Glosten-Milgrom Zero-Profit Bid-Ask Quotes
A binary asset has fundamental value $V \in \{0, 1\}$ with prior $P(V = 1) = p$, where $p \in (0, 1)$.
A single counterparty arrives to trade one unit. With probability $1 - q$ the counterparty is perfectly informed (buys if $V = 1$, sells if $V = 0$). With probability $q$ the counterparty is a noise trader who buys or sells each with probability $\frac{1}{2}$.
You are the market maker. You must post a bid $b$ and an ask $a$ (unit size) such that your expected profit is exactly zero conditional on being hit on that side.
- Derive the zero-profit ask $a(p, q)$ and bid $b(p, q)$.
- Compute the implied spread $a - b$ and simplify.
- Interpret the limiting cases $q \to 0$ (all informed) and $q \to 1$ (all noise).
Hints
- The zero-profit condition means each quote equals the expected value of $V$ conditional on being hit on that side. What does a buy signal tell you about whether the counterparty is informed?
- Write out $P(\text{buy} \mid V = 1)$ and $P(\text{buy} \mid V = 0)$ by mixing over the informed and noise components, then apply Bayes' theorem to get $P(V = 1 \mid \text{buy})$.
- For the spread simplification, note that the two denominators $D_a = 2p + q - 2pq$ and $D_b = 2 - 2p - q + 2pq$ sum to 2. Cross-multiply and collect terms to get $4p(1-p)(1-q)$ in the numerator.
Worked Solution
How to Think About It: You are a market maker posting quotes against a mix of informed and uninformed flow. The informed traders pick you off -- they only buy when the asset is worth 1 and only sell when it is worth 0. Noise traders are your profit center since they trade randomly. The zero-profit condition means you set each quote equal to the expected value of the asset conditional on being hit on that side. This is just Bayes' theorem: a buy tells you it is more likely the counterparty is informed and $V = 1$, while a sell tells you $V = 0$ is more likely. The ask will be above $p$ (adverse selection on the buy side) and the bid will be below $p$ (adverse selection on the sell side), with the gap driven by how much informed flow there is.
Quick Estimate: Take $p = 0.5$, $q = 0.5$ (half informed, half noise, fair coin prior). If someone buys, the informed component pushes $P(V=1)$ up. By symmetry the ask should be above 0.5 and the bid below by the same amount. Rough guess: $a \approx 0.7$, $b \approx 0.3$, spread $\approx 0.4$. As we will see, the exact answer gives $a = 3/4$, $b = 1/4$, spread $= 1/2$.
Approach: Apply Bayes' theorem separately to the buy event and the sell event, then set each quote equal to the posterior expectation of $V$.
Formal Solution:
*Step 1: Ask quote (zero-profit on the sell side of the book).*
When someone buys from you, compute the conditional probabilities:
- $P(\text{buy} \mid V = 1) = (1 - q) \cdot 1 + q \cdot \tfrac{1}{2} = 1 - \tfrac{q}{2}$
- $P(\text{buy} \mid V = 0) = (1 - q) \cdot 0 + q \cdot \tfrac{1}{2} = \tfrac{q}{2}$
By Bayes' theorem:
$$P(V = 1 \mid \text{buy}) = \frac{p\,(1 - q/2)}{p\,(1 - q/2) + (1 - p)\,q/2} = \frac{p(2 - q)}{2p + q - 2pq}$$
The zero-profit ask is $a = E[V \mid \text{buy}] = P(V = 1 \mid \text{buy})$:
$$\boxed{a(p, q) = \frac{p(2 - q)}{2p + q - 2pq}}$$
With $p = q = \tfrac{1}{2}$: $P(\text{buy}\mid V=1) = 1 - \tfrac{1}{4} = \tfrac{3}{4}$ and $P(\text{buy}\mid V=0) = \tfrac{1}{4}$, so
$$a = \frac{\tfrac{1}{2}\cdot\tfrac{3}{4}}{\tfrac{1}{2}\cdot\tfrac{3}{4} + \tfrac{1}{2}\cdot\tfrac{1}{4}} = \frac{3/8}{3/8 + 1/8} = \frac{3/8}{1/2} = \frac{3}{4}.$$
Equivalently from the formula, $a = \frac{(1/2)(3/2)}{1 + 1/2 - 1/2} = \frac{3/4}{1} = \frac{3}{4}$.
*Step 2: Bid quote (zero-profit on the buy side of the book).*
When someone sells to you:
- $P(\text{sell} \mid V = 1) = (1 - q) \cdot 0 + q \cdot \tfrac{1}{2} = \tfrac{q}{2}$
- $P(\text{sell} \mid V = 0) = (1 - q) \cdot 1 + q \cdot \tfrac{1}{2} = 1 - \tfrac{q}{2}$
By Bayes' theorem:
$$P(V = 1 \mid \text{sell}) = \frac{p\,q/2}{p\,q/2 + (1 - p)(1 - q/2)} = \frac{pq}{pq + (1 - p)(2 - q)}$$
The zero-profit bid:
$$\boxed{b(p, q) = \frac{pq}{pq + (1 - p)(2 - q)}}$$
With $p = q = \tfrac{1}{2}$: $b = \frac{(1/2)(1/2)}{(1/2)(1/2) + (1/2)(3/2)} = \frac{1/4}{1/4 + 3/4} = \frac{1/4}{1} = \frac{1}{4}$.
*Step 3: Implied spread.*
Let $D_a = 2p + q - 2pq$ and $D_b = pq + (1 - p)(2 - q) = 2 - 2p - q + 2pq$. Note $D_a + D_b = 2$.
$$a - b = \frac{p(2 - q)}{D_a} - \frac{pq}{D_b}$$
Putting over the common denominator $D_a D_b$ and expanding the numerator:
$$p(2 - q) D_b - pq \, D_a = p\bigl[(2 - q)(2 - 2p - q + 2pq) - q(2p + q - 2pq)\bigr]$$
Expanding and collecting terms, the bracket simplifies to $4(1 - p)(1 - q)$, so:
$$\boxed{a - b = \frac{4p(1 - p)(1 - q)}{(2p + q - 2pq)(2 - 2p - q + 2pq)}}$$
With $p = q = \tfrac{1}{2}$: $a - b = \frac{4(1/2)(1/2)(1/2)}{(1)(1)} = \frac{1/2}{1} = \frac{1}{2}$, matching $a - b = \tfrac{3}{4} - \tfrac{1}{4} = \tfrac{1}{2}$.
*Step 4: Limiting cases.*
$q \to 0$ (all informed): $a \to \frac{2p}{2p} = 1$ and $b \to \frac{0}{(1-p) \cdot 2} = 0$. The spread goes to 1, the maximum possible. Every counterparty is informed, so the market maker must quote the widest possible spread to avoid being picked off. Trading effectively shuts down.
$q \to 1$ (all noise): $a \to \frac{p \cdot 1}{0 + 1} = p$ and $b \to \frac{p}{p + 0} = p$. The spread collapses to zero. With no informed traders, a buy or sell carries no information about $V$, so both quotes equal the prior $p$. The market is perfectly liquid.
Answer: The zero-profit quotes are $a(p, q) = \frac{p(2 - q)}{2p + q - 2pq}$ and $b(p, q) = \frac{pq}{pq + (1-p)(2-q)}$. The spread is $\frac{4p(1-p)(1-q)}{(2p+q-2pq)(2-2p-q+2pq)}$. For $p = q = \tfrac{1}{2}$ this gives $a = \tfrac{3}{4}$, $b = \tfrac{1}{4}$, spread $= \tfrac{1}{2}$. The spread ranges from 0 (pure noise) to 1 (pure informed), and is maximized in the prior at $p = 1/2$ for any fixed $q$.
Intuition
This is the foundational model of why bid-ask spreads exist: adverse selection. The market maker faces a mix of counterparties -- some who know more than she does and some who trade for reasons unrelated to value. Every time she is hit, she must ask: is this person trading because they know something, or is it just noise? The spread is the price she charges for bearing that uncertainty. It is not a transaction cost in the usual sense -- it is an information rent extracted by informed traders from the market maker, passed through to noise traders via wider quotes.
The key quantitative insight is that the spread is proportional to $(1 - q)$, the fraction of informed flow, and to $p(1 - p)$, the prior uncertainty about the asset. If you already know the asset is worth 1 (or 0), informed traders cannot hurt you, so the spread vanishes. If half the traders are informed and you have maximum uncertainty ($p = 1/2$), the spread is widest. This is why you see spreads blow out around earnings announcements (high $1 - q$, high $p(1-p)$) and tighten for heavily traded blue chips (low $1-q$ relative to noise volume). In practice, market makers estimate $q$ from flow toxicity metrics like VPIN, and the Glosten-Milgrom framework is the theoretical backbone behind those tools.