Signal Aggregator

5-10 signals is typical. Too few (2-3) doesn't provide enough confirmation. Too many (15+) creates noise and complexity. Choose one or two from each category: trend, momentum, volume, volatility, sentiment.

You can, but choose wisely. Best combinations use independent indicators from different categories. Avoid multiple similar indicators (like 3 moving averages) - they're redundant. Aim for indicators that measure different things.

Start with 60-70%. Higher thresholds (80%+) give fewer but higher-quality signals. Lower thresholds (50-60%) give more signals but lower quality. Adjust based on your experience and risk tolerance.

Frequent conflicts often mean: (1) You're using correlated signals that should agree, or (2) Market is genuinely uncertain. Review your signal selection for independence. Accept that some market periods have no clear direction.

You can, but reduce position size. Full consensus = full size. 70% consensus = 70% size. This manages risk while still participating. Never go full size on weak consensus.

Methods: (1) Equal weights - simplest. (2) Performance-based - weight by historical accuracy. (3) Category-based - allocate across categories. (4) Regime-based - different weights for different conditions. Start with equal, then optimize based on backtest.

Convert all to common scale (e.g., -100 to +100). For RSI: map 30-70 to -100 to +100. For MACD: divide by ATR or price to standardize. For bounded indicators: linear scaling. For unbounded: z-score normalization.

Calculate correlation between signal outputs over time. Correlation > 0.7 = redundant. Use correlation matrix to see all pairings. Cluster correlated signals and use one representative per cluster.

Both! Leading indicators (sentiment, breadth) give early warning. Coincident indicators (price-based) confirm current state. Lagging indicators (moving averages) confirm trend. Combine all three for best picture.

Review quarterly. Track signal performance continuously. Make small adjustments (10-20% weight changes) rather than dramatic shifts. Major recalibration only when market regime clearly changes or signal consistently fails.

Start with prior probability (e.g., 50% bullish). For each signal, calculate likelihood ratio: P(Signal|Bull)/P(Signal|Bear). Update posterior using Bayes: P(Bull|Signal) = Likelihood × Prior / Normalizer. Chain updates through all signals.

Train ML model where features are individual signals and target is future return. Methods: Random Forest (handles non-linearity), Gradient Boosting (learns optimal weighting), Neural Networks (complex patterns). Use proper train/test split and walk-forward validation.

Maintain regime detection indicators (ADX for trend strength, VIX for volatility). Define regime thresholds. When regime changes detected, smoothly transition weights over several days. Don't make abrupt changes. Log regime transitions.

Ensemble diversity measures how different your signals are. Low diversity (correlated signals) = little benefit from combining. High diversity = aggregation reduces error. Maximize diversity by choosing signals that are wrong at different times.

Architecture: Data ingestion → Signal calculation → Normalization → Weighting → Aggregation → Output. Key: real-time updates, error handling, monitoring, logging. Store historical signals for analysis. Version control weights. Alert on anomalies.