Interference and Spillover Effects

Most causal analysis assumes each unit's outcome depends only on whether that unit was treated. Interference (also called spillover) is what happens when one unit's treatment leaks into another's outcome. A trading example: you test a new order-routing algorithm on a subset of stocks by trading them differently, expecting the rest of your universe to be an unaffected control group. But if your new routing changes how much liquidity you take from a shared dark pool, the "control" stocks' execution costs shift too, because everyone is drawing on the same finite liquidity. The comparison you built the test around is now contaminated — the control group was never really untouched.

Spillover shows up constantly in market experiments because markets are networks, not isolated boxes: order flow, capital, and information move between related names. A/B tests on market-making quoting logic can leak through correlated pairs; a promotional trading signal pushed to some accounts can move prices that affect other accounts' fills; even academic event studies on one stock can bleed into its close peers through common analysts or index membership.

The core diagnostic question is simple: could the treatment given to unit A plausibly change the outcome measured for unit B? If yes — through a shared market, a network tie, or a common resource — treat the "control" group with suspicion. Fixes generally work by changing the unit of comparison rather than patching the statistics: randomize at the level where interference can't cross (e.g. by whole market or whole day rather than by individual stock), or explicitly model the spillover pathway and estimate it rather than assuming it away.