zkarbie/pim-layer2-rl-agents-MeanReversion_qdrant

MeanReversion_qdrant - Layer 2 RL Agent

Part of the PassiveIncomeMaximizer (PIM) trading system.

Model Description

Layer 2 RL agents for signal filtering (PPO-trained)

This is a Proximal Policy Optimization (PPO) reinforcement learning agent trained to filter trading signals from FinColl predictions. The agent evaluates prediction confidence and signal quality based on meanreversion criteria.

Architecture

• —Algorithm: Proximal Policy Optimization (PPO)
• —Input: 414-dimensional SymVectors from FinColl
• —Output: Confidence score (0-1) and action recommendation
• —Training: Trained on historical market data with profit-based rewards
• —Framework: PyTorch with custom RL implementation

Layer 2 System

PIM uses 9 Layer 2 RL agents that collaborate to filter predictions:

1. 1.MomentumAgent - Price momentum patterns
2. 2.TechnicalAgent - Chart patterns and indicators
3. 3.RiskAgent - Volatility and drawdown assessment
4. 4.OptionsAgent - Options flow analysis
5. 5.MacroAgent - Economic indicators
6. 6.SentimentAgent - News and social sentiment
7. 7.VolumeAgent - Trading volume patterns
8. 8.SectorRotationAgent - Sector strength
9. 9.MeanReversionAgent - Overbought/oversold detection

Usage

import torch
from pim.learning.agents.layer2_mlp import Layer2MLPAgents

# Load model
agents = Layer2MLPAgents(device='cuda')
agents.load_trained_agents('path/to/trained_agents/')

# Evaluate a SymVector
import numpy as np
symvector = np.random.rand(414)  # 414D feature vector from FinColl
scores = agents.evaluate(symvector)  # Returns dict of agent scores

# Aggregate scores
composite, confidence = agents.aggregate_scores(scores)
print(f"Composite score: {composite:.3f}, Confidence: {confidence}")

Training Data

• —Period: 2024 historical equity data (35,084 SymVectors)
• —Symbols: 332 equities from diversified portfolio
• —Features: 414-dimensional vectors (price, sentiment, fundamentals, technical indicators)
• —Source: FinColl API with TradeStation market data

Performance Metrics

Based on January 2024 backtests:

• —Directional Accuracy: 71.88% (10-day horizon)
• —Sharpe Ratio: 7.24 (annualized)
• —Profit Factor: 3.45
• —Win Rate: 71.9%

Limitations

• —Trained on 2024 equity data only (not tested on other asset classes)
• —Requires FinColl SymVectors (414D) as input
• —Performance may degrade in unprecedented market conditions
• —Best used as part of complete PIM dual-layer system

Intended Use

This model is intended for:

• —Signal filtering in automated trading systems
• —Research into RL-based trading strategies
• —Educational purposes in quantitative finance

Not intended for:

• —Standalone trading decisions (use full PIM system)
• —Financial advice or recommendations
• —Unmonitored autonomous trading

Citation

@software{pim_layer2_meanreversion,
  author = {PassiveIncomeMaximizer Team},
  title = {MeanReversion_qdrant - Layer 2 RL Agent},
  year = {2025},
  url = {https://github.com/yourusername/PassiveIncomeMaximizer}
}

More Information

• —Repository: https://github.com/yourusername/PassiveIncomeMaximizer
• —Documentation: See LAYER2_README.md in docs/architecture/layer2/
• —License: MIT