Adverse selection

The market maker’s core risk: a disproportionate share of your fills come from traders who know something you don’t. Glosten–Milgrom (1985) and Kyle (1985) show it is the reason spreads exist at all.

See it move

What is adverse selection?

Adverse selection is the risk that the counterparties who trade with a market maker are, on average, better informed than the maker. Because the informed trade in the direction the price is about to move, the maker's fills systematically precede losses. It is unavoidable: the maker cannot distinguish informed from uninformed flow until the price has already moved.

Picture yourself standing in the market, offering to buy at 99 and sell at 101. The people who choose to deal with you are not random. The one who buys from you at 101 just before the price jumps to 103 knew something; the one rebalancing a fund did not. You filled both at the same price, but only the first one cost you. Over many trades the informed ones are selected against you, hence "adverse" selection. This is the winner's curse of market making: the trades you win are disproportionately the ones you should have lost, because the very act of being filled is weak evidence the fill was bad.

Adverse selection is distinct from inventory risk (Market Making I). Inventory risk is price risk on the position you accumulate; it would exist even if all flow were uninformed, just from holding the asset while the market moves. Adverse selection is the risk that the flow itself is informationally toxic; it would exist even if you could flatten instantly, because the fill itself preceded the move. A real maker carries both at once.

How does adverse selection set the spread?

In the Glosten–Milgrom model (1985) the competitive market maker sets its ask to the expected value given the next order is a buy, and its bid to the expected value given a sell. The one-line gloss: the ask is what the asset is worth given that someone wants to buy it from you, and a buyer is, on average, slightly informed, so that posterior sits above the unconditional value.

The model puts a fraction $\mu$ of traders in the informed camp (they know the true value $v$, high or low, and trade accordingly) while the rest are uninformed and trade at random. A buy order is therefore more likely under the high value than the low one, so the posterior mean given a buy exceeds the prior. That gap is the half-spread on the ask side, and symmetrically on the bid. Because a buy is more likely to come from an informed high-value trader, the ask sits above the bid, and the whole spread is exactly the expected adverse-selection cost, with no inventory or fee component at all.

The spread widens with the informed fraction $\mu$ (this is the PIN, the probability of informed trading; see PIN & VPIN) and with the size of the value innovation, how far the price can jump if the informed are right. It narrows as the maker grows more confident about value. A maker facing 60% informed flow must quote roughly twice as wide as one facing 30%, all else equal. The deep point: this spread exists with zero inventory cost and zero fees, purely to recoup information asymmetry. It is the irreducible floor under any market's spread; inventory and fees stack on top, but you can never quote below the adverse-selection floor and survive.

Toxic versus benign flow

Benign flow is uninformed (liquidity trades, rebalances, retail noise) and it is the market maker's profit: you earn the spread and the price does not move against you. Toxic flow is informed: it precedes adverse moves, so each toxic fill costs you the value move minus the spread. The whole craft is earning enough on benign flow to cover the toxic flow you cannot avoid.

The maker's P&L is a tug-of-war: a + half-spread on every fill against a − value-move on every informed fill. Profit per benign fill is the half-spread; loss per toxic fill is roughly the information edge minus the half-spread. The book is profitable only if the benign volume earns more than the toxic volume bleeds.

You cannot identify toxicity at fill time. That is the entire difficulty. A toxic order looks identical to a benign one as it arrives; you only learn which it was when the price does (or does not) move against you afterwards. So you cannot refuse the toxic fills selectively; you can only price as if a known fraction of flow is toxic, and adjust that fraction as conditions change. What you can do is estimate toxicity in real time and re-quote. The downstream tools in these guides are all toxicity estimators feeding back into your spread and your fair value: order-flow imbalance flags when flow is one-sided and often informed; PIN & VPIN give a running estimate of the informed fraction; and the microprice is a fair value that already leans toward where imbalance says the price is going, so you are picked off less around it.

When toxicity spikes (a news event, a venue dislocation) the right response is to widen and skew, or pull quotes. Better to miss benign flow than to keep feeding informed flow at a stale spread. Makers that do not widen fast enough during toxic bursts are how a quiet book turns into a flash event loss.

Why can't you just widen your way out?

Widening the spread cuts adverse-selection cost per fill, but it also kills your fill rate, because the benign flow that pays you walks to a tighter competitor. There is an optimal spread: wide enough to survive the informed, tight enough to still win the uninformed. Quote too tight and you bleed to toxic flow; too wide and you earn nothing.

The trade-off is exactly the one Avellaneda–Stoikov formalises on the inventory side, mirrored here on the information side: the wider you quote, the lower your fill intensity $\lambda(\delta)$ (fewer benign fills, less spread earned) but the lower your per-fill adverse-selection cost. The optimum is interior, neither zero nor infinite spread.

In a competitive market the spread is pinned near the adverse-selection floor: quote above it and a rival quotes between you and the touch to take the benign flow; quote below it and you bleed. So the market spread reflects the consensus informed fraction. Your edge is being right about toxicity when the consensus is wrong: quoting tight when flow is actually benign, pulling fast when it is actually toxic. This is why the durable 2026 edge is estimation, not blunt widening. A maker with a better toxicity model can run a tighter average spread than a rival, winning more benign volume, while avoiding more toxic fills through faster re-quoting. Same average spread on the screen, very different realised P&L: the theme of the order-flow & information guides.

Worked example

Take a synthetic market-making session driven by adverse selection, as of 2026. Reproduce it by moving the sliders in the interactive above. The value $v$ is $100 \pm 1$, so $v_H = 101$ and $v_L = 99$ are equally likely and the unconditional value is $\mathbb{E}[v] = 100$. The informed fraction is $\mu = 30\%$; the rest is noise, 50/50.

With $P(\text{buy}\mid v_H) = 0.65$ and $P(\text{buy}\mid v_L) = 0.35$, a buy implies $P(v_H \mid \text{buy}) = 0.65$, so the ask and bid are the two posterior means and the break-even spread is their gap.

A benign fill. A noise trader buys from you at $a = 100.30$. The value turns out to be $v_L = 99$: they were uninformed and wrong-way. You are short at 100.30 against a true 99, so you make $+1.30$ when it settles. On average across noise fills you earn the half-spread, $+0.30$.

A toxic fill. An informed trader buys from you at $a = 100.30$. They only buy when $v = 101$. You are short at 100.30 against a true 101, so you lose $-0.70$, the value move of 1.00 minus the half-spread of 0.30. Every informed fill costs you 0.70 at this spread.

The book nets out. Per 100 incoming buys at $\mu = 30\%$, about 30 are informed ($-0.70$ each, $-21.0$) and about 70 are noise, contributing roughly $+0.30$ each ($+21.0$) once the right-way and wrong-way noise fills average out. The net is approximately zero at the competitive break-even spread, exactly the zero-profit condition. A real maker quotes 0.70 instead of 0.60 and keeps about $+0.10$ per fill, but only if its 30% estimate is right.

Dial toxicity up to $\mu = 60\%$ (the "toxic" preset) and the adverse-selection bar in the P&L panel grows; if you keep quoting the old 0.70 spread you go firmly net-negative. That is the visual lesson that flow toxicity, not inventory, is what blows up a maker on event days. The numbers here are synthetic and rounded; informed fractions, value gaps and break-even spreads must be estimated per instrument and dated. The estimation problem (knowing $\mu$ before the price tells you) is the real job, and it is what OFI, VPIN and the microprice attack.