How I validated cross-exchange lead/lag — 90.7% on ETH, then BTC and LINK — with live price data

Cross-exchange lead/lag in crypto is academically known: Binance Futures moves first, spot exchanges follow within milliseconds, and a fast bot can capture the gap. The phrase shows up in market microstructure papers, quant blogs, and the occasional X thread.

But the numbers people quote are usually old, simulated, or pulled from someone else's data. I wanted real measurements, and I wanted to know exactly when the strategy stops working. So I built the data collection myself.

This post is the methodology and the numbers. The deep dive below is ETH: Binance Futures → Coinbase Spot at a 0.10% threshold gives 90.7% follow-through, ~300ms median follow time, and profitability that collapses past 0.12% round-trip fees. I've since run the same methodology on BTC and LINK — both cleared the bar (more on that below). Full ETH breakdown first.

The setup

I connected to four exchange WebSocket feeds simultaneously and recorded every ETH price update with a millisecond timestamp:

• Binance Futures (the leader candidate)
• Coinbase Spot (potential follower)
• Bybit Spot (potential follower)
• Bybit Futures (potential follower)

All four streams ran 24/7 against a Postgres-backed signal engine. Over 7 days the engine inserted just under 9.4 million price updates with a 0.0008% error rate. No backtesting, no replay — every datapoint is from live markets.

The engine also times itself. From detecting a leader move to the signal hitting the wire is about 2ms on the engine clock; delivery to a co-located client in US-East lands in single-digit milliseconds — I measured a ~6ms median from an NTP-synced box, backed by a 525-sample heartbeat baseline. That latency is the part I control. The follower-lag numbers below are the part the market gives you — and the gap between the two is your tradeable window.

Two definitions matter:

• Lead event: Binance Futures moves ≥0.1% in either direction within a 5-second window.
• Follow-through: the follower exchange moves the same direction within 60 seconds of the lead event.

The 0.1% threshold filters out routine noise. The 60-second window is generous on purpose — I wanted to capture the full follow move, not just the initial reaction. Tighter windows give different numbers, which I'll cover below.

What the data shows

Binance Futures → Coinbase Spot at the 0.10% threshold:

The pattern is the classic precision/recall tradeoff. Lower thresholds catch more events but include more noise. Higher thresholds are more reliable but rarer.

0.10% sits at a useful middle: ~150 actionable signals per week, 90%+ accuracy, predictable timing. That's the threshold most retail bots can realistically work with — high enough to filter noise, frequent enough to amortize fixed costs of running infrastructure.

Where it stops being profitable

Follow-through rate doesn't tell you whether you'd make money. The follower has to move further than your round-trip fees to actually profit. So I ran the same dataset against four common fee tiers:

This is the most important table in the whole writeup. The strategy is highly profitable at maker fees and completely dead at taker fees. There's no in-between — the curve is sharp.

What this means in practice: if you don't have maker rebate access on the follower exchange, this isn't your edge. Cross-exchange lead/lag isn't competing on direction prediction. It's competing on execution speed and fee tier. The signal is reliable; the only question is whether your infrastructure and fee tier let you capture the spread before it equalizes.

What this data doesn't tell you

A 90.7% follow-through rate on a clean dataset is not the same as 90.7% net profitability in production. Some things this measurement deliberately doesn't account for:

• Slippage and partial fills. The data assumes you can place a maker order at the right price and have it filled at that price. In reality, during fast moves, the order book reshapes around you. Some fills will be worse than the "expected" price. Some won't fill at all.
• Queue position. Maker orders sit in a FIFO queue. If you place an order at the right level but the price moves through it before your order reaches the front of the queue, you don't get filled.
• Cost of capital. To capture follower moves on Coinbase, you need capital sitting on Coinbase. That capital isn't earning yield while it waits.
• Adverse selection. If you become a notable participant, market makers on the follower exchange may start adjusting their quotes to penalize your order flow specifically. The edge can decay as you scale.
• Time window. The data covers 7 days in spring 2026. ETH microstructure can change. A volatile week, a regime shift in market makers, or a new HFT firm arbitraging the gap could all change the numbers.

None of these invalidate the underlying signal. They just mean a 92.7% "profitable trades" number is the upper bound — your actual capture rate will be lower. How much lower depends on your execution stack, capital, and timing.

Why I started with ETH — and what BTC and LINK showed

ETH was the cleanest first case: liquid enough for continuous price discovery on every venue, but it lags BTC-led moves closely enough that there's a real cross-exchange gap to capture. So I validated it first and deepest — every table above is ETH.

I originally expected BTC to be a worse follower — it's often the market's lead asset rather than a follower, and its cross-exchange spreads are tighter because more capital arbs them. The first version of this post said exactly that. Then I ran the same methodology on BTC, and it held up better than I'd predicted: about 88.6% follow-through on Bybit Spot.

LINK cleared the bar too — roughly 83% follow-through on Bybit Spot at the 0.05% threshold. One pattern stood out: for both BTC and LINK, Bybit Spot was the cleaner follower venue, while ETH's flagship follower is Coinbase Spot. Different assets lead into different venues, so each is validated and positioned independently rather than assuming one venue fits all. (I also tested Bybit Futures as a follower and dropped it — two perpetual-futures venues co-move within milliseconds, so there's no capturable lag, just correlation.)

The live roster is now ETH, BTC, LINK, XRP, and DOGE. XRP and DOGE cleared the same bar on Bybit Spot (~85.6% and ~81.7% follow-through); other candidates I tested — SOL, SUI, APT — didn't, and were dropped. New assets get added only once they clear that bar: high same-direction follow-through, low opposite-direction rate, a real capturable lag, and enough signals to trust the numbers.

What's next

The signal engine is still running. Live performance is tracked at /api/v1/stats/performance on a rolling 24h, 7d, and 30d basis — anyone with a free API key can verify the numbers above haven't drifted. They will drift, of course. Markets move. The point is that the methodology is reproducible and the live data is inspectable.

Things on the roadmap, prioritized by what bot devs ask for:

• Done — Bybit Spot follower predictions, now live for BTC, LINK, XRP, and DOGE
• Done — BTC validation (~88.6% follow-through on Bybit Spot)
• Done — lower-threshold (0.05%) signal stream
• More follower assets as new candidates clear validation
• Per-strategy backtesting endpoints

If you're running cross-exchange strategies, the questions I'd love to compare notes on:

• What latency do you measure between Binance Futures and your follower exchange from your infrastructure?
• What's your fill rate on maker orders during fast moves at the 0.10%+ magnitude band?
• What fee tier are you operating at, and how does it shape your strategy selection?

DMs and email are open. The whole point of publishing the methodology is to invite scrutiny — if the numbers don't hold up to your testing, I want to know.