RayMelius/stockex-nn-agent

StockEx Transformer Trading Agent

Advanced neural network trading agent for the StockEx Clearing House.

Architecture

• Feature Extractor: Temporal Transformer Encoder (3L, 4H, d=128)
• Portfolio Encoder: 2-layer MLP (8 -> 64)
• Actor-Critic: MLP heads (192 -> 128 -> 64)
• Algorithm: PPO (Proximal Policy Optimization)
• Training: 500,000 timesteps on real ATHEX/NYSE/Euronext data

Observation Space

3,008-dimensional vector:

• Market state (3,000): 60 OHLCV bars x 50 technical indicators
• Portfolio state (8): capital, holdings, net worth, returns, ratios

Action Space

2-dimensional continuous:

• action[0]: 0=Hold, 1=Buy, 2=Sell
• action[1]: Position size (5%-50% of capital/holdings)

Usage

Set CH_RL_MODEL_REPO=RayMelius/stockex-nn-agent in your StockEx deployment.

Files

• final_model.zip: PPO model (stable-baselines3 format)
• scaler.pkl: StandardScaler for observation normalization

Performance

• Avg PnL: -1,204.17 EUR
• Win Rate: 40%
• Avg Trades: 63 securities/episode