Natron Transformer V2

Natron is a multi-phase GPU-first trading intelligence stack that learns market structure, supervised signals, and reinforcement policies end-to-end from OHLCV sequences. The system ships with a production-ready data pipeline, Transformer backbone, pretraining + fine-tuning loops, PPO reinforcement adapter, realtime inference bridge, monitoring utilities, and an MQL5 Expert Advisor.

Repository Layout

• configs/natron_config.yaml – master parameters for data, training, RL, and serving.
• dataset_loader.py – feature engineering, institutional labeling, sequence builders, and NatronDataModule.
• model_natron.py – Transformer encoder, multi-task heads, masked/contrastive blocks, and init helpers.
• losses.py – masked reconstruction, InfoNCE, and multi-task classification losses.
• train_natron.py – orchestrated pretraining, supervised fine-tuning, and optional PPO reinforcement.
• rl_trainer.py – trading environment wrapper and PPO trainer built on the Natron encoder.
• server_natron.py – Flask + TCP hybrid inference server that scores sequences and returns JSON signals.
• monitor_natron.py – lightweight system/GPU/health monitor for deployed servers.
• start_natron.sh – boot script that launches the server (and optional monitor) with a given config.
• natron_ea.mq5 – MetaTrader 5 Expert Advisor that streams candles to the socket bridge and manages trades.
• Dockerfile, requirements.txt – containerized GPU runtime for training and inference.
• docs/natron_v2_whitepaper.md – architecture, research rationale, and operating philosophy.

Quick Start

1. Environment Setup

   python3 -m venv .venv
   source .venv/bin/activate
   pip install --upgrade pip
   pip install torch==2.1.0 --extra-index-url https://download.pytorch.org/whl/cu121
   pip install -r requirements.txt

2. Prepare Data
   • Place data_export.csv under data/.
   • Update paths and data sections of configs/natron_config.yaml if required.
3. Train the Model

   python train_natron.py --config configs/natron_config.yaml

   • Phase 1: masked modelling + contrastive pretraining over engineered features.
   • Phase 2: supervised multi-task fine-tuning (buy / sell / direction / regime).
   • Phase 3 (optional): PPO reinforcement tuning (rl.enabled: true).
   • Checkpoints and the feature scaler are saved under model/.
4. Serve Inference

   bash start_natron.sh configs/natron_config.yaml

   • REST endpoint: POST /predict with the latest 96 OHLCV candles.
   • TCP bridge: newline-delimited JSON on the configured socket (server.socket).
   • Monitor logs stream CPU/GPU health while the service runs.
5. Connect MetaTrader
   • Load natron_ea.mq5 into the MetaTrader 5 terminal.
   • Set the EA inputs to match the socket host/port and thresholds.
   • Enable algo-trading; the EA will poll Natron and manage positions.

Training Pipeline

• Feature Matrix (100+ signals) – trend, momentum, volume, volatility, structure, smart money concept, market profile, and calendar embeddings.
• Label Suite – buy/sell heuristics, binary direction, and six-state market regime following institutional rules.
• Sequences – sliding windows of (96, 100) features aligned with next-step labels.
• Phase 1 (Pretraining) – random masking + InfoNCE on jittered/time-warped augmentations to learn latent structure.
• Phase 2 (Supervised) – multi-task heads with AdamW + ReduceLROnPlateau, encoder freezing warm-up, and early stopping.
• Phase 3 (RL) – PPO actor-critic on top of the shared encoder, optimizing profit minus turnover/drawdown penalties.

Deployment and Ops

• REST JSON Output

  {
    "buy_prob": 0.71,
    "sell_prob": 0.24,
    "direction_up": 0.69,
    "regime": "BULL_WEAK",
    "confidence": 0.82
  }

• Socket Protocol – Send a JSON payload with candles array terminated by \n; receive a single-line JSON response.
• Monitoring – monitor_natron.py polls /health, reports CPU/RAM/GPU usage, and can be run standalone or via start_natron.sh.
• Docker Build

  docker build -t natron:v2 .
  docker run --gpus all -p 8000:8000 -p 9000:9000 natron:v2

Extending Natron

• Update FeatureEngineer to add domain-specific indicators; the scaler automatically adapts via caching.
• Customize loss weights or schedules in the YAML config.
• Swap PPO for SAC by adding a new trainer module; the environment is self-contained.
• Integrate alternative brokers by mirroring the EA’s socket contract.

For a deeper explanation of the architecture, data philosophy, and experimentation roadmap, read docs/natron_v2_whitepaper.md.