Asymmetric Market Making with Adverse Selection

Market Microstructure · Hard · Free problem

A single asset has true value $V \sim N(0, \sigma^2)$ with $\sigma^2 = 2$. Each period, one trader arrives:

  • With probability $q$, the trader is informed and observes $V$. The informed trader buys one unit if $V > a$, sells one unit if $V < b$, and passes otherwise.
  • With probability
    - q$, the trader is noise -- buys or sells one unit with equal probability $\tfrac{1}{2}$, independent of $V$.

You are the market maker. You quote an ask $a > 0$ and a bid $b < 0$ (one unit each). Noise traders always execute at your posted quote.

  1. Derive the market maker's expected one-period P&L $\pi(a, b)$.
  2. Find the optimal quotes $(a^{*}, b^{*})$ that maximize $\pi(a, b)$.
  3. State the first-order conditions and show that the optimum is unique by verifying concavity of $\pi$ in $(a, b)$.

Hints

  1. Decompose the expected P&L by trader type -- noise traders give you a clean spread profit, informed traders impose an adverse selection cost. What are the conditional expectations $E[V \mid V > a]$ and $E[V \mid V < b]$ for a normal distribution?
  2. Use the identity $E[V \cdot \mathbf{1}_{V > a}] = \sigma\,\phi(a/\sigma)$ for $V \sim N(0,\sigma^2)$. After writing $\pi(a,b)$, check whether it separates into independent functions of $a$ and $b$.
  3. Differentiate and notice an exact cancellation: $g'(a) = \frac{1-q}{2} + q(1-\Phi(a/\sigma)) > 0$ always. This means no interior maximum exists for the monopolist -- the competitive zero-profit condition $g(a^{*}) = 0$ is the economically relevant solution.

Worked Solution

How to Think About It: You are a market maker posting a bid and an ask. You make money off noise traders (they trade regardless of value) and lose money to informed traders (they cherry-pick you when the price is wrong). The spread is your compensation for adverse selection -- wider spreads reduce informed losses but you still collect from noise flow. Before doing any math, think about what the answer should look like: since $V$ is symmetric around 0, the optimal quotes should satisfy $a^{*} = -b^{*}$, i.e., the spread is symmetric. The ask should be positive (you charge a premium to sell) and the bid negative (you demand a discount to buy). The key tension is: wider quotes reduce adverse selection losses but the noise profit grows linearly in the spread -- so we need to figure out exactly what constrains quote-widening.

Quick Estimate: With $\sigma = \sqrt{2} \approx 1.41$ and say $q = 0.3$, a reasonable ask might be around $a \approx 0.8$. The informed buy probability is $P(V > 0.8) = 1 - \Phi(0.8/\sqrt{2}) \approx 1 - \Phi(0.57) \approx 0.29$. The conditional expected value given $V > 0.8$ is roughly $0.8 + \sigma \cdot \phi(0.57)/(1-\Phi(0.57)) \approx 0.8 + 1.41 \cdot 0.34/0.29 \approx 2.45$. Noise profit from ask side: $(0.7/2) \times 0.8 = 0.28$. Informed loss: $0.3 \times (2.45 - 0.8) \times 0.29 \approx 0.14$. Net from ask: $\approx 0.14$. This is positive, and it would be even larger for wider spreads, suggesting the monopolist wants to widen the spread as much as possible.

Approach: We derive $\pi(a,b)$ in closed form. An important structural result emerges: with perfectly inelastic noise demand, $\pi$ separates into independent functions of $a$ and $b$, each monotonically increasing in the spread width. This means no interior maximum exists for a monopolist. The economically meaningful solution is the competitive (zero-profit) equilibrium where Bertrand competition among market makers drives $a^{*} = E[V \mid \text{buy}]$ and $b^{*} = E[V \mid \text{sell}]$. We derive both.

Formal Solution:

Let $\phi(\cdot)$ and $\Phi(\cdot)$ denote the standard normal PDF and CDF. For $V \sim N(0, \sigma^2)$, the density is $f(v) = \frac{1}{\sigma}\phi\!\left(\frac{v}{\sigma}\right)$.

Part 1: Deriving $\pi(a,b)$.

When the market maker sells at $a$, the P&L is $a - V$. When the market maker buys at $b$, the P&L is $V - b$.

*Noise trader contribution.* With probability

-q$, a noise trader buys (prob $\frac{1}{2}$) or sells (prob $\frac{1}{2}$). Since noise trades are independent of $V$ and $E[V]=0$:

$\pi_{\text{noise}} = \frac{1-q}{2}\,a + \frac{1-q}{2}(-b) = \frac{1-q}{2}(a - b)$

*Informed trader contribution.* With probability $q$, an informed trader arrives.

  • Informed buy (when $V > a$): Market maker sells, P&L $= a - V < 0$.

$\pi_{\text{inf,ask}} = q\int_a^{\infty}(a - v)\,f(v)\,dv = q\bigl[a(1 - \Phi(a/\sigma)) - \sigma\,\phi(a/\sigma)\bigr]$

where we used the identity $E[V \cdot \mathbf{1}_{V > a}] = \sigma\,\phi(a/\sigma)$ for $V \sim N(0,\sigma^2)$.

  • Informed sell (when $V < b$): Market maker buys, P&L $= V - b < 0$.

$\pi_{\text{inf,bid}} = q\int_{-\infty}^{b}(v - b)\,f(v)\,dv = q\bigl[-\sigma\,\phi(b/\sigma) - b\,\Phi(b/\sigma)\bigr]$

using $E[V \cdot \mathbf{1}_{V < b}] = -\sigma\,\phi(b/\sigma)$.

Combining all terms:

$\pi(a,b) = \frac{1-q}{2}(a-b) + q\bigl[a(1-\Phi(a/\sigma)) - \sigma\phi(a/\sigma)\bigr] + q\bigl[-\sigma\phi(b/\sigma) - b\,\Phi(b/\sigma)\bigr]$

This separates as $\pi(a,b) = g(a) + h(b)$ where:

$g(a) = a\!\left[\frac{1-q}{2} + q(1 - \Phi(a/\sigma))\right] - q\sigma\,\phi(a/\sigma)$

$h(b) = -b\!\left[\frac{1-q}{2} + q\,\Phi(b/\sigma)\right] - q\sigma\,\phi(b/\sigma)$

The separability is key: the optimal ask and bid can be chosen independently.

Part 2: First-order conditions and the monopolist's problem.

Differentiate $g(a)$ using $\frac{d}{da}\Phi(a/\sigma) = \frac{1}{\sigma}\phi(a/\sigma)$ and $\frac{d}{da}\phi(a/\sigma) = -\frac{a}{\sigma^2}\phi(a/\sigma)$:

$g'(a) = \frac{1-q}{2} + q(1-\Phi(a/\sigma)) - \frac{qa}{\sigma}\phi(a/\sigma) + \frac{qa}{\sigma}\phi(a/\sigma) = \frac{1-q}{2} + q(1-\Phi(a/\sigma))$

The terms $-\frac{qa}{\sigma}\phi(a/\sigma)$ (from the product rule on $a \cdot (1-\Phi)$) and $+\frac{qa}{\sigma}\phi(a/\sigma)$ (from differentiating $-q\sigma\phi(a/\sigma)$) cancel exactly. Since $\frac{1-q}{2} > 0$ and $q(1-\Phi(a/\sigma)) > 0$ for all finite $a$, we have $g'(a) > 0$ for all $a$.

Similarly, $h'(b) = -\frac{1-q}{2} - q\,\Phi(b/\sigma) < 0$ for all $b$, so $h$ is decreasing in $b$ (making $b$ more negative increases profit).

Critical observation: With perfectly inelastic noise demand, $\pi(a,b)$ is strictly increasing in $a$ and strictly decreasing in $b$. There is no interior maximum -- a monopolist market maker would set the spread infinitely wide. This is a fundamental result: when noise traders are price-insensitive, the monopolist extracts unlimited rents.

Part 3: The competitive equilibrium (Glosten-Milgrom zero-profit condition).

In practice, Bertrand competition among market makers drives profit to zero. The competitive ask satisfies $a^{*} = E[V \mid \text{someone buys at } a^{*}]$ and the competitive bid satisfies $b^{*} = E[V \mid \text{someone sells at } b^{*}]$.

The probability someone buys at the ask is:

$P(\text{buy}) = \frac{1-q}{2} + q(1 - \Phi(a/\sigma))$

The expected value of $V$ conditional on a buy:

$E[V \mid \text{buy}] = \frac{q \cdot E[V \cdot \mathbf{1}_{V>a}] + \frac{1-q}{2} \cdot E[V]}{P(\text{buy})} = \frac{q\,\sigma\,\phi(a/\sigma)}{\frac{1-q}{2} + q(1-\Phi(a/\sigma))}$

The zero-profit condition $a = E[V \mid \text{buy}]$ gives:

$a\left[\frac{1-q}{2} + q(1-\Phi(a/\sigma))\right] = q\,\sigma\,\phi(a/\sigma)$

This is equivalent to $g(a) = 0$. Note $g(0) = -q\sigma\phi(0) < 0$ and $g(a) \to +\infty$ as $a \to \infty$ (since $g'(a) > 0$). By the intermediate value theorem and strict monotonicity, there is a unique $a^{*} > 0$ solving $g(a^{*}) = 0$.

By the symmetry of $N(0, \sigma^2)$, the bid satisfies $b^{*} = -a^{*}$, giving a symmetric spread.

Concavity in the competitive setting. To verify this is a proper equilibrium, consider the market maker's profit when deviating from the competitive quotes. If a market maker undercuts (sets $a < a^{*}$), she attracts more informed adverse flow and $g(a) < 0$: she loses money. If she sets $a > a^{*}$, she is overcut by competitors and gets no flow. Thus $a^{*}$ is the unique equilibrium ask in the competitive game. Formally, the second-order condition on $g$ evaluated at the zero-profit point confirms it is a crossing from below (since $g'(a^{*}) > 0$, $g$ crosses zero from negative to positive), ensuring uniqueness.

Numerical example. With $\sigma^2 = 2$ and $q = 0.3$, solve $g(a) = 0$ numerically. The condition is:

$a\left[0.35 + 0.3(1-\Phi(a/\sqrt{2}))\right] = 0.3\sqrt{2}\,\phi(a/\sqrt{2})$

Numerical solution gives $a^{*} \approx 0.52$, so the competitive spread is approximately $a^{*} - b^{*} \approx 1.04$, or about $0.74\sigma$.

Answer:

The expected P&L is: $\pi(a,b) = \frac{1-q}{2}(a-b) + q[a(1-\Phi(a/\sigma)) - \sigma\phi(a/\sigma)] + q[-\sigma\phi(b/\sigma) - b\,\Phi(b/\sigma)]$

This separates into $g(a) + h(b)$ with $g'(a) > 0$ and $h'(b) < 0$, so a monopolist sets an infinite spread. Under competition, the unique equilibrium is the zero-profit condition $a^{*} = E[V|\text{buy at }a^{*}]$, $b^{*} = E[V|\text{sell at }b^{*}]$, with $b^{*} = -a^{*}$ by symmetry. The zero-profit ask solves $a[\frac{1-q}{2} + q(1-\Phi(a/\sigma))] = q\sigma\phi(a/\sigma)$, which has a unique positive root.

Intuition

This problem is a clean version of the Glosten-Milgrom (1985) adverse selection model and it teaches the single most important idea in market microstructure: the bid-ask spread exists to compensate the market maker for being picked off by informed traders. Noise traders lose money on average (they pay the spread), and that revenue offsets the losses from selling cheap to someone who knows the asset is worth more, or buying dear from someone who knows it is worth less. The deeper lesson is structural: with perfectly inelastic noise demand, a monopolist market maker would widen the spread without limit. Only competition -- the threat of being undercut -- pins the quotes to the zero-profit Glosten-Milgrom condition $a = E[V|\text{buy}]$. This is why real markets need competition or regulation to keep spreads tight.

The normal distribution plays a specific role here. The exact cancellation in $g'(a)$ -- where the marginal adverse selection cost of widening the ask perfectly offsets the marginal decrease in informed trade frequency -- is a consequence of the normal tail structure. The partial expectation identity $E[V \cdot \mathbf{1}_{V>a}] = \sigma\phi(a/\sigma)$ is doing all the heavy lifting. In practice, asset value distributions have fatter tails than normal, which makes adverse selection worse and spreads wider. This is why you see wider spreads in illiquid, volatile, or information-sensitive names -- the $q$ is effectively higher and the tails are fatter.

Open the full interactive solver →