Running an ONNX model on crypto data — what's different from forex.

Most of this site is about ONNX inside MT5 against forex data. The pattern transfers cleanly to crypto on Bybit, Binance, or Kraken — same model export, same runtime, similar architectures. But the data has structural differences that matter for model design. This article walks through the four big ones.

24/7 markets — no weekly close

Forex markets close on weekends; crypto doesn't. Two consequences for ML:

• No daily-bar effect: the typical "open of London/NY session" features used in forex don't translate. Crypto markets have continuous flow.
• More data per year: ~8760 hourly bars vs ~6000 for forex. Training datasets are larger.
• Different distribution: weekend volume is much lower than weekday; weekend price moves are sometimes outsized. Don't normalize weekend bars against weekday stats.

Volatility regimes

Crypto has volatility regimes that don't exist in forex majors. BTC can move 5% in an hour (a 6-sigma event in EURUSD); a stable-coin de-peg event can break any model trained on normal-regime data.

Practical implications:

• Normalize within rolling windows, not over the full history. A 30-day rolling z-score captures the current regime; a full-history z-score smears the recent regime into ancient volatility.
• Train on multiple regime samples. A model trained only on 2024 calm-period data will malfunction in 2025 trend-up data.
• Cap position sizing per-volatility. Static lot sizes are dangerous in crypto. Scale inversely to recent realized vol.

Order book depth as a feature

Crypto exchanges expose level-2 order books over WebSocket. The depth at various price levels — how much resting liquidity there is at 0.1% above and below the mid-price — is a real predictive feature. Forex retail traders don't get this; crypto does.

Adding L2 features to your feature vector:

• Bid/ask volume ratio (top-of-book).
• Depth-imbalance at +/− 0.5%, +/− 1%, +/− 2%.
• Trade flow imbalance over the last N seconds.

These features capture the supply/demand structure that price alone doesn't. Useful additions to any sequence model.

Funding rates as a feature

Perpetual futures (the dominant crypto derivative) charge a funding rate every 8 hours, calculated from the basis between perp and spot. The rate is a noisy proxy for "leveraged sentiment" — high positive funding means longs are paying shorts (bullish sentiment, often crowded), negative means the opposite.

Funding-rate features that work as inputs:

• Current funding rate.
• Rolling 7-day average of funding.
• Funding minus 30-day-mean (deviation from baseline).

Putting it together

A crypto-specific ONNX feature vector might look like: 120 hourly returns + 60 hourly volumes + current bid/ask imbalance + current funding + 7-day funding average. Train as you would forex; export with opset 17; run inference as in the Bybit algo article.