/edge-signal-aggregator

Signal Weighting Framework

This document explains the rationale behind default signal weights, the scoring methodology, and guidance for customizing weights based on your trading style.

Default Skill Weights

Total: 1.00

Weight Rationale

edge-candidate-agent (0.25)

Highest weight because:

• Quantitative, data-driven anomaly detection
• Produces actionable tickets with specific entry/exit levels
• Least susceptible to narrative bias
• Signals are time-stamped and verifiable

edge-concept-synthesizer (0.20)

Second highest because:

• Integrates multiple upstream inputs
• Produces structured edge concepts with testable hypotheses
• Requires corroboration from multiple sources

theme-detector, sector-analyst, institutional-flow-tracker (0.15 each)

Equal moderate weights because:

• Each provides a distinct analytical lens
• Themes capture narrative momentum
• Sectors capture rotational flows
• Institutional flow captures smart money
• None is inherently more reliable than others

edge-hint-extractor (0.10)

Lowest weight because:

• Hints are suggestive, not definitive
• Often require validation from other skills
• Useful for idea generation, less for conviction

Composite Score Calculation

The composite conviction score for an aggregated signal is calculated as:

base_score = Σ(skill_weight × normalized_score) / Σ(skill_weight)
composite  = min(max_score, (base_score + agreement_bonus + merge_bonus) × recency_factor)

Normalization

Raw scores from different skills are normalized to [0, 1]:

• For 0-1 scale inputs (value <= 1.0): used as-is
• For 0-100 scale inputs (value <= 100.0): divided by 100
• For categorical grades: A=1.0, B=0.8, C=0.6, D=0.4, F=0.2
• Missing values: 0.0 (no contribution)

Agreement Bonus (additive)

When multiple skills agree on the same signal (after dedup merge):

• 2 skills agree: +0.10 added to base_score
• 3+ skills agree: +0.20 added to base_score

Merge Bonus (additive)

When duplicates are merged, each merged duplicate adds +0.05 to base_score.

Recency Factor (multiplicative)

Applied as a multiplier to the combined score:

• Within 24 hours: ×1.00
• 1-3 days old: ×0.95
• 3-7 days old: ×0.90
• 7+ days old: ×0.85

Final composite is capped at 1.0.

Deduplication Logic

Similarity Detection

Two signals are considered duplicates if direction matches AND either condition is met:

1. Ticker overlap >= 30% -- Jaccard overlap of ticker sets (default 0.30)
2. Title similarity >= 0.60 -- Word-based Jaccard similarity of titles (default 0.60)

Note: OR logic is used -- a high ticker overlap alone or a high title similarity alone is sufficient.

Merge Strategy

When duplicates are detected:

1. Keep the signal with the highest raw score as primary
2. Aggregate contributing skills from all duplicates
3. Boost composite score by 5% per merged duplicate (indicates consensus)
4. Log all merged signals for audit trail

Contradiction Detection

Definition

A contradiction exists when:

1. Same ticker or sector appears in multiple signals
2. Directions are opposite (LONG vs SHORT)
3. Time horizons overlap

Severity Levels

Resolution Hints

The aggregator provides resolution hints:

• Timeframe mismatch -- Check if signals apply to different horizons
• Sector vs stock -- Sector bearish but specific stock bullish within sector
• Flow vs price -- Institutional buying but price declining (accumulation phase?)

Customizing Weights

For Momentum Traders

Increase weights for skills that capture short-term moves:

weights:
  edge_candidate_agent: 0.30
  theme_detector: 0.25
  sector_analyst: 0.20
  edge_concept_synthesizer: 0.15
  institutional_flow_tracker: 0.05
  edge_hint_extractor: 0.05

For Value/Position Traders

Increase weights for skills that capture longer-term positioning:

weights:
  institutional_flow_tracker: 0.30
  edge_concept_synthesizer: 0.25
  edge_candidate_agent: 0.20
  sector_analyst: 0.15
  theme_detector: 0.05
  edge_hint_extractor: 0.05

For Thematic Investors

Increase weights for narrative and theme detection:

weights:
  theme_detector: 0.30
  edge_concept_synthesizer: 0.25
  sector_analyst: 0.20
  edge_candidate_agent: 0.15
  institutional_flow_tracker: 0.05
  edge_hint_extractor: 0.05

Quality Indicators

Signal Confidence Factors

Each aggregated signal includes a confidence breakdown:

Minimum Thresholds

Recommended minimum conviction thresholds:

Limitations

1. Garbage In, Garbage Out -- Aggregation quality depends on upstream skill quality
2. Weight Sensitivity -- Small weight changes can shift rankings significantly
3. No Fundamental Override -- The aggregator doesn't validate fundamental thesis
4. Temporal Lag -- Some skills (institutional flow) have inherent reporting delays

Best Practices

1. Regular Weight Tuning -- Review and adjust weights quarterly based on backtested performance
2. Contradiction Review -- Always manually review HIGH severity contradictions
3. Provenance Audit -- Periodically trace high-conviction signals back to source data
4. Diverse Inputs -- Run at least 3 upstream skills before aggregating for meaningful consensus