count pending predictions toward the denominator.
6. Until we have at least 30 graded high-confidence calls, the headline number is suppressed. The standard error on a percentage with N=10 is too wide to publish honestly; at N=30 it drops below 9%.

Source code for the grader is open in our repo. No human edits the outcome column — once the AI made the call, the price tape decides.