Gate 1: Composite Threshold
Weighted composite score must be above +0.10 for BUY or below -0.10 for SELL.
Methodology and Transparency
This page documents exactly how Q Signals computes outputs so users can interpret results with context, not blind trust.
Signal Decision Framework
Gate 2: Vote Consensus
At least 60% of active modules must agree on direction, reducing one-module false positives.
Confidence (0–100%)
Confidence measures signal decisiveness, not probability of profit. It combines two factors:
- Threshold distance: How far the composite score exceeds ±0.10 threshold (normalized 0–1)
- Module agreement: What fraction of modules voted in the same direction vs the 60% minimum
confidence = threshold_distance × (0.5 + 0.5 × agreement_ratio)
0–15% = near noise · 15–40% = low conviction · 40–70% = moderate · 70–100% = high conviction. A 5% confidence signal barely cleared the gates — treat it as borderline.
Runtime and Fallback Transparency
DQN Runtime Mode
Primary mode is a Deep Q-Network (PyTorch) with 10 core signal inputs and BUY/HOLD/SELL Q-value output.
Fallback Mode
If DQN runtime dependencies are unavailable, tabular fallback mode is used so analysis remains available.
Operational Principle
Fallback behavior prioritizes continuity over perfect feature parity. Endpoints may return simplified or partial results under degraded conditions.
What the 41 Modules Measure
| Module | Primary Input |
|---|---|
| Fibonacci | Retracement and extension structure from recent price swings |
| Wave Finder | Wave-pattern structure and trend phase |
| Swing Strategy | SMA crossover, RSI, Bollinger Band context |
| Sentiment / Social | News and social tone direction |
| Insider / Institutional / Congress | Ownership and disclosed transaction behavior |
| Supply Chain / Competitor | Peer stress and dependency signals |
| Volume + Multi-Timeframe | Volume confirmation and cross-timeframe trend alignment |
| Sector / Correlation / VIX Regime | Macro regime and cross-asset flow context |
| Economic / Analyst | Macro indicator direction and analyst consensus |
| Price Trend / Volatility / 52w / Moving Avg | Core technical direction and overextension |
| Candle Patterns (Lao-Tzu) | 61 TA-Lib candlestick patterns with confidence scoring (pattern strength, volume, breakout potential, 52w context) |
| RL Agent (DQN) | Learned BUY/HOLD/SELL Q-values from historical outcomes |
Targets, Stops, and Holding Period
Targets and stop-loss values are ATR-scaled estimates, not guarantees. Every signal includes conservative, moderate, and aggressive targets plus a suggested hold duration that expands or contracts with volatility and selected horizon.
RL Training Stages
| Steps | Stage | Interpretation |
|---|---|---|
| 0-500 | Infant | Mostly exploratory behavior |
| 500-5,000 | Learning | Early pattern formation |
| 5,000-25,000 | Developing | Improving but still unstable by regime |
| 25,000-100,000 | Intermediate | Stronger pattern recognition and consistency |
| 100,000+ | Experienced | High training volume, still requires ongoing validation |
How KPI Pages Are Calculated
| Metric | Definition | What "Good" Looks Like |
|---|---|---|
| Win Rate | Evaluated signals where at least one target (T1/T2/T3) was hit before stop-loss or expiration, divided by total evaluated signals. | Above 55% with positive expectancy |
| Expectancy | Average return (%) per evaluated signal. Each signal's return is clipped to ±200% to prevent outlier distortion. | Positive (higher is better) |
| Profit Factor | Sum of all winning returns ÷ sum of all losing returns (absolute). Capped at 99.99 for display. | Above 1.5; above 2.0 is strong |
| Alpha vs B&H | Signal return minus buy-and-hold return over the same period. Positive = signals outperformed passive holding. | Positive |
| Confidence | Signal decisiveness (0–100%): how far composite score exceeds threshold × module agreement strength. Not a probability of profit. | Above 40% for moderate conviction |
KPIs are computed from evaluated outcomes in signal history. Win rate, expectancy, and alpha are retrospective statistics and can change as more outcomes settle. Leaderboards are grouped by horizon and market regime over a rolling lookback window.
Signal Outcome Evaluation Rules
| Check | Rule |
|---|---|
| Targets | Each directional signal tracks three targets (T1/T2/T3). If any target is hit before deadline, outcome is WIN (with level recorded). |
| Stop | If stop-loss is breached before any target, outcome is LOSS (Stopped). |
| Expiration | If deadline passes without target/stop hit, outcome is LOSS (Expired). |
| Pending Window | Signals remain pending until enough time has passed for evaluation under selected horizon. |
Interpretation Guidance
Highest-conviction setups generally have: threshold-clearing composite, strong module agreement, acceptable risk/reward, and RL agreement. Even then, all results remain probabilistic. Treat outputs as structured research evidence, not guaranteed outcomes.
Markets are non-stationary. Methodology transparency improves interpretation but does not eliminate risk, uncertainty, or model drift.