RL Trader

Multi-component research and live-trading sandbox for FX, crypto, and equities using reinforcement learning, genetic algorithms, and walk‑forward optimization.

Top-level layout

• forex-rl/ – Python trading research & execution stack (actor‑critic, GA, TCN policies, walk‑forward optimization, tools).
• local-broker/ – Go-based local broker that emulates an OANDA‑style REST/streaming surface for fast offline and paper trading.
• streamer/ – Market data ingestion and account state utilities (OANDA FX, Bitunix crypto) with QuestDB integration.
• bkup-forex-rl-YYYYMMDD/ – Snapshots of earlier forex-rl trees kept for reference.
• ref/ – Reference code and vendor SDKs used during experiments (e.g. Alpaca examples).

Python stack (forex-rl/)

Key pieces:

• Offline and live trading runners (e.g. online_trader.py, multi_instrument_box_trading.py, tpsl_trader.py).
• Actor‑critic and TCN research code under actor-critic/ plus GA stacks under ga-multi20/, ga-ness/, and ga-trailing-20/.
• Walk‑forward optimization framework under wfo/ (see forex-rl/wfo/README.md for detailed CLI docs).
• Feature engineering and hybrid models under hybrid-agnostic/ and the autoencoder scripts in supervised-ae/ and unsupervised-ae/.
• Tools for universe construction, account PnL, and OANDA/Alpaca connectivity under tools/.

There is intentionally no single pinned requirements.txt for the Python stack; dependencies are managed per experiment. A typical setup is:

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
# Core libraries (adjust as needed for a given script)
pip install numpy pandas scipy torch matplotlib oandapyV20 requests

Running tests (quick smoke over loaders, simulators, and strategies):

cd forex-rl
pytest -q

For walk‑forward experiments, see the examples in forex-rl/wfo/README.md. For GA Trailing‑20 work, see forex-rl/ga-trailing-20/ROADMAP.md.

Local broker (local-broker/)

A Go service that exposes an OANDA‑compatible REST + streaming surface backed by an in‑memory engine and pluggable pricing sources.

High level:

• Pricing sources for OANDA streams and synthetic ticks (internal/pricing).
• In‑memory engine with TP/SL/trailing‑stop enforcement (internal/state).
• File‑backed store for equity/time‑series snapshots (internal/storage).
• HTTP server wiring + config loader under cmd/broker and internal/server.

Development loop:

cd local-broker
go test ./...
go run ./cmd/broker

More details live in local-broker/LOCAL_BROKER_PLAN.md.

Market data and streaming (streamer/)

Utilities to ingest and explore market data, with scripts to push OANDA FX and Bitunix crypto into QuestDB and to track account state.

Examples:

• forex-stream.py – OANDA FX tick/stream recorder.
• bitunix_to_questdb.py, oanda_to_questdb.py – ingestion to QuestDB.
• questdb_ingest.py, questdb_admin.py – admin/helpers for QuestDB tables.
• orders.py, account_state.py – simple account/order abstractions used by the streaming tools.

Systemd unit examples for long‑running ingestion live under streamer/systemd/.

Secrets and environment

• Local secrets (API keys, account IDs, etc.) live under .secrets/ and are not intended to be committed.
• Most Python entrypoints read credentials from environment variables or .env‑style files in .secrets/; see the relevant script’s docstring for exact names.

Optional: OANDA tick recorder

• See streamer/forex-stream.py (requires oandapyV20 and env vars like OANDA_DEMO_ACCOUNT_ID, OANDA_DEMO_KEY).