feat: quantitative learning engine — AI learns from its own trade numbers

New module: skopaq/learning/tracker.py
- Signal accuracy tracking per symbol (win rate, avg P&L)
- Confidence calibration (AI overconfident/underconfident detection)
- Sector performance breakdown (IT, banking, energy, etc.)
- Regime performance (high/low VIX, trending/ranging)
- Time-of-day entry patterns (best hours for trades)
- Stop-loss effectiveness analysis (too tight? too loose?)
- Holding period optimization per symbol
- generate_learning_insights() — comprehensive AI self-assessment

New MCP tools (36 total):
- get_learning_insights: Full learning report with all patterns
- get_symbol_stats: Per-symbol AI performance stats

Database: signal_records table (Fly.io Postgres)
Docs: architecture/learning.md added

The AI now knows its own strengths, weaknesses, and blind spots.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: samuelvinay91

docs/architecture/learning.md

@@ -0,0 +1,46 @@
+# Quantitative Learning Engine
+
+The learning engine tracks every AI prediction and its outcome, building a data-driven understanding of what works and what doesn't.
+
+## What It Tracks
+
+| Metric | Question It Answers |
+|--------|-------------------|
+| **Symbol Accuracy** | Which stocks is the AI best/worst at? |
+| **Confidence Calibration** | When AI says 75%, does it win 75%? |
+| **Sector Performance** | Banking vs IT vs Energy — where's the edge? |
+| **Regime Performance** | High VIX vs low VIX — different strategies? |
+| **Timing Patterns** | Morning entries vs afternoon — which works? |
+| **Stop-Loss Analysis** | Are stops too tight (hit before recovery)? |
+| **Holding Period** | Optimal days to hold per stock? |
+
+## How It Works
+
+```
+Trade Executed → Record: symbol, confidence, entry, sector, regime, hour
+        ↓
+Trade Closed → Update: exit price, P&L, won/lost, holding days
+        ↓
+Learning Engine → Analyze patterns across all trades
+        ↓
+Generate Insights → "OVERCONFIDENT on IT stocks", "Stops too tight on RELIANCE"
+        ↓
+Inject into Agent → Next analysis uses these insights
+```
+
+## MCP Tools
+
+- `get_learning_insights` — Full learning report with all patterns
+- `get_symbol_stats RELIANCE` — Per-symbol performance stats
+
+## Database Tables
+
+- `signal_records` — Every signal with outcome (Fly.io Postgres)
+
+## Self-Correction
+
+The learning engine enables **self-correction**:
+
+- If AI is overconfident (says 80% but wins 50%), it learns to lower confidence
+- If stops are hit 60% of the time, it widens stops automatically
+- If a sector consistently loses, the AI deprioritizes it in scans

skopaq/learning/__init__.py

@@ -0,0 +1,5 @@
+"""Quantitative Learning Engine — the AI learns from its own trading numbers.
+
+Tracks signal accuracy, regime performance, confidence calibration,
+sector patterns, and timing — all fed back into the analysis pipeline.
+"""