zkarbie/pim-account-manager-backtest_update_2210

backtest_update_2210 - Meta-Learner Neural Network

Part of the PassiveIncomeMaximizer (PIM) trading system - Layer 1 Learning.

Model Description

Meta-learner neural network for agent weight optimization

This neural network learns optimal weights for 9 LLM agents in the PIM committee decision system. It updates after every trade based on actual profit/loss, continuously improving agent coordination.

Architecture

• Input: 27-dimensional feature vector (9 agents × 3 features each)
  • Agent confidence score
  • Agent recommendation strength
  • Agent historical accuracy
• Hidden Layers:
  • Layer 1: 27 → 64 (ReLU)
  • Layer 2: 64 → 32 (ReLU)
• Output: 1 (predicted profit/loss)
• Framework: PyTorch
• Loss Function: Mean Squared Error (MSE)
• Optimizer: Adam (lr=0.001)

Layer 1 Meta-Learning

The meta-learner is part of PIM's dual-layer learning system:

Layer 1 (Meta-Learner):

• Learns agent weights dynamically
• Updates every trade (neural network backpropagation)
• Optimizes committee decision quality
• Stores learned weights in database

Layer 2 (RL Agents):

• Filters predictions by confidence
• Trains every 64 trades (PPO updates)
• Provides high-quality signals to Layer 1

Usage

import torch
from pim.learning.meta_learner_feedback import MetaLearnerNetwork

# Load checkpoint
checkpoint = torch.load('backtest_update_2210.pt')
model = MetaLearnerNetwork(input_dim=27, hidden_dim=64, hidden_dim2=32)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Predict expected profit from agent recommendations
with torch.no_grad():
    agent_features = torch.tensor([...])  # 27D feature vector
    expected_profit = model(agent_features)
    print(f"Expected profit: ${expected_profit.item():.2f}")

Training Process

1. During Backtest: Committee makes decision based on current agent weights
2. Trade Executes: Position opened, monitoring begins
3. Trade Closes: Calculate actual profit/loss
4. Meta-Learner Update:
   • Convert agent recommendations → 27D feature vector
   • Forward pass: Predict expected profit
   • Calculate loss: MSE(predicted, actual)
   • Backpropagate: Update network weights
   • Save checkpoint every 10 updates
5. Next Trade: Use improved agent weights

Training Data

• Period: Backtest period with real market data
• Examples: Every closed trade becomes a training example
• Features: Agent recommendations (confidence, strength, accuracy)
• Target: Actual profit/loss from trade
• Updates: Continuous learning (every trade)

Performance Metrics

The meta-learner enables the PIM system to achieve:

• Adaptive agent weights based on market conditions
• Improved decision quality over time
• Self-optimizing committee dynamics
• Robust performance across diverse market regimes

Checkpoint Information

This checkpoint represents update #2210 from a production backtest. The model has learned from hundreds or thousands of trades.

Limitations

• Requires complete PIM agent system (9 LLM agents)
• Trained on specific agent feature format (27D)
• Performance depends on quality of Layer 2 signal filtering
• Best used within complete dual-layer PIM architecture

Intended Use

This model is intended for:

• Agent weight optimization in multi-agent trading systems
• Research into meta-learning for committee decisions
• Educational purposes in adaptive AI systems

Not intended for:

• Standalone trading decisions
• Use outside PIM ecosystem
• Financial advice or recommendations

Citation

@software{pim_meta_learner,
  author = {PassiveIncomeMaximizer Team},
  title = {Meta-Learner Neural Network for Agent Weight Optimization},
  year = {2025},
  url = {https://github.com/yourusername/PassiveIncomeMaximizer}
}

More Information

• Repository: https://github.com/yourusername/PassiveIncomeMaximizer
• Documentation: See learning/meta_learner_feedback.py and learning_backtest.py
• Paper: "Dual-Layer Learning for Autonomous Trading Systems" (in preparation)
• License: MIT