Legend AI converts published price, volume, trend, and risk concepts into deterministic research rules. The scanner does not predict a winner and no named trader endorses the product.
Attribution identifies the published framework that inspired a rule. It does not reproduce a book, impersonate a trader, or imply affiliation. Pattern ownership follows the scanner's canonical pattern map.
Scanner measurements: 50/150/200-day moving-average alignment, rising long-term average, 52-week high/low location, relative strength, VCP contraction count and depth, volume dry-up, and fresh pivot distance. Owned replay setups include VCP, Power Play, and Three Weeks Tight.
Public methodology reference: Minervini workshop preview ↗Scanner measurements: Cup & Handle and Flat Base depth/length, pivot and breakout volume, Pocket Pivot volume, relative-strength rank, quarterly earnings growth, and general-market state. The code applies its own explicit thresholds shown in the scanner and tests.
Public methodology reference: IBD chart-pattern guide ↗Scanner measurements: price confirmation around a pivotal level, direction of the broader trend, leading-stock behavior, continuation structure, and volume expansion. The pattern map groups Bull Flag, Pennant, and Falling Wedge observations under this lens.
Public methodology reference: Livermore trading rules ↗Scanner measurements: the slope and ordering of weekly-equivalent moving averages, price above or below those averages, breakout volume, stage classification, and pullbacks toward support. Stage labels are context, not outcome forecasts.
Public methodology reference: Stage Analysis overview ↗Scanner measurements: a defined recent range, confirmed box top and bottom, breakout above the range, volume participation, and invalidation below support. The replay requires a later executable trigger rather than assuming a fill inside the box.
Public methodology reference: Darvas book overview ↗Scanner measurements: trading-range support/resistance, a fresh spring below support and recovery, Sign of Strength above resistance, price spread, and relative volume. The rules approximate observable price/volume events; they do not claim to infer an operator's intent.
Public methodology reference: Wyckoff SMI glossary ↗The historical replay runs the same detector on only the bars available at each historical step. It then applies explicit eligibility gates before a setup enters any displayed sample.
| Gate | Current replay rule | Why it exists |
|---|---|---|
| History | At least 220 stored sessions before evaluation | Moving averages and trend context need enough prior data. |
| Detector | Signal confidence ≥ 0.60 and pattern validation passes | Weak or structurally invalid fires are excluded. |
| Scanner quality | Composite score ≥ 70, trend score ≥ 6, relative-strength score ≥ 70 | The sample represents scored scanner setups, not every raw detector event. |
| Freshness | Signal close is between 3% below and 5% above its pivot and is not extended | Persistent mid-trend states are not repeatedly counted as fresh breakouts. |
| Executable entry | Starting the next session, the pivot +0.1% trigger must trade within 5 sessions; gaps fill at the higher open | No below-market or never-traded fill is invented. |
| Overlap | The next entry for that detector waits until the prior replay outcome resolves | One persistent signal is not counted every day. |
Every displayed replay percentage must travel with its sample, source-bar window, target, stop, horizon, and definition.
The share of eligible, filled setups whose intraday high reached +20% before the intraday low reached −8%. If both levels occur in the same bar, the stop wins the tie. This is not a generic win rate.
40 trading sessions after the executable entry. If neither boundary is reached, the session-40 close supplies the return. An unresolved setup without all 40 sessions is right-censored and excluded.
n is the number of eligible, executable, non-overlapping setup observations. A rate is hidden below n=30; the page says the sample is still building instead of showing a zero or a small-sample percentage.
The earliest through latest stored OHLCV dates used by the artifact. It describes the historical data span, while “generated at” states when the replay last ran. Stale or incompatible artifacts do not publish numbers.
Average return includes target exits, stop exits, and full-window timeouts. R divides that average by the configured 8% stop distance. It describes this replay sample, not an expected personal return.
No unfilled trigger, thin-sample percentage, incomplete unresolved horizon, stale artifact, personalized sizing, brokerage slippage, fees, taxes, liquidity constraints, or discretionary trade management is represented.
The artifact is regenerated off-hours each Sunday and is displayable for at most 14 days only while its source bars are also current. The committed artifact is the durable deploy baseline; an operator refresh must pass validation before replacing it. Run python -m jobs.backtest_stats, review the summary and diff, run the focused/full tests, then commit the validated JSON. Full instructions live in docs/BACKTEST_STATS.md.
Historical and illustrative, not predictive. The replay tests deterministic rules on stored daily bars. It is not a brokerage simulation, audited portfolio record, guarantee, or recommendation. Daily bars cannot establish the sequence of target and stop touches inside one bar, so ties are resolved against the setup. Pattern and legend aggregates may observe the same symbol in different, separately detected setups. Research and education only; not investment advice.
A setup whose current reward-to-risk is below 2.0:1 is not presented as actionable; the buy-zone label remains only while reward-to-risk is still ≥ 2.0:1. This live risk gate is separate from the fixed +20%/−8% historical replay definition.