Adverse Selection

Adverse selection is the single deepest idea in market microstructure: the act of trading against you is itself information. A market maker posts a two-sided quote to anyone. Uninformed traders (rebalancers, indexers, retail) trade for reasons unrelated to value. But informed traders trade only when the quote is wrong in their favor. So conditional on being filled, your counterparty is disproportionately someone who knows more than you, and you have, on average, just made a losing trade. The spread exists to make that survivable.

The market maker's winner's curse

This is the microstructure form of the winner's curse. In a common-value auction, winning means you bid the most, i.e., you were the most optimistic, so conditional on winning your estimate was probably too high. For a market maker the analog is: getting your offer lifted is bad news, because the buyer lifted it precisely when the true value exceeded your ask.

Formalize it. Let value $V$ be uncertain. You post an ask $a$. A buyer arrives; with probability $\theta$ they are informed and know $V$, with probability $1-\theta$ they are a noise trader who buys regardless. An informed trader lifts your ask only if $V > a$. Your expected profit conditional on being lifted is

$\mathbb{E}\big[a - V \mid \text{ask lifted}\big] = (1-\theta)\big(a - \mathbb{E}[V]\big) \;+\; \theta\,\mathbb{E}\big[a - V \mid V > a\big].$

The first term is your gain from uninformed flow (you sold above the unconditional mean); the second is negative, against informed flow you sold below the truth. Setting the whole thing appropriately positive requires $a > \mathbb{E}[V]$: you must post above fair value. By symmetry the bid sits below fair value. The spread is the wedge that lets the gain from the uninformed pay for the loss to the informed. No inventory cost, no processing cost, pure information, and you still get a spread. This is the engine of The Glosten-Milgrom Model.

Quotes must be regret-free (zero-profit Bayesian updating)

Under competition, a market maker earns zero expected profit per trade type. That means the quotes must satisfy a conditional-expectation (regret-free) condition:

$a = \mathbb{E}[V \mid \text{buy order}], \qquad b = \mathbb{E}[V \mid \text{sell order}].$

You quote the ask equal to your posterior value given that someone is buying, already incorporating the bad news that a buy conveys. Each execution then updates your beliefs by Bayes' Theorem, and quotes ratchet toward the informed traders' information. The spread $a - b = \mathbb{E}[V\mid \text{buy}] - \mathbb{E}[V\mid \text{sell}]$ is entirely the information difference between a buy and a sell.

Measuring it: effective vs. realized spread

Adverse selection is not just theory; you can read it off trades. Three spreads, all as half-spreads relative to the mid $m_t$ at trade time, with $q_t = \pm 1$ the trade sign:

• Effective (half-)spread: what the taker paid, $\;\text{ES}_t = q_t (P_t - m_t)$.
• Realized (half-)spread: what the maker kept after the value moved, using the mid a short time $\Delta$ later, $\;\text{RS}_t = q_t (P_t - m_{t+\Delta})$.
• Price impact (adverse selection): the permanent move the trade predicted, $\;\text{PI}_t = q_t (m_{t+\Delta} - m_t)$.

By construction these add up:

$\underbrace{q_t(P_t - m_t)}_{\text{effective}} \;=\; \underbrace{q_t(P_t - m_{t+\Delta})}_{\text{realized}} \;+\; \underbrace{q_t(m_{t+\Delta} - m_t)}_{\text{adverse selection}}.$

The maker's gross revenue is the effective spread, but after the mid drifts in the trade's direction (up after a buy, down after a sell), the maker only keeps the realized spread. The difference is the adverse-selection cost, the part of the spread that leaks straight back to informed counterparties. On liquid names the realized spread is often a small fraction of the effective spread; almost all of the quoted spread is compensation for adverse selection. This decomposition is the empirical heart of Bid-Ask Spread Decomposition.

Worked example

A market maker quotes 100 bid / 100.10 ask; mid is 100.05. A buy prints at the ask, 100.10, so $q_t = +1$ and the effective half-spread is \text{ES} = 100.10 - 100.05 = \0.05$. Ten seconds later the mid has drifted up to 100.09 because the buy was (partly) informed. Then the realized half-spread is$\text{RS} = 100.10 - 100.09 = $0.01$, and the adverse-selection component is$\text{PI} = 100.09 - 100.05 = $0.04$. The maker "earned" 5 cents but kept only 1: 80% of the quoted edge was adverse selection. Widen the quote and you keep more per trade but capture less flow; the equilibrium spread balances the two.

Failure modes and where it bites

• Toxic flow / order-flow toxicity. When the informed fraction $\theta$ spikes, around news, at the open, during a regime shift, realized spreads go negative and makers lose. Metrics like VPIN try to estimate $\theta$ in real time so makers can widen or pull quotes.
• The market-maker's death spiral. If adverse selection is severe enough, widening the spread drives away the uninformed (who won't pay it) faster than the informed (who still have an edge), raising $\theta$ further. Glosten–Milgrom show markets can break down entirely, no quote clears.
• Passive execution is adversely selected. Your resting limit orders fill mostly when the market is about to move against them (see Market vs. Limit Orders); "spread capture" measured gross is an illusion until you net out post-trade drift.
• Selection in your own fills. Any strategy that provides liquidity, from index rebalancing to option market-making, is short this option and must price it.

In interviews

Adverse selection is the concept interviewers use to see whether you think like a market maker. The framing that lands: "if someone wants to trade with me at my price, why is that bad news?", because informed traders trade only when your price is wrong for them, so a fill is evidence your price was wrong. Be able to state the regret-free quote condition $a = \mathbb{E}[V \mid \text{buy}]$, connect it to the winner's curse, and derive the effective = realized + adverse-selection identity. A classic prompt: "you're making a market on a coin's value and a genius who's seen the coin can trade with you, how wide do you quote?" You quote wide enough that the noise traders pay for what the genius takes; that is exactly The Glosten-Milgrom Model.