Detecting Stale Quotes and Crossed Books in Raw Feeds
A crossed book is a data error until proven otherwise. The integrity checks every raw order-book feed needs before it can support research.
In a functioning continuous order book, the best bid is below the best ask. When a reconstructed book shows the opposite — a crossed book — the cause is rarely the market. Far more often, a delete message was missed, updates were applied out of order, or a snapshot was merged incorrectly with incremental data.
Common failure modes
- Dropped messages that leave cancelled orders in the reconstructed book.
- Out-of-sequence updates applied in arrival order rather than sequence order.
- Snapshots taken at a different sequence point from the incremental stream they are combined with.
- Venue-side incidents, which do occur but are rare and usually brief.
Checks we apply
Every reconstructed book state is tested for crossing and locking, for sequence continuity, and against venue checksums where venues provide them. Quote age is tracked for each level, so levels that persist far beyond typical lifetimes can be flagged as potentially stale.
The principle is simple: a research dataset should make its imperfections visible. A flagged gap can be excluded or modelled; an invisible one quietly contaminates every result built on top of it.
This publication is provided for informational purposes only and does not constitute investment, legal, or tax advice, or an offer or solicitation to buy or sell any asset. Live figures are computed from third-party public market data and may be delayed, incomplete, or inaccurate.
