Algorithmic trading on Bybit with ONNX models (2026 guide)

Bybit is one of the largest crypto exchanges by spot and derivatives volume, with a developer-friendly REST and WebSocket API that algorithmic traders can wire ONNX models into. Unlike forex where the MT5 ecosystem is mandatory, on Bybit you write your own trading loop in Python (or any language) and call onnxruntime directly.

This article covers the integration pattern, why Bybit specifically works well for algo traders, and the geographic restriction that affects some readers.

What's in this article

Why Bybit specifically
Geographic restrictions (Brazil, US)
The integration stack
A worked example: ONNX signal → Bybit order

Why Bybit specifically

Three reasons Bybit comes up in algo-trader conversations:

Liquidity. Top-3 globally by derivatives volume. Slippage on majors (BTC, ETH) is minimal even at non-trivial size.
API documentation and uptime. The V5 unified API is well-documented; rate limits are generous; their status page tracks uptime publicly.
Affiliate program. Up to 50% lifetime commission on referred trading fees — one of the most aggressive in the industry.

Geographic restrictions (Brazil, US)

Bybit has progressively restricted access in certain regions. As of 2026:

United States: derivatives effectively unavailable to US persons; spot may be available via partner exchanges.
Brazil: derivatives products are restricted due to local regulation. Spot trading remains available.
UK, Canada, Singapore, Japan: partial restrictions; check current local terms.

If derivatives are blocked in your jurisdiction, you can still run an ONNX-driven spot-trading bot — the API surface is similar.

The integration stack

Unlike the MT5 ecosystem (where MQL5 hosts everything), the Bybit-side stack is just Python:

requirements.txt

pybit==5.7.0 # official Bybit SDK onnxruntime==1.17.0 # or onnxruntime-gpu for CUDA numpy<2.0 pandas

This stack runs equally well on a local machine, a Linux GPU cloud instance, or a Windows VPS. No MetaTrader involved.

A worked example: ONNX signal → Bybit order

bot.py — minimal Bybit + ONNX

import numpy as np import onnxruntime as ort from pybit.unified_trading import HTTP # Load model once sess = ort.InferenceSession("btc_signal.onnx") # Bybit client (testnet first) client = HTTP(testnet=True, api_key=API_KEY, api_secret=API_SECRET) def fetch_features(): # Pull last 120 1-hour candles, return feature vector klines = client.get_kline(category="spot", symbol="BTCUSDT", interval="60", limit=120) closes = [float(k[4]) for k in klines["result"]["list"]][::-1] features = np.array(closes, dtype=np.float32).reshape(1, 120, 1) # ... apply same normalization as training ... return features def predict_and_trade(): x = fetch_features() y = sess.run(None, {"input": x})[0] signal = float(y[0]) if signal > 0.55: client.place_order( category="spot", symbol="BTCUSDT", side="Buy", orderType="Market", qty="0.001", ) elif signal < 0.45: client.place_order( category="spot", symbol="BTCUSDT", side="Sell", orderType="Market", qty="0.001", ) # Run every hour at minute 0:01 while True: try: predict_and_trade() except Exception as e: print(f"Error: {e}") time.sleep(3600)

That's the whole pattern. The ONNX model and the Bybit SDK are independent — you can swap the model file without touching the trading logic, and vice versa. Production-grade versions add error retry, position-tracking, drawdown limits, and structured logging.

Open Bybit account → (affiliate link — up to 50% lifetime on fees)

// related

how-to

ONNX on crypto data

upstream

Export PyTorch to ONNX

strategy

CNN-LSTM price forecast

infrastructure

GPU cloud for training