Tyche — Crypto HFT Research System

1-second bars · PyTorch PPO · MongoDB · Streamlit

What the dashboard shows

Panel	Refresh
P&L ($10,000 → $X)	every 1 second
Live trade feed (BUY/SELL + exact $ P&L)	every 1 second
Live prices (5 assets)	every 1 second
MongoDB trade count	every 5 seconds
Episode P&L chart	every 10 seconds
Adversary mode (FLASH CRASH etc.)	every 1 second

Architecture

Binance S3 (1-second OHLCV)
        ↓
  data_loader.py  →  tyche/environment.py  (HFTEnv)
                              ↓
                   tyche/agent_gpu.py  (PPO, RTX 4070)
                              ↓
                   tyche/trainer.py  (episode loop, persistent weights)
                              ↓
          ┌────────────────────────────────────┐
          │  memory/trainer_status.json        │
          │  memory/recent_trades.json         │
          │  memory/agent_weights.pt           │
          │  MongoDB (trades, episodes)        │
          └────────────────────────────────────┘
                              ↓
                         server.py  (Flask :5000)
                              ↓
                      dashboard_app.py  (Streamlit :8501)

File map

tyche/
├── setup.bat               one-click install
├── train_overnight.bat     overnight training
├── start.bat               launch dashboard
├── download_data.py        Binance S3 1-second downloader
├── generate_weights.py     bootstrap agent weights
├── main.py                 training entry point
├── server.py               Flask REST API (:5000)
├── dashboard_app.py        Streamlit UI (:8501)
├── requirements.txt
├── memory/
│   ├── agent_weights.pt    latest weights (auto-saved)
│   ├── best_weights.pt     best episode weights
│   ├── training_log.json   episode history
│   ├── trainer_status.json live status (read by dashboard)
│   └── recent_trades.json  last 200 trades
└── tyche/
    ├── config.py
    ├── data_loader.py      loads parquet or synthetic GBM
    ├── environment.py      HFTEnv (OBS_DIM=35, N_ACTIONS=243)
    ├── agent_gpu.py        PyTorch actor-critic PPO
    ├── trainer.py          episode loop
    ├── adversarial.py      5-mode stress engine
    └── mongo_store.py      MongoDB + memory fallback

Key numbers

Parameter	Value
Starting capital	$10,000
Assets	BTC ETH BNB SOL ADA
Data granularity	1 second
Observation dim	35
Actions	243 (3^5 per-asset)
PPO clip ε	0.2
Learning rate	3×10⁻⁴
Trade fee	0.1% (Binance taker)
Execution latency	~2s mean