AlphaTRADER Academy
Pattern Engine
AlphaTRADER runs three independent pattern detectors in parallel: a geometric harmonic scanner, a deep-learning candlestick classifier, and an AI synthesis layer that ranks the full output. Each answers a different question. Mistaking one for another is how traders blow up.
"A pattern by itself is a coincidence. A pattern at the right place is a thesis." — anonymous prop trader
Three Engines at a Glance
Same input (price + volume bars) → three orthogonal interpretations. Each output is a candidate, not a trade.
Geometric Fibonacci ratios. Detects XABCD shapes whose legs match Bat / Gartley / Butterfly / Crab / Cypher / Shark proportions.
- Method: deterministic ratio matcher
- Output: PRZ + entry/SL/TP1-4 (Carney multipliers)
- Edge source: reflexive symmetry, level confluence
Deep-learning candlestick classifier. CNN trained on labeled bar windows recognises Hammer, Engulfing, Star, Doji, etc., with calibrated confidence.
- Method: neural net over OHLCV windows
- Output: pattern + confidence 0.25–1.00
- Edge source: microstructure of single bars
AI synthesis layer. Curates the best Detector + Harmonic candidates, adds Wyckoff context, ranks by composite probability.
- Method: LLM-driven multi-signal review
- Output: ranked setups + thesis text
- Edge source: filtered confluence
Mental model: Harmonic finds the geometric where. Detector finds the microstructural when. Hidden Gems finds the contextual why. The trader's job is to see all three agree.
Choosing The Right Timeframe
| Timeframe | Pattern Reliability | Best For Trader Style |
|---|---|---|
| M15 / H1 | Lower — high noise, false signals frequent | Day / scalp execution; only with strong HTF context |
| H4 | Strong — intraday structure with enough sample size | Swing setups, primary execution timeframe for most traders |
| Daily | Premium — clean structure, slow but high-conviction | Position trades, primary HTF anchor for any style |
| Weekly / Monthly | Highest — patterns rarely fail, but rare to form | Multi-week swing, macro positioning, cycle harmonic |
Lower TF patterns get false signals; higher TF patterns are rarer but more reliable. The classic top-down workflow: identify HTF Wyckoff phase, then drop to LTF for pattern entry inside that context.
Engine 1 — Harmonic Patterns
Origin: Scott Carney / H.M. Gartley. The thesis is that markets reflect Fibonacci proportions because traders react to the same measured retracements at scale.
XABCD Anatomy
Five pivot points labeled X → A → B → C → D. Each leg's length is a Fibonacci ratio of the previous leg. Pattern is "complete" when point D is hit — that's the entry zone (PRZ — Potential Reversal Zone).
Pattern Catalog (Ratios)
| Pattern | XA→AB | AB→BC | BC→CD | XA→AD | Character |
|---|---|---|---|---|---|
| Bat | .382–.500 | .382–.886 | 1.618–2.618 | .886 | Tightest SL, highest win rate |
| Gartley | .618 | .382–.886 | 1.272–1.618 | .786 | Classic, mid trend retracement |
| Butterfly | .786 | .382–.886 | 1.618–2.618 | 1.272–1.414 | Extension beyond X, exhaustion |
| Crab | .382–.618 | .382–.886 | 2.618–3.618 | 1.618 | Aggressive overshoot — high R:R |
| Cypher | .382–.618 | 1.130–1.414 | 1.272–2.000 | .786 | Stronger BC extension, fast PA |
| Shark | — | 1.130–1.618 | 1.618–2.240 | .886–1.130 | 5-0 variant, liquidity sweep |
PRZ — Potential Reversal Zone
Where multiple Fibonacci levels cluster around point D. Tighter cluster = stronger PRZ. The platform displays PRZ as the entry zone, not a single line.
- ▸Entry: at the D-point (the published entry price).
- ▸Stop loss: beyond X with a 10% × XD buffer — the pattern invalidates if price closes past origin. For ABCD without an X-point, SL falls back to
entry ± 0.382 × AD. Direction-safety guard ensures SL is always on the opposite side of entry from the trade direction. - ▸TP ladder: Carney targets — fixed multiples of
XDprojected past D. Per-pattern table below.
TP Ladder by Pattern (multiples of XD)
| Pattern | TP1 | TP2 | TP3 | TP4 | Reading |
|---|---|---|---|---|---|
| Bat / Gartley / ABCD | .382 | .618 | 1.000 | 1.272 | TP3 = back to X. TP4 = mild extension past X. |
| Butterfly / Crab / Cypher | .382 | .618 | 1.272 | 1.618 | Extension patterns — TP3/4 sit BEYOND X. Higher R:R, lower hit rate. |
| Shark | .500 | .886 | 1.130 | 1.618 | 5-0 variant — first take at midpoint, deepest reach is full extension. |
H&S, Flag, Pennant, ABC use generic projected-height targets: TP1 = 0.5×, TP2 = 1.0×, TP3 = 1.5×, TP4 = 2.0× of the pattern's own projection range (head→neckline distance for H&S, pole height for Flag, BD distance for ABC).
Worked Example — Bullish Bat XAU
Freshness Rule (CRITICAL)
Harmonic signals decay fast. A pattern is only actionable within roughly one candle duration after the candle closes — past that, the level still exists but price has had enough time to invalidate the entry trigger.
Example: an H4 candle opens 08:00, closes 12:00 — entry trigger is fresh until ~16:00. After that, the geometry is historical reference only. Older signals can still be watched as levels but should not be entered as if the pattern just completed.
Engine 2 — Deep-Learning Candle Detector
A trained convolutional network classifies candlestick formations the same way an expert eye does — but at scale, across all assets and TFs, every candle close.
Architecture Brief
- ▸Input: rolling window of OHLCV bars (typically 5-20 bars, depending on pattern).
- ▸Network learns spatial features (body/wick ratios) and temporal features (sequence dependencies).
- ▸Output: pattern label + softmax probability (calibrated to confidence).
- ▸Re-runs every candle close. Each pattern surfaces with its own SL anchor and TP ladder.
Reading Confidence
- ≥ 0.85Textbook formation. Take it if context agrees.
- 0.65–0.85Likely valid. Demand additional confirmation.
- 0.45–0.65Borderline. Use only as confluence, never alone.
- < 0.45Filtered out by default (min threshold 0.25 in DB).
Confidence is relative, not a literal win-rate. A 0.90 Hammer in chop is still chop.
Detected Patterns & Their Real Meaning
- Hammer — long lower wick rejection at support; spring-like
- Inverted Hammer — pre-reversal probe; needs confirmation
- Bullish Engulfing — sellers absorbed; aggressive buyer step-in
- Morning Star — 3-bar exhaustion; high-conviction reversal
- Piercing Line — gap-down rejection; weaker than engulfing
- Tweezer Bottom — twin-low rejection; level retest signal
- Shooting Star — long upper wick rejection at resistance
- Hanging Man — same shape as Hammer but at top — context flip
- Bearish Engulfing — buyers absorbed; institutional unwind
- Evening Star — 3-bar topping; mirror of Morning Star
- Dark Cloud Cover — gap-up failure; sell pressure returns
- Tweezer Top — twin-high rejection; resistance hold
- Doji — equilibrium; signal value depends on prior trend
- Spinning Top — small body, two wicks; loss of momentum
- Three White Soldiers — strong continuation up
- Three Black Crows — strong continuation down
- Harami — inside bar; volatility compression — breakout pending
- Harami Cross — Doji inside prior bar; sharper reversal odds
- Marubozu — bodyless wick; institutional flow signal
Stop Loss — APR-Floored Buffer
SL is anchored beyond the signal candle extreme with a buffer that respects period-level volatility:
SL_long = bar_low − max(0.5 × bar_range, 0.5 × APR(14)) SL_short = bar_high + max(0.5 × bar_range, 0.5 × APR(14))
APR(14) = mean (high − low) of the last 14 closed bars at the signal granularity. Acts as a volatility floor — guarantees the stop is never tighter than half the period's typical range, even when the signal candle itself is unusually small.
TP Ladder Logic (1R / 2R / 3R / 4R)
The platform calculates TPs as fixed multiples of the initial risk window R = |entry − stop|. Each TP is meant as a scale-out level, not a single all-in target.
Engine 3 — Hidden Gems (Neural Dashboard)
A synthesis layer. Reviews fresh Detector + Harmonic candidates, cross-references Wyckoff structure, and writes a thesis. The output is an opinionated short-list, not a feed.
What the AI Actually Does
- Pulls all fresh Detector + Harmonic signals across the watchlist.
- Pulls Wyckoff phase, COT, sentiment, MTF context for each candidate.
- Asks the language model to score thesis quality, not just pattern presence.
- Returns a curated set with composite confidence, levels, and a written analysis.
- Each setup carries an expiry timestamp — once stale, it disappears from the active list.
How to Read a Gem Card
- Confidence % — composite, not a single-model probability. 70%+ is the realistic ceiling.
- Pattern type — the structural anchor (e.g., "Bat", "Bullish Engulfing").
- Deep dive analysis — the AI's written thesis, including the invalidation case.
- Key insights — bullet hits on confluence sources (COT, sentiment, MTF).
- Expires — past this timestamp the thesis is no longer valid.
Important: confidence is comparative, not predictive.
A 75% Hidden Gem isn't promising 75 wins out of 100. It's saying this setup is in the top decile of what the engine has seen lately. Calibrate against your own backtest, not the number itself.
When Pattern Meets Schematic
Pattern Engine output isolated = noise. Pattern Engine output at the right Wyckoff phase = thesis. These are the cross-products worth flagging.
| Pattern Signal | Wyckoff Context | Confluence Read | Conviction |
|---|---|---|---|
| Bullish Bat at .886 | Spring (Schematic #1) | Geometric retest matches stop-hunt low — composite operator absorption. | A+ |
| Hammer + high vol | LPS / BUEC (post-test) | Pullback ends with absorption candle — continuation entry. | A |
| Bearish Butterfly | UTAD (Schematic #1) | Geometric overshoot beyond TR confirms upthrust trap. | A+ |
| Bearish Engulfing | PSY / SC top zone | First sign of supply emerging at distribution start. | B+ |
| Crab pattern | Phase E (markup running) | Pattern correct, but counter-trend in Phase E — fade only with explicit weakness signs. | C |
| Doji on no volume | Mid-range consolidation | Indecision in noise — no actionable signal regardless of confidence. | D |
| Hidden Gem (compound) | Aligned with Schematic + COT | By definition the AI already filtered for context — treat as elevated baseline. | A |
Operating heuristic: never enter a pattern signal you cannot map onto a Wyckoff phase. If you can't say "this is the Spring / LPS / UTAD of Schematic X", you're trading shape, not story.
Failure Modes — How Each Engine Lies
Knowing how a system fails is more valuable than knowing how it succeeds. These are the documented blind spots.
Harmonic
- ✗Busted D: price overshoots PRZ by >3 ATR — pattern is invalid even if the candle later reverses.
- ✗Counter-trend in markup: Bearish Bat in Phase E gets steamrolled. Geometry without context = death.
- ✗Over-fitting to noise: intraday harmonic on M5 in chop is statistical illusion.
- ✗Missed freshness: taking a 6-hour-old H4 signal that already ran without you.
Detector
- ✗Pattern in trend: Bullish Engulfing during a parabolic uptrend is continuation noise, not a reversal.
- ✗News-window confidence: a 0.92 Hammer 30 sec before NFP is a coincidence, not a setup.
- ✗Confidence ≠ context: 0.85 in chop << 0.65 at HTF demand.
- ✗Volume mismatch: reversal candle on no volume — institutions absent — likely fades.
Hidden Gems
- ✗Stale thesis: setup expired but you found it later — invalidation conditions may have already triggered.
- ✗Compounding own bias: AI may anchor on Wyckoff phase classification even when phase is ambiguous.
- ✗Fewer ≠ better: empty board doesn't mean no opportunity — it means nothing crossed the threshold this run.
- ✗"Confidence theatre": a 78% number lulls you out of independent verification.
Operational Cheatsheet
Cut-out reference. Print it. Tape it next to your monitor.
| If you see… | Confirm with… | Invalidate at… | Action |
|---|---|---|---|
| Fresh Harmonic at PRZ | HTF context + volume on entry candle | Close beyond X-point | Enter at the D-point, SL beyond X with buffer |
| Detector ≥ 0.75 conf | Wyckoff phase agreement | Stop trigger | Take, scale at TP1/2/3 |
| Detector 0.45–0.65 | Two other independent signals | Stop trigger | Confluence only — never standalone |
| Hidden Gem 70%+ | Read invalidation case in thesis | Documented invalidation level | Take if expiry not breached |
| Pattern in opposite-trend phase | Demands additional reversal evidence | Continuation past signal high/low | Default: skip |
| Stale signal (past freshness) | — | — | Treat as historical reference only |
| Conflicting signals across engines | Defer to higher-TF, higher-conviction | — | Stand aside — cleanest setups self-announce |
Closing principle: the engines surface candidates. The trader supplies context. A pattern at the right level with the right phase is a setup. A pattern by itself is decoration.
Connect the Dots
Test Your Understanding
4 questions — instant feedback, no scoring stored.