Spread Regime

A venue’s spread is not a number; it is a mixture of distributions. Sampled across a week, the bid-ask spread of a liquid pair does not scatter around one centre — it occupies distinct states with sharp transitions between them: a tight calm-session regime, a wider regime around session opens and the daily rollover, an event regime around scheduled releases in which the spread is a multiple of its calm value, and a stress regime in which quoting thins toward withdrawal. Each regime is persistent while it lasts, and each has its own distribution.

This is why the “typical spread” of marketing pages is an actively misleading statistic. An average taken across regimes describes a spread that no individual trade ever pays: it understates the event regime by construction, because calm periods dominate the clock even though events dominate the cost. The distinction matters doubly because retail order flow is not uniform across regimes — stops, pending orders, and breakout entries cluster precisely at the transitions and events where the spread is widest. The regime a trader’s orders experience is systematically worse than the regime the average describes.

Honest measurement is therefore regime-conditional: a spread distribution per regime, the transition timing between regimes, and how a given venue’s regime structure compares to a multi-venue reference at the same timestamps. Spread widening is the transition into the event regime seen close up; censoring is one way a feed can disguise its true regime structure by suppressing the ticks that would reveal it.

The backtest implication is the same one that recurs across execution cost: a simulation fed a flat spread assumption pays the calm regime on every trade, including the trades that fired into the event regime. That single flattering constant is one of the quieter roads into backtest overfitting — optimism laundered through an input nobody audits.