Splitting the Spread in Two: Glosten and Harris on What You Are Actually Paying For

Roll taught us how to measure how wide the bid-ask spread is. But knowing the size of a bill is not the same as knowing what is on it. If a market maker quotes you a spread of ten cents, what are those ten cents for?

By the mid-1980s there were two rival stories, and they had very different implications. Glosten and Harris, in 1988, built the machinery to ask the data which story was right, and by how much.

The problem: one number, several causes

Two theories of the spread had been circulating.

The first is the cost and inventory story, associated with Demsetz, Stoll, and Ho. A market maker runs a business. There is paperwork, exchange fees, clearing, staff, and the risk of getting stuck holding a pile of stock they did not want. To stay in business they charge a markup. Think of it as a shopkeeper's margin.

The second is the adverse selection story, associated with Copeland and Galai, Kyle, and Glosten's own earlier work with Milgrom. A market maker quotes prices to the whole world, and some of the people hitting those quotes know something they do not. Every time an informed trader buys from the market maker, the market maker loses. To survive, the market maker must widen the spread enough that the profits from ordinary, uninformed traders cover the losses to the informed ones.

These sound similar because both produce a spread, but they behave completely differently. The cost markup is basically a fixed toll: pay it, and the price snaps back. The adverse selection charge is not a toll at all. It is a permanent revision of belief. If a big buy order makes the market maker suspect the buyer knows something, the market maker does not just charge them more, they move the price up and leave it there.

Regulators and exchanges cared a lot about which was which. If the spread is mostly cost markup, more competition among market makers will shrink it. If it is mostly adverse selection, competition will not help at all, because the adverse selection charge is not profit, it is a loss reimbursement.

The key idea via analogy: which part of the price move comes back?

Here is the trick that makes the whole thing work.

Imagine you buy a stock and it ticks up. Wait a while, and watch what happens.

• If the tick-up was transitory, meaning the price falls back to where it was, then you just paid a toll. That is the cost and inventory component. The market maker charged you for immediacy and then the price returned to fair value.
• If the tick-up is permanent, meaning the price stays at the new higher level, then the market went "hmm, that buyer might know something," and revised its estimate of the stock's worth. That is the adverse selection component. Your own trade taught the market something and the lesson stuck.

So: the transitory part is the fee, the permanent part is the information. The whole art is separating the two, and Glosten and Harris did it by writing a model of the trade-by-trade price where a buy order pushes the price via two distinct channels, one that decays and one that does not, and then estimating how big each channel is from actual NYSE transaction data.

They added a second piece of realism that matters enormously. They allowed both components to depend on trade size. Intuitively, a bigger order is scarier: nobody hides a large informed bet in a hundred-share trade. So the adverse selection charge should scale with size. Meanwhile the fixed overhead per ticket does not care whether you bought 100 shares or 10,000. Their model lets the data speak on how each component grows with size, rather than assuming the spread is one flat number for everyone.

Why it mattered

• It turned a philosophical debate into an estimation problem. Before this paper, "how much of the spread is adverse selection?" was an argument. After it, it was a regression you could run on transaction data. That is a huge upgrade.
• It established that adverse selection is real and economically meaningful, and that it grows with trade size. This is the empirical backbone of a claim quants now take for granted: large orders cost more per share, and the extra cost is not a fee you can negotiate away, it is the market updating its beliefs about you. Every execution algorithm that slices a parent order into small children is, in effect, paying respect to Glosten and Harris.
• It gave microstructure a durable template. The "decompose the price change into a permanent, information-bearing piece and a transient, mechanical piece" move became the standard operating procedure. You see it again in Hasbrouck's vector autoregression work, in Madhavan, Richardson and Roomans, in Huang and Stoll, and in essentially every modern transaction cost analysis dashboard, which reports "permanent impact" and "temporary impact" as separate line items.
• It reframed policy. If a chunk of the spread is compensation for adverse selection, then squeezing market makers with tighter tick sizes or more competition does not simply hand savings to investors. It changes who bears the information risk, and possibly whether anyone is willing to quote at all.

The honest limitations

• You have to guess the direction of every trade. The model needs to know whether each transaction was a buy or a sell, and the raw tape does not say. In practice you infer it with a rule of thumb (compare the price to the prevailing quote midpoint), and those rules misclassify a meaningful share of trades. Errors in trade direction feed straight into the estimated components.
• It is a structural model, and structural models are only as good as their structure. The decomposition is not a fact read off the data. It is a fact read off the data through a specific model of how trades move prices. Change the assumed model, and the split can change. Different papers using different structures have produced noticeably different estimates of the adverse selection share for the same market.
• Inventory and order processing get lumped together. The paper cleanly separates information from everything else, but "everything else" is a bucket that mixes clearing costs, the specialist's market power, and inventory risk. Those have different economics and the model does not fully pull them apart.
• It was built for a world with a single human specialist. The estimation used NYSE data from the early 1980s. In today's fragmented, automated markets, with dozens of venues and no designated monopolist, the same decomposition still runs, but the story it tells about "the market maker's costs" needs translating.

The one-line takeaway

Glosten and Harris showed that the bid-ask spread is really two bills stapled together: a transitory fee for immediacy, which the price gives back, and a permanent adverse selection charge, which it does not, and they gave the profession a repeatable way to estimate how much of your trading cost is each.