Crypto Arbitrage

Python 3 trading bot that scans for triangular, cross-exchange (inventory), and spot/perp basis edges. Paper trading is the default. Live trading is triple-locked.

This is a rewrite of the 2017 Python 2 Bittrex bot. Exchange adapters are gone; CCXT provides unified markets, order books, and order placement.

What changed

The old bot fired three innermost limit orders and hoped they filled. This version:

• Uses full order-book depth (VWAP), not just the top level
• Subtracts taker fees + a slippage buffer before calling anything an opportunity
• Binary-searches size so larger clips that walk the book still clear the profit floor
• Pre-filters triangles from a single fetch_tickers call, then hydrates books only for candidates
• Treats cross-exchange arb as pre-positioned inventory (no on-trade withdrawals)
• Adds a cash-and-carry / funding scanner (execution off unless you enable it)
• Refuses live orders unless paper: false, --live, and LIVE_TRADING=1 are all set
• Cancels a sequential live path with a best-effort unwind if a later leg fails

Retail HFT this is not. Fees, latency, and inventory still eat most prints. Use it as a scanner first.

Strategies

Mode	Idea
triangular	Same-exchange 3-leg cycle (e.g. USDT → BTC → ETH → USDT). Needs inventory of all three assets if execution.style is parallel.
cross	Buy the cheap venue, sell the rich one. Capital stays on both exchanges; rebalance later off-bot.
basis	Spot vs USDT-perp: rich perp → cash-and-carry signal; also surfaces funding APR. Scan-only by default.

Setup

Python 3.11+ (developed on 3.13).

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
copy .env.example .env

API keys are optional for scanning public books. Add them to .env only if you will fetch private balances or place live orders.

Usage

python -m arbitrage doctor
python -m arbitrage scan
python -m arbitrage scan --strategy triangular --once
python -m arbitrage run
python -m arbitrage run --strategy cross

main.py is a thin wrapper around the same CLI.

Live trading (real money):

1. Set paper: false in config.yaml
2. Set LIVE_TRADING=1 in .env
3. Run python -m arbitrage run --live

IOC limit orders are used by default (execution.order_type: ioc) at the worst book price the simulator consumed.

Config

See config.yaml. Useful knobs:

• triangular.min_profit_bps — net edge after fees and the slippage buffer
• triangular.max_cycles / min_24h_volume_usd — how many triangles to track
• risk.max_notional_usd / max_daily_loss_usd / max_trades_per_hour
• risk.stale_book_ms / max_book_skew_ms — drop cross-venue prints from mismatched clocks
• basis.execute — keep false unless you accept futures liquidation risk

Paper fills update data/paper_state.json. Logs go to data/arbitrage.log.

Tests

pytest

Difficulties that still apply

1. Fees dominate. A 6–10 bps floor is optimistic on 10 bps taker venues; VIP fees or maker flow change the math.
2. Latency. A REST poll cannot beat colocated takers. Opportunities you see are often already gone.
3. Inventory. Parallel triangular and spatial arb both need coins on the right books. Transfers are slow and are not done in the hot path.
4. Partial fills. Sequential live mode tries to unwind; it can still leave you with residual risk.

Disclaimer

Educational software. You can lose money, including more than the intended notional when futures are enabled. No warranty.

Original project © 2017 Kelvin Au (MIT).