Change-point detection isolates moments when the statistical behavior of an order book shifts, flagging structural breaks that matter for traders, risk managers, and regulators. By modeling time series of mid-price returns, bid-ask spread, depth at top levels, and arrival rates of limit and market orders, researchers can detect changes that indicate liquidity stress, informed trading, or exchange-specific events.
Methods and algorithms
Classical econometric approaches that test for multiple breaks provide a foundation. Jushan Bai Columbia University and Pierre Perron Boston University developed frameworks to estimate break dates and the number of regimes in time series, useful when structural changes are persistent. Sequential and online methods based on CUSUM statistics detect sudden mean or volatility shifts with low latency and are practical for real-time surveillance. For computational efficiency on high-frequency order-book streams, the PELT algorithm proposed by Rebecca Killick Lancaster University Paul Fearnhead Lancaster University and Iain A. Eckley Lancaster University offers near-linear segmentation, enabling fast identification of multiple change points across thousands of observations per second.
Relevance, causes, and measurable signals
Change points appear in measurable order-book features: a sudden widening of the bid-ask spread signals liquidity withdrawal; a persistent shift in top-of-book depth reflects a new regime of market-making; alterations in message rates indicate a change in participation. Common causes include large block trades by whales, exchange outages or reconfigurations, major news or regulatory announcements, and design changes in matching engines. Cultural and territorial factors influence causes and interpretation because market participant composition varies by venue and jurisdiction; an Asian spot venue may show different response patterns to regional news than a U.S. venue.
Consequences and operational use
Detecting structural breaks enables adaptive execution strategies, dynamic risk limits, and post-event forensics. Early detection can help liquidity providers adjust quoting algorithms to avoid adverse selection, and regulators can identify systemic events spreading across exchanges. Environmental aspects such as mining-related network events or scheduled protocol upgrades on decentralized exchanges can also create predictable change points that deserve distinct handling. Nuanced interpretation is essential because not every flagged change implies malfunction; some reflect regime shifts in trader behavior that persist and reshape market microstructure.