OPEN-SOURCE SCRIPT
AetherEdge - Pattern Edge Learner
🖊️ Overview
Most pattern tools hard-code the textbook ("engulfing = bullish"). AE-PAT throws that out and measures each shape's forward edge from this market's own history. It encodes recent price action as a normalized signature and aggregates what past analogues did next via kernel regression — telling you which shapes actually work now, as an up-probability, expected move, and confidence (effective neighbors). It also shows learned win-rates for five classic candlestick patterns.
🔶 Key Features
Learned shape signatures — encodes the last L bars as an ATR-normalized shape and learns the forward edge from analogues
Non-parametric (kernel regression) — estimates "this shape did X" from data, no hard-coded lore
Up-probability + expected move + confidence — including how many analogues exist (effective neighbors)
Measured win-rates of classic patterns — bull/bear engulfing, hammer, shooting star, doji (the record, not the textbook)
Pattern signals — fire only when the learned edge is strong with enough analogues
Bar coloring for the edge; bar-close training and signals — no repaint
🧠 Technical Architecture
Shape signature: the last L bars encoded as four features each (body / upper wick / lower wick / return, normalized by each bar's ATR = scale-invariant), F = L×4 dims.
ML (Nadaraya-Watson kernel regression): a memory bank stores, each bar, the signature from N bars ago plus its realized outcome (up/down + forward return; no lookahead). The current signature is weighted against the whole memory by a Gaussian kernel exp(−dist²/2h²), yielding P(up), expected move, and effective neighbors (confidence). The estimate adapts as shapes change.
Stats (Beta-Bernoulli): online win-rates for five classic patterns, updated from the outcome N bars after each occurrence.
Honest scope: kernel regression + Bernoulli counting (no RL). Not deep learning, not a guarantee.
⚙️ Recommended Settings & Tuning Guide (crypto 15m–4H)
Key parameters: signature window L, outcome horizon N, up threshold (ATR), memory size, kernel bandwidth, bull/bear edge thresholds, min effective neighbors.
Window L ≈ 3–5 (larger captures complex shapes but yields fewer analogues)
N is the bars to measure the outcome; 8–15 for 15m–1H
Bandwidth controls match tolerance (smaller = stricter/fewer neighbors, larger = smoother)
Min effective neighbors ensures firing only with enough analogues
⚠️ It's a memory × dimension scan; if you hit a "script takes too long" error on long timeframes, lower the memory size
💡 How to Use in Practice
Use the pattern triangles for shape-backed entries
Read P(up) and expected move for direction/size, and stand aside when effective neighbors is low
Use the classic-pattern table to check "is the hammer working now / what's the engulfing win-rate" and drop assumptions
Combine with AE-MRB (mean reversion) or AE-CORR (drivers) to layer shape × context
⚠️ Important Notes
Needs a learning period (warmup) and memory to build; memory resets on timeframe/parameter change
Past analogues are not guaranteed to resolve the same now (regime shifts break it)
Rare shapes have few neighbors and lower confidence
Probability, not a guarantee — always do your own due diligence and use risk management
🚨 Disclaimer
This indicator is for educational and informational purposes only and is not financial advice or a recommendation to buy or sell. No method guarantees future profits; past performance does not indicate future results, and trading carries the risk of loss. All decisions are your own — use proper validation and disciplined risk management.
Most pattern tools hard-code the textbook ("engulfing = bullish"). AE-PAT throws that out and measures each shape's forward edge from this market's own history. It encodes recent price action as a normalized signature and aggregates what past analogues did next via kernel regression — telling you which shapes actually work now, as an up-probability, expected move, and confidence (effective neighbors). It also shows learned win-rates for five classic candlestick patterns.
🔶 Key Features
Learned shape signatures — encodes the last L bars as an ATR-normalized shape and learns the forward edge from analogues
Non-parametric (kernel regression) — estimates "this shape did X" from data, no hard-coded lore
Up-probability + expected move + confidence — including how many analogues exist (effective neighbors)
Measured win-rates of classic patterns — bull/bear engulfing, hammer, shooting star, doji (the record, not the textbook)
Pattern signals — fire only when the learned edge is strong with enough analogues
Bar coloring for the edge; bar-close training and signals — no repaint
🧠 Technical Architecture
Shape signature: the last L bars encoded as four features each (body / upper wick / lower wick / return, normalized by each bar's ATR = scale-invariant), F = L×4 dims.
ML (Nadaraya-Watson kernel regression): a memory bank stores, each bar, the signature from N bars ago plus its realized outcome (up/down + forward return; no lookahead). The current signature is weighted against the whole memory by a Gaussian kernel exp(−dist²/2h²), yielding P(up), expected move, and effective neighbors (confidence). The estimate adapts as shapes change.
Stats (Beta-Bernoulli): online win-rates for five classic patterns, updated from the outcome N bars after each occurrence.
Honest scope: kernel regression + Bernoulli counting (no RL). Not deep learning, not a guarantee.
⚙️ Recommended Settings & Tuning Guide (crypto 15m–4H)
Key parameters: signature window L, outcome horizon N, up threshold (ATR), memory size, kernel bandwidth, bull/bear edge thresholds, min effective neighbors.
Window L ≈ 3–5 (larger captures complex shapes but yields fewer analogues)
N is the bars to measure the outcome; 8–15 for 15m–1H
Bandwidth controls match tolerance (smaller = stricter/fewer neighbors, larger = smoother)
Min effective neighbors ensures firing only with enough analogues
⚠️ It's a memory × dimension scan; if you hit a "script takes too long" error on long timeframes, lower the memory size
💡 How to Use in Practice
Use the pattern triangles for shape-backed entries
Read P(up) and expected move for direction/size, and stand aside when effective neighbors is low
Use the classic-pattern table to check "is the hammer working now / what's the engulfing win-rate" and drop assumptions
Combine with AE-MRB (mean reversion) or AE-CORR (drivers) to layer shape × context
⚠️ Important Notes
Needs a learning period (warmup) and memory to build; memory resets on timeframe/parameter change
Past analogues are not guaranteed to resolve the same now (regime shifts break it)
Rare shapes have few neighbors and lower confidence
Probability, not a guarantee — always do your own due diligence and use risk management
🚨 Disclaimer
This indicator is for educational and informational purposes only and is not financial advice or a recommendation to buy or sell. No method guarantees future profits; past performance does not indicate future results, and trading carries the risk of loss. All decisions are your own — use proper validation and disciplined risk management.
开源脚本
秉承TradingView的精神，该脚本的作者将其开源，以便交易者可以查看和验证其功能。向作者致敬！您可以免费使用该脚本，但请记住，重新发布代码须遵守我们的网站规则。
免责声明
这些信息和出版物并非旨在提供，也不构成TradingView提供或认可的任何形式的财务、投资、交易或其他类型的建议或推荐。请阅读使用条款了解更多信息。
开源脚本
秉承TradingView的精神，该脚本的作者将其开源，以便交易者可以查看和验证其功能。向作者致敬！您可以免费使用该脚本，但请记住，重新发布代码须遵守我们的网站规则。
免责声明
这些信息和出版物并非旨在提供，也不构成TradingView提供或认可的任何形式的财务、投资、交易或其他类型的建议或推荐。请阅读使用条款了解更多信息。