Liquidity Is Not There When You Need It

Price impact per dollar traded in SPY reached roughly twice its own recent normal in October 2008, in March 2020 and in the first week of August 2024, and about a third above it in April 2025. Each of those readings sits above the 87th percentile of a twenty-eight-year sample, and all four coincide with the sessions on which a holder would most have wanted to reduce a position.

Figure 1 measures that in its upper panel and simulates the mechanism behind it in the lower one. The two panels have different evidentiary status and the figure says which is which.

Section 01Two kinds of liquidity

An order-book snapshot answers a narrow question: what limit orders were displayed at the instant it was taken. It does not answer how much can be executed after an order reveals direction, consumes queue, and invites the remaining resting size to reconsider.

A resting limit order is a contingent commitment. It can be cancelled, and in a fast market it usually is. So a large order does more than consume static levels: it changes the information set of every observer and alters the state it was meant to measure. The pre-trade book is a benchmark. The book the order actually meets is a response.

Section 02What the measurement shows

The Amihud illiquidity ratio is the standard way to get at this from daily data: absolute return per dollar of trading volume, averaged over a window. A high reading means price moved a great deal per dollar traded, which is the operational definition of thin liquidity.

Its raw level is not comparable across decades, because SPY's dollar volume grew by orders of magnitude over the sample and that trend swamps everything else. Figure 1 therefore normalises by a trailing one-year median, which turns the series into a stress index: a reading of two means impact per dollar is twice its own recent normal. That transformation is the reason the figure is legible and it is also the reason the figure cannot speak to whether markets have become more or less liquid over time. It has that information divided out by construction.

What survives is the conditional statement, which is the one the article needs. Liquidity behaves as a state variable rather than as a constant of the market. It deteriorates by roughly a factor of two in stress, and it does so in exactly the episodes when an exit is wanted.

Section 03Why it evaporates

Two classical results supply the mechanism. Glosten and Milgrom showed that a spread arises even among risk-neutral competitive market makers when some incoming orders may be informed: the quote has to be wide enough to recover from uninformed flow what it loses to informed flow. Kyle gave the companion intuition for impact, with price moving in proportion to net order flow and the constant of proportionality set by how much informed trading the market suspects.

Both imply that liquidity is a function of perceived information, and that a large order is itself evidence. The lower panel of Figure 1 puts that into a cost curve. It walks a marketable order of increasing size through a book with geometric depth decay, in a calm state and in a stressed state where the spread is wider, depth is thinner, and 35 per cent of resting size cancels ahead of the order.

The result is not that stressed execution is more expensive, which is obvious, but how fast the gap widens. At one thousand shares the stressed book costs six times the calm one. At five thousand, eleven times. At eight thousand, fourteen times. Past about ten thousand shares the stressed book has no displayed size left at all, and the curve stops rather than quoting an average price that was never available.

That shape is the practical warning. A desk sizing from average conditions is not wrong by a constant factor; it is wrong by a multiple that grows with the order, and the order it most wants to send in a stress is the large one.

Amihud (2002) illiquidity, as computed here:

    raw_t   =  mean over 21 sessions of  |return| / dollar_volume
    index_t =  raw_t / median( raw over trailing 252 sessions )

    the normalisation removes the secular growth in dollar volume,
    which otherwise dominates the series; it also removes any
    ability to compare liquidity levels across decades.

measured, SPY, 1998 to 2026:

    median                     0.91
    Oct 2008                   1.82      97th percentile
    Mar 2020                   1.95      98th
    Aug 2024                   1.76      97th
    Apr 2025                   1.33      87th

execution cost, against a declared decision price:

    shortfall  =  (fill_price - decision_price) * signed_size
               =  spread + delay + impact + adverse selection

Section 04The stress case, and what backtests miss

The clearest documented illustrations come from venues normally considered among the deepest anywhere. On 15 October 2014 the US Treasury market experienced an extreme move in a narrow window, and the official Joint Staff Report found no single cause, describing instead a confluence of factors including a sharp decline in depth. A market that is the benchmark for risk-free liquidity produced a price path that its own depth could not explain.

Almgren and Chriss framed the resulting problem as a trade-off with no free corner: trade quickly and pay impact, trade slowly and accept the risk that the price moves away. Both sides of that trade-off worsen in the states Figure 1 identifies, which is why a strategy calibrated on average liquidity is not merely optimistic but optimistic in a correlated way, across exactly the days its risk model cares about.

The liquidity in your risk model is the liquidity you will not have on the day you need to trade.

The useful reframing is to treat liquidity as a response function rather than a stock. Executable size depends on order size, urgency, information content, participation rate, and the state of the market at the moment of trading. None of those is a property of the book you can see, and the one that matters most, the state of the market, is the one that moves against you.

This research is analysis and commentary for general information. It is not investment advice, an offer, or a solicitation, and it contains no price forecasts. The upper panel of Figure 1 is the author's own calculation from the source cited; the lower panel is simulated.