/signal-postmortem

Feedback Integration Guide

Overview

This document explains how signal postmortem results integrate with downstream systems: edge-signal-aggregator weight calibration and the skill improvement backlog.

Feedback Targets

1. Edge-Signal-Aggregator Weight Calibration

Purpose: Adjust the weight of each contributing skill based on historical accuracy.

Integration Point: reports/weight_feedback_YYYY-MM-DD.json

Consumption:

1. edge-signal-aggregator reads feedback file on startup
2. Applies weight adjustments to skill contribution scores
3. Logs adjustment application for audit trail

Weight Adjustment Formula:

new_weight = current_weight * (1 + adjustment_factor)

adjustment_factor = (accuracy - baseline) * sensitivity

where:
- accuracy = true_positives / (true_positives + false_positives)
- baseline = 0.55 (expected random accuracy)
- sensitivity = 0.5 (dampening factor to avoid overreaction)

Constraints:

• Minimum weight: 0.3 (never fully disable a skill)
• Maximum weight: 2.0 (never over-amplify)
• Minimum sample size: 20 signals for adjustment
• Rolling window: 90 days of postmortem data

2. Skill Improvement Backlog

Purpose: Generate actionable improvement tasks for the skill improvement loop.

Integration Point: reports/skill_improvement_backlog.yaml

Consumption:

1. Skill improvement loop reads backlog entries
2. Prioritizes by severity and sample size
3. Creates improvement branches for high-severity issues

Issue Types:

Backlog Entry Format:

- skill: vcp-screener
  issue_type: false_positive_cluster
  severity: medium
  evidence:
    false_positive_rate: 0.18
    sample_size: 45
    regime_correlation: RISK_OFF
    sector_correlation: Technology
  suggested_action: "Add RISK_OFF regime filter or reduce confidence"
  priority_score: 72  # Calculated from severity * sample_size * impact
  generated_by: signal-postmortem
  generated_at: "2026-03-17T10:35:00Z"
  status: pending

Feedback Frequency

Real-Time (Per Signal)

• Postmortem record created immediately after trade closure
• Stored in reports/postmortems/

Daily Batch

• Weight feedback regenerated daily at 06:00
• Improvement backlog updated with new entries
• Old entries marked as addressed when improvements deployed

Weekly Review

• Summary statistics by skill, sector, regime
• Trend analysis (rolling 4-week accuracy)
• Human review flagged for significant changes

Data Flow Diagram

Signal Generated
       |
       v
Trade Executed (or skipped)
       |
       v
Holding Period Completes
       |
       v
postmortem_recorder.py
       |
       +---> postmortem record (JSON)
       |
       v
postmortem_analyzer.py
       |
       +---> weight_feedback.json --> edge-signal-aggregator
       |
       +---> skill_improvement_backlog.yaml --> skill improvement loop
       |
       +---> summary report (Markdown)

Idempotency and Deduplication

Postmortem Records

• postmortem_id = pm_ + signal_id
• If postmortem already exists, update instead of duplicate
• Version field tracks updates

Weight Feedback

• Regenerated fresh each run (not cumulative)
• Based on rolling 90-day window
• Old feedback files archived to reports/archive/

Backlog Entries

• Keyed by skill + issue_type + month
• New evidence updates existing entry instead of creating duplicate
• Entry moved to addressed when skill version changes

Minimum Thresholds

To avoid noisy feedback from small samples:

Manual Override Integration

When human review identifies an issue not caught by automated analysis:

# Manual backlog entry
- skill: earnings-trade-analyzer
  issue_type: manual_review
  severity: high
  evidence:
    description: "Systematic gap fade failure in biotech earnings"
    reviewer: "tradermonty"
  suggested_action: "Exclude biotech from gap fade signals"
  generated_by: human_review
  generated_at: "2026-03-17T14:00:00Z"

Manual entries have generated_by: human_review and are prioritized higher.

Conflict Resolution

When automated and manual feedback conflict:

1. Manual feedback takes precedence for immediate action
2. Automated feedback triggers investigation
3. Resolution documented in backlog entry

Audit Trail

All feedback actions are logged:

{
  "action": "weight_adjustment_applied",
  "skill": "vcp-screener",
  "old_weight": 1.0,
  "new_weight": 0.85,
  "reason": "15% FP rate in RISK_OFF",
  "applied_at": "2026-03-17T06:00:00Z",
  "applied_by": "edge-signal-aggregator"
}

Logs stored in logs/feedback_audit.log with 90-day retention.