Trading Gates & Pattern Filter - Research Notes

Experiment Overview

Problem Statement

The live trader was showing "Pattern filter rejected - no high-win-rate pattern match" for all symbols, even when the dashboard showed valid BUY/SELL signals. Investigation revealed two issues:

1. Missing predator/prey values: getpredprey_validation() wasn't returning predator/prey values
2. Misaligned thresholds: Pattern filter used absolute thresholds but Lotka-Volterra dynamics decay values to near-zero

Failed Attempts

Solution: Normalized Dominance Ratios

The key insight is that while raw predator/prey values can be any magnitude, their ratio indicates regime dominance:

# In filter_by_patterns() - alpaca_trading/signals/pattern_filter.py

# Extract raw values
raw_prey = regime_context.get('prey_strength', 0.5)
raw_predator = regime_context.get('predator_strength', 0.5)

# Normalize to dominance ratios (0-1 range)
total_strength = raw_predator + raw_prey + 1e-9  # Avoid division by zero
prey_strength = raw_prey / total_strength  # Higher = more mean-reverting
predator_strength = raw_predator / total_strength  # Higher = more trending

Example

• Raw values: predator=0.015, prey=0.17
• Normalized: predatordominance=8%, preydominance=92%
• Interpretation: Strong mean-reverting regime (VWAP reversion pattern applies)

Trading Gates Architecture

Orders must pass ALL gates before execution:

Pattern Filter Requirements

VWAP Reversion (68% win rate):

• prey_dominance >= 0.4 (mean-reverting regime)
• Price 2+ standard deviations from VWAP
• Signal direction matches expected reversion

Momentum Continuation (71% win rate):

• predator_dominance >= 0.6 (trending regime)
• trendprobup > 0.7 OR trendprobdown > 0.7
• Pullback in trend direction
• Signal aligns with trend

Dashboard Gate Display

Added gate status indicators to plotsymbolsignals():

# In dashboard.py - alpaca_trading/visualization/dashboard.py

gate_statuses = symbol_signals.get('gate_statuses', [])
for i, sym in enumerate(symbols):
    gate = gate_statuses[i]
    status = gate.get('final_status', 'UNKNOWN')

    if status == 'READY':
        status_color = 'green'
        status_text = 'READY'
    elif status == 'BLOCKED':
        status_color = 'red'
        status_text = gate.get('block_reason', 'blocked')[:12]
    elif status == 'HOLD':
        status_color = 'gray'
        status_text = 'HOLD'

Key Files Modified

Verification

After fix:

• AAPL: PASS (vwapreversion) - preydominance=92%
• AMD: PASS (vwapreversion) - preydominance=91.9%
• AVGO: FAIL (no_match) - SELL signal in prey-dominant regime (expects BUY)
• BTCUSD: FAIL (cryptonoshort) - Alpaca doesn't support crypto shorts

Key Learnings

1. Lotka-Volterra dynamics decay: Both predator and prey values decay to near-zero over time; use ratios, not absolutes
2. Crypto limitations: Alpaca doesn't support short selling for crypto; must block SELL signals for crypto entries
3. Fractional shares: Executor must use float qty, not int, for proper fractional share support
4. Dashboard visibility: Showing gate status helps debug why signals don't result in orders

Related Skills

• position-reconciliation: Broker state sync
• markov-regime-features: Debugging constant Markov features
• drawdown-guardrails-pattern: Drawdown control across systems