Cascading Liquidity Withdrawal In A Selloff

On May 6, 2010, the S&P 500 fell roughly 5% and then recovered almost all of it within about twenty minutes. During the worst of it, some individual stocks traded for pennies or for hundreds of dollars, not because their fundamental value had actually changed that much, but because the order books that normally supplied prices had emptied out. That event — the "Flash Crash" — is the canonical illustration of a liquidity withdrawal cascade: a self-reinforcing loop where falling prices cause liquidity providers to quote less, which causes prices to fall further on the same order size, which causes liquidity providers to quote even less.

The mechanism, step by step

Start with a market maker quoting a healthy two-sided book. A wave of selling arrives — large enough that the market maker starts to suspect it's informed, not random. Two things happen inside the market maker's own risk model at once: their inventory grows more long than they want (having absorbed shares on the bid), and their estimate of how much further the price might fall goes up, because persistent one-directional flow is exactly the pattern that precedes further declines. Standard practice under both effects is to widen the quoted spread and shrink the quoted size, precisely the behavior described in The Two Sides Of Liquidity: Supply And Demand. But now the book is thinner than before the wave arrived. The next seller who needs to trade the same size moves the price further than the first seller did, for an unchanged amount of selling. That bigger price move is itself new information to every other market maker watching the tape — it looks like confirmation that something is wrong — so they widen too, even in unrelated but liquidity-correlated names, per Commonality In Liquidity. The loop repeats, each round pushing the price further per unit of order flow than the round before.

A worked example

A market maker is quoting 5,000 shares deep at a $0.02 spread. A sell program routes 5,000 shares, clearing the bid; the market maker reprices, now quoting only 2,000 shares deep at a $0.05 spread, having marked down fair value by 4 cents on the inference that the seller knows something. A second 5,000-share sell order arrives: it clears the 2,000-share bid and then walks through the next two price levels to fill the remaining 3,000 shares, moving the price down 11 cents — nearly three times the impact of the identical-sized first order, purely because the book had less depth to absorb it. If the market maker's model treats this second, larger move as further confirmation of informed flow, the next quote might be only 500 shares deep at a $0.12 spread, and a third identically-sized sell order could move the price by 30 cents or more. Same order size each time; wildly escalating impact, because depth is being depleted faster than it's replenishing — precisely a book whose recovery is slower than the arrival rate of new sell orders, as described in Measuring Resilience: The Liquidity Half-Life.

Where this gets used

• Circuit breaker design: trading halts and limit-up/limit-down bands exist specifically to break this loop by forcing a pause, giving market makers time to re-assess without the pressure of continuously arriving orders, rather than to fix any underlying fundamental problem.
• Stress-testing execution plans: a desk sizing how much it can trade "safely" needs to model that its own available liquidity shrinks as it trades, not just that price moves — using a static depth assumption throughout a large unwind understates cost precisely when the cascade is starting to bite.
• Systemic risk monitoring: regulators and risk teams watch for early cascade signatures — synchronized depth withdrawal across correlated names — as a leading indicator distinct from realized volatility, which only shows up after the loop has already run.

In interviews

If asked to explain a flash-crash-style event, walk through the loop rather than naming a single cause: falling price → market maker caution → thinner book → bigger impact per order → feeds back into price. Naming all four links, and explaining why they reinforce rather than offset each other, is what distinguishes a mechanistic answer from a vague "panic selling" answer.