Adaptive Predator-Prey Dynamics (v3.3.0)

Overview

Item	Details
Date	2025-01-21
Goal	Make Lotka-Volterra predator-prey coefficients adaptive to market conditions
File	`alpaca_trading/signals/predator_prey.py`
Status	Success

Context

The predator-prey Markov model uses Lotka-Volterra dynamics to detect regime transitions:

•Predator (P): Represents trend followers / momentum traders
•Prey (N): Represents value / mean-reversion traders

Previously, coefficients were fixed:

•a = 0.8 (predator growth rate)
•b = 0.05 (interaction rate)
•Cooldown was binary (0 or 1)

This didn't adapt to changing market conditions.

Solution: Adaptive Coefficients

Predator Coefficient (a) - Adapts to Volatility

python

# Higher volatility = faster predator growth (momentum strategies thrive)
# vol_ratio = current_vol / baseline_vol
a = 0.8 + 0.3 * max(0, vol_ratio_local - 1.0)
a = min(1.1, a)  # Cap at 1.1

# Effect:
# - vol_ratio = 1.0: a = 0.8 (baseline)
# - vol_ratio = 1.5: a = 0.95 (faster growth)
# - vol_ratio = 2.0+: a = 1.1 (capped)

Interaction Coefficient (b) - Adapts to Momentum

python

# Stronger momentum = stronger interaction (more decisive regime shifts)
# momentum_strength = abs(returns[-5:].mean())
b = 0.05 + 0.1 * min(0.3, momentum_strength)

# Effect:
# - momentum = 0: b = 0.05 (baseline)
# - momentum = 0.15: b = 0.065 (moderate)
# - momentum = 0.30+: b = 0.08 (strong interaction)

Exponential Decay Cooldown

python

# OLD: Binary gate (either 0% or 100% allowed)
if bars_since_flat < cooldown_bars:
    entry_allowed = 0
else:
    entry_allowed = 1

# NEW: Exponential decay (gradual unlock)
cooldown_factor = min(1.0, bars_since_flat / cooldown_bars)
effective_edge = edge * cooldown_factor

# Effect:
# - Just exited (0 bars): cooldown_factor = 0.0
# - Halfway (cooldown/2): cooldown_factor = 0.5
# - Full cooldown: cooldown_factor = 1.0

Code Location

python

# predator_prey.py - _compute_edge method (~line 240)
def _compute_edge(
    self,
    N: float,
    P: float,
    N_prime: float,
    P_prime: float,
    bars_since_flat: int,
    cooldown_bars: int,
    vol_ratio_local: float = 1.0,
    momentum_strength: float = 0.0,
) -> Tuple[float, int, float, float]:
    """Compute trading edge with adaptive dynamics."""

    # v3.3.0: Adaptive predator coefficient based on volatility
    a = 0.8 + 0.3 * max(0, vol_ratio_local - 1.0)
    a = min(1.1, a)

    # v3.3.0: Adaptive interaction coefficient based on momentum
    b = 0.05 + 0.1 * min(0.3, momentum_strength)

    # v3.3.0: Exponential decay cooldown (was binary gate)
    cooldown_factor = min(1.0, bars_since_flat / cooldown_bars)
    effective_edge = edge * cooldown_factor

Macro Regime Influence (v3.3.0)

Increased Influence Weight

python

# harmonized_gpu.py (~line 362)
# OLD: Weak influence (12%)
bull_trend_boost = p_bull * 0.12
bear_trend_boost = p_bear * 0.12

# NEW: Stronger influence (30%)
bull_trend_boost = p_bull * 0.30
bear_trend_boost = p_bear * 0.30

This makes macro regime (bull/sideways/bear) have 2.5x more influence on fine-scale chains.

Dynamic Volatility Gate

python

# OLD: Static threshold
vol_gate_threshold = 0.6

# NEW: Dynamic threshold based on current volatility
vol_gate_base = 0.3      # Minimum threshold
vol_gate_scale = 0.5     # Scaling factor
vol_gate_threshold = vol_gate_base + vol_gate_scale * (current_vol / vol_p90)

Key Insights

Why Adaptive Coefficients?

Fixed coefficients assume market conditions are constant. In reality:

•High volatility: Momentum strategies work better, predators should grow faster
•Strong momentum: Regime transitions are more decisive, interaction should be stronger
•After exits: Need time to reassess, but binary cooldown is too restrictive

Exponential vs Binary Cooldown

Binary cooldown: "No trading for N bars, then full trading"

•Problem: Misses opportunities near the end of cooldown
•Problem: No gradual re-entry based on confidence

Exponential decay: "Gradually increase position sizes as cooldown expires"

•Advantage: Can take partial positions during cooldown
•Advantage: Smooth transition from cautious to normal trading

Macro Influence Strength

The 4-chain hierarchy (macro → vol/trend/momentum) was too weakly coupled. Increasing macro influence from 12% to 30% makes:

•Bull macro regime → trend chain more likely to show "up"
•Bear macro regime → trend chain more likely to show "down"

Failed Attempts

Attempt	Why it Failed	Lesson Learned
Fixed a=0.8, b=0.05	Didn't adapt to changing conditions	Make coefficients functions of market state
Binary cooldown	Too restrictive, missed opportunities	Use exponential decay for gradual unlock
12% macro influence	Fine-scale chains operated nearly independently	Increase to 30% for stronger hierarchy
a coefficient unbounded	Could grow infinitely in high vol	Cap at reasonable max (1.1)

References

•Lotka-Volterra equations (predator-prey dynamics)
•Hierarchical Markov Models for regime detection
•arXiv 2508.01173 - MARS Meta-Adaptive Framework