Cross-exchange price discovery in ETH and BTC

[Build note — remove before publishing] You’ve already run crypto_cross_exchange_demo.py on real data. Replace every [bracketed] value below with your actual numbers from dma_demo_summary.csv, export one figure (ETH-vs-Kraken looked liveliest) to public/research/crypto-cross-exchange.png, and set the chart: field in the frontmatter above. Do not invent numbers — use what the run produced.

The question

When the price of ETH or BTC moves, do the major exchanges move together — and is the visible gap between them a real, gettable dislocation, or just a structural artifact?

Data

Public one-minute market data for BTC and ETH collected from Binance, Coinbase, and Kraken over a [12-hour] window, aligned to a common clock. No private or proprietary feeds are used; figures shown are our own derived analysis, not redistributed raw data.

Method

For each asset we reconstruct the price path on every venue, align them minute-by-minute, and then: measure the raw cross-venue spread in basis points; remove a rolling-median structural baseline (the persistent offset — largely the USDT/USD stablecoin basis, since one venue quotes USDT and the others quote USD); and study the residual — the genuine, short-lived dislocations around that baseline. Separately we measure the lead-lag relationship between venues’ returns.

Findings

Across the window, BTC and ETH were effectively synchronised at one-minute resolution — return correlation of [0.7–0.95], with no measurable lead at this granularity. The apparent [~8 bps] gap between venues was almost entirely a persistent structural basis, consistent with the USDT/USD stablecoin spread, not a tradeable dislocation. After removing that basis, genuine transient dislocations were small and short-lived — [N] episodes beyond ±5 bps, median [X] minutes.

What this means

The honest read is a modest one, and that is the point: most of the raw cross-venue “gap” is a structural artifact, and separating it out is what keeps an analysis from overstating an edge. Detecting true sub-second lead-lag would require trade-level (tick) data rather than one-minute bars — an active research direction.

Reproducibility: this analysis was produced by our own pipeline (ingest → reconstruct → baseline- adjust → lead-lag), with a self-test on synthetic data that recovers a known injected lead and a known dislocation before any real data is touched.