Switchback experiment

A switchback experiment applies treatment and control to an entire system, a city's driver-matching algorithm, a warehouse's routing logic, alternating between them across defined time windows, instead of splitting individual users into separate groups at the same time. A ride-hailing market might run its current pricing algorithm from 8-9am, a new one from 9-10am, and so on, then compare outcomes across the treatment and control windows.

This design exists because many marketplace and operational systems have strong network effects or shared resources, drivers, inventory, matching pools, that make a standard user-level A/B test invalid: treating half of users with a new matching algorithm inevitably changes outcomes for the untreated half too, since they compete for the same drivers or inventory. Switchbacks avoid this "interference" by treating the whole system at once, so there is no cross-contamination between arms during any given window.

Analysis has to account for autocorrelation between adjacent time windows and any time-of-day or day-of-week pattern that could be confused with a treatment effect, typically by randomizing which windows get which treatment and using models that account for the time structure. Switchback experiments are common in ride-hailing, delivery logistics, and other two-sided marketplaces where individual-level randomization would leak.