←BACK TO BLOG
>

The Bid-Ask Spread Is Three Different Prices

Market MicrostructureTradingLiquidity
JUL 2026·6 MIN READ

#ONE NUMBER, THREE STORIES

Every trading desk has a version of the same slide: assets ranked by bid-ask spread, narrowest to widest, labeled "liquidity." It treats a two-cent spread and a twenty-cent spread as differing in degree, not in kind, as if a market maker widens a quote for one reason, and more of that reason produces a wider number. That's not what's happening, and the literature has known it since at least 1985.

A dealer widens a quote for three economically distinct reasons: the fixed cost of processing an order, the risk of being stuck holding inventory before offsetting it, and the fear of trading against someone who knows more than they do. Each has a different name in the literature (order-processing costs, inventory-holding costs, adverse-selection costs), and each responds to a completely different lever: technology, capital, and information, respectively. Read the spread as one aggregate score and you lose the ability to tell which lever actually moved.

#THREE REASONS A DEALER WIDENS A QUOTE

Order-processing costs are the boring one: running a matching engine, paying exchange fees, compensating the desk for the mechanics of making a market at all (Demsetz 1968, Tinic 1972). It's the cost of existing as a dealer, and it falls with better technology (colocation, cheaper execution, tighter operations) almost independent of what's being traded.

Inventory-holding cost is different: it's the compensation a dealer demands for carrying risk on their book between the moment they take one side of a trade and the moment they lay it off (Stoll 1978, Amihud & Mendelson 1980). This is a capital and volatility problem: the wider the swings in the underlying, or the thinner the dealer's balance sheet, the more they charge to warehouse the position, even if every counterparty they trade with is a completely uninformed liquidity taker.

Adverse selection is the strangest of the three, and the one most people misunderstand. Glosten and Milgrom's 1985 paper showed that a positive spread can exist purely because some fraction of order flow is informed, even with a risk-neutral, zero-profit specialist and zero explicit costs of any kind. The specialist loses money on every trade with an informed trader and makes it back on noise traders; the spread is the insurance premium, not a fee for service. It's a pure information tax, and it would exist even in a costless, infinitely well-capitalized market.

#THE MODELS DON'T AGREE WITH EACH OTHER

Glosten and Harris built the first empirical machinery to actually split observed spreads into these components using trade-by-trade data in 1988. Since then, a half-dozen competing decomposition models (George, Kaul & Nimalendran (1991), Lin, Sanger & Booth (1995), Huang & Stoll (1997), Madhavan, Richardson & Roomans (1997)) have each estimated the same three components from the same kind of data and produced materially different splits.

Van Ness, Van Ness and Warr (2001) put a number on the disagreement: for large stocks, a two-way decomposition attributed only about 11% of the spread to combined adverse-selection-plus-inventory costs, while a three-way model attributed roughly 28.7% to inventory costs alone, for the same stocks, over the same period. That's not measurement noise; it's evidence that the three components are real, separable economic quantities that our estimation tools are still bad at cleanly isolating. Treating the spread as one number doesn't just simplify the truth, it erases the actual disagreement about what the truth is.

Same Shock, Different Spread Response Across the Treasury Curve (March 2020)

1.55-Year Note210-Year Note6.230-Year Bond

In the March 2020 Treasury turmoil the 30-year bond's bid-ask spread blew out past 6x its post-crisis average while the 5-year note barely widened 1.5x, the fingerprint of duration and inventory risk rather than one uniform liquidity shock.

#WHEN THE COMPONENTS DIVERGE, THE SPREAD ITSELF TELLS YOU WHICH ONE MOVED

March 2020 is the cleanest natural experiment I know of. Treasury bid-ask spreads widened across the curve during the COVID selloff, but not uniformly: the 30-year bond's spread rose to more than six times its post-crisis average, the 10-year roughly doubled, and the 5-year widened only about 50%, per Federal Reserve analysis. If spread were one "liquidity" score, you'd conclude the whole curve seized up equally. It didn't; the divergence tracks duration, meaning dealers were pricing the risk of warehousing long-duration exposure they couldn't offload, not a sudden information asymmetry specific to 30-year yields. The Fed's response confirms the diagnosis: it pledged to add at least $500 billion in Treasury holdings and $200 billion in agency MBS, offered $1.5 trillion in funding to primary dealers on March 12, and ultimately bought roughly $700 billion in Treasury notes and bonds. That's a capital fix, because the disease was inventory risk, not adverse selection.

Compare that to GameStop in January 2021. Spreads on GME options and shares widened just as dramatically, but the SEC's staff report on the episode points to a different cause: order flow that market makers had every reason to treat as informed or coordinated, from retail and hedge funds alike. Same visual signature (a much wider quote), but a completely different lever moved. No amount of dealer capital fixes adverse selection, and no amount of information about who's trading fixes an inventory problem. Mistake one for the other and you'll call an informed market merely risky, or a risky market suspiciously informed.

#READ THE LEVER, NOT THE NUMBER

This isn't just an academic distinction: Back and Baruch (2004) showed that Glosten-Milgrom's sequential-trade model and Kyle's (1985) continuous strategic-trading model converge to the same equilibrium as trade sizes shrink, unifying the two canonical ways economists model informed trading. The math holds together because the underlying object (information asymmetry as a distinct, quantifiable force on price) is real, not a modeling convenience.

So when a spread widens, the right question isn't "how much did liquidity deteriorate," it's "which of three unrelated diagnostics fired."

  • ▸Order-processing widening: a small, structural move with no accompanying volatility or news; fix it with technology, not capital
  • ▸Inventory widening: correlates with volatility and duration, hits capital-constrained dealers hardest, gets fixed by balance sheet or a backstop like the Fed's March 2020 facilities
  • ▸Adverse-selection widening: concentrated in a specific name or contract amid a genuine information event, and no amount of dealer capital makes it go away

Two Models, Two Very Different Answers About the Same Spread

11Two-way model (adverse selection + inventory)28.7Three-way model (inventory only)

On the same large-cap stock sample, a two-way decomposition assigns just 11% of the spread to combined adverse-selection and inventory costs while a three-way model assigns 28.7% to inventory costs alone, showing these components are real but hard to cleanly isolate.

#Further Reading