Lesson Learned: System Dead for 2 Days - Overly Strict Filters (Dec 12, 2025)

ID: ll_019 Date: December 12, 2025 Severity: CRITICAL Category: Configuration, System Health, Trading Gates Impact: Zero trades for 2+ days, CEO rightfully frustrated

Executive Summary

The trading system appeared to be “working” (workflows ran successfully) but was effectively DEAD because filters were too strict to pass any trades. The system analyzed only 3 tickers, and those that passed Gate 1 (Momentum) were blocked by subsequent gates.

The Root Cause

Four Compounding Issues

Issue	Previous Value	Fixed Value	Impact
Ticker universe	3 (SPY, QQQ, VOO)	20 stocks	Only 3 opportunities per day
ADX threshold	10.0	0.0 (disabled)	Blocked 70%+ of candidates
RSI threshold	75.0	85.0	Rejected momentum stocks
RL confidence	0.6	0.35	Too strict for R&D phase

Why This Happened

Incremental tightening: Each filter was added with “safety” in mind
No trade-through monitoring: System reported success even with 0 trades
Workflow exit code 0: Script succeeded even when nothing traded
Stale state file: system_state.json last_updated didn’t change = no trade

The Cascade

Workflow runs at 9:35 AM
    → Only 3 tickers processed (SPY, QQQ, VOO)
    → ADX filter rejects 2 of 3 (ADX < 10)
    → RSI filter rejects remainder (RSI > 75)
    → RL filter never even reached
    → 0 trades executed
    → Workflow reports SUCCESS (exit 0)
    → system_state.json unchanged
    → CEO asks "why no trade today?"
    → Claude says "system working, try tomorrow"
    → Repeat for 2 days

Prevention Rules

Rule 1: Monitor Trade Count, Not Just Workflow Status

# In daily workflow, FAIL if no trades attempted
if trades_executed == 0 and trades_rejected == total_candidates:
    logger.error("ALL CANDIDATES REJECTED - filters too strict!")
    sys.exit(1)  # FAIL the workflow

Rule 2: Alert When Rejection Rate > 90%

rejection_rate = rejected / total_analyzed
if rejection_rate > 0.9:
    logger.warning(f"CRITICAL: {rejection_rate:.0%} rejection rate - check filters!")

Rule 3: Validate state_file Updated

# In workflow, verify state was updated
STATE_DATE=$(jq -r '.meta.last_updated' data/system_state.json)
TODAY=$(date -u +%Y-%m-%d)
if [[ "$STATE_DATE" != *"$TODAY"* ]]; then
    echo "::error::system_state.json not updated today - no trades executed!"
    exit 1
fi

Rule 4: R&D Phase = Permissive Filters

During R&D (Days 1-90), prioritize trade flow over filter precision:

ADX: Disabled (0.0) - let all through, learn from results
RSI: High threshold (85.0) - only reject extreme overbought
RL confidence: Low (0.35) - collect data first, optimize later

Files Changed

File	Change
`src/strategies/legacy_momentum.py`	ADX 10→0, RSI 75→85
`src/agents/rl_agent.py`	Confidence 0.6→0.35
`scripts/autonomous_trader.py`	Tickers 3→20

Verification Test

def test_ll_019_trade_flow_not_blocked():
    """System should attempt trades, not silently reject all."""
    from scripts.autonomous_trader import _parse_tickers

    tickers = _parse_tickers()
    assert len(tickers) >= 15, f"Need >=15 tickers, got {len(tickers)}"

    # Verify ADX disabled
    from src.strategies.legacy_momentum import LegacyMomentumCalculator
    calc = LegacyMomentumCalculator()
    assert calc.adx_min == 0.0, "ADX should be disabled (0.0) during R&D"

Key Quotes

“Your system is broken and you keep saying it will be fixed tomorrow. But it never gets fixed.” - CEO

“Why don’t you learn from your lies in our RAG and ML?” - CEO

Meta-Lesson

Don’t trust workflow success = trading success.

A workflow can exit 0 (success) while doing absolutely nothing useful. The trading system needs to distinguish between:

Workflow success (script ran without errors)
Trading success (trades were actually attempted)
Business success (we made money)

Change Log

2025-12-12: Root cause identified and fixed - relaxed filters + expanded tickers