Kyle's Lambda: Price Impact and Strategic Trading

Kyle (1985): Strategic Trading and Price Impact

glosten-milgrom-model shows that the spread exists because of adverse selection. But it treats informed traders as mechanical — they simply buy when $V$ is high and sell when low. Kyle (1985) asks the next question: what if the informed trader is strategic?

The answer produces the single most important parameter in execution analysis: $\lambda$, the price impact coefficient.

The Three Agents

Kyle’s model has exactly three participants:

1. The insider: knows the true asset value $V$ with certainty. Submits a market order of size $x$.
2. Noise traders: submit random aggregate order flow $u\sim \mathcal{N}(0,{\sigma }_u^2)$. They trade for liquidity reasons unrelated to $V$.
3. The market maker: observes total order flow $y=x+u$ but cannot distinguish insider flow from noise. Sets price $P$ to break even in expectation.

The true value is $V\sim \mathcal{N}(\mu ,{\sigma }_v^2)$, independent of $u$.

The Linear Equilibrium

Kyle shows there exists a unique linear equilibrium with two components:

The insider’s optimal strategy

$x=\beta (V-\mu )$

The insider trades proportionally to their information advantage $(V-\mu )$. The coefficient $\beta$ governs how aggressively they trade.

The market maker’s pricing rule

$P=\mu +\lambda \cdot y$

The price is the prior mean plus a linear adjustment for observed order flow. The coefficient $\lambda$ is the price impact — how much the price moves per unit of net order flow.

Solving for $\lambda$ and $\beta$

The market maker sets $P=E[V\mid y]$. Since $V$ and $u$ are jointly normal and $y=\beta (V-\mu )+u$:

$E[V\mid y]=\mu +\frac{\text{Cov}(V,y)}{\text{Var}(y)}\cdot y$

Computing each term:

$\text{Cov}(V,y)=\text{Cov}(V,\beta (V-\mu )+u)=\beta {\sigma }_v^2$

$\text{Var}(y)={\beta }^2{\sigma }_v^2+{\sigma }_u^2$

Therefore:

$\lambda =\frac{\beta {\sigma }_v^2}{{\beta }^2{\sigma }_v^2+{\sigma }_u^2}$

The insider maximizes expected profit $E[x\cdot (V-P)]$:

$E[\beta (V-\mu )(V-\mu -\lambda (\beta (V-\mu )+u))]$ $=\beta (1-\lambda \beta ){\sigma }_v^2$

First-order condition in $\beta$:

$(1-2\lambda \beta ){\sigma }_v^2=0⟹\beta =\frac{1}{2\lambda }$

Substituting back and solving the system of two equations:

$\boxed{\lambda =\frac{{\sigma }_v}{2{\sigma }_u}}$

$\beta =\frac{{\sigma }_u}{{\sigma }_v}$

Interpreting $\lambda$

The price impact $\lambda =\frac{{\sigma }_v}{2{\sigma }_u}$ has clean comparative statics:

• $\lambda$ increases with ${\sigma }_v$: more uncertainty about the true value means each unit of order flow is more informative. The maker must adjust prices more aggressively.
• $\lambda$ decreases with ${\sigma }_u$: more noise trading provides cover for the insider. The maker cannot distinguish signal from noise, so each unit of flow moves the price less.

The ratio $\frac{{\sigma }_v}{{\sigma }_u}$ is the signal-to-noise ratio of order flow. High signal-to-noise means high price impact.

Key Results

1. The insider trades gradually

With $\beta =\frac{{\sigma }_u}{{\sigma }_v}$, the insider scales their trade to the noise level. More noise means larger trades (more cover). The insider never reveals all information at once — they optimally spread their trades to avoid moving the price too much.

2. Half the information is incorporated

In this single-period model, the price after trading incorporates exactly half of the insider’s private information:

$\text{Var}(V\mid y)=\frac{1}{2}{\sigma }_v^2$

The remaining half is “saved” — in the multi-period version, the insider spreads information incorporation across all trading rounds.

3. Noise traders are essential

Without noise (${\sigma }_u=0$), $\lambda \to \infty$ and the market collapses. The maker would perfectly infer $V$ from any trade, so the insider cannot profit, so they do not trade, so there is no price discovery. Noise traders are not parasites — they are the necessary substrate for markets to function.

4. Insider profit

$E[\text{profit}]=\frac{1}{2}{\sigma }_v{\sigma }_u$

Profit increases with both ${\sigma }_v$ (more information to exploit) and ${\sigma }_u$ (more cover to exploit it behind).

From Kyle to Execution Algorithms

Kyle’s $\lambda$ is the foundation of modern execution cost analysis. When an institution needs to sell 1M shares, the execution desk faces exactly the Kyle problem: their order will move the price, and the price impact is roughly linear in order size (for moderate sizes).

The VWAP, TWAP, and implementation shortfall algorithms all attempt to minimize $\lambda$-mediated impact by:

• Splitting orders across time (spreading like Kyle’s insider)
• Trading during high-volume periods (when ${\sigma }_u$ is large)
• Randomizing timing (to look like noise)

Almgren-Chriss (2001) formalized this as an optimal control problem: trade off the urgency cost (risk of price drift while waiting) against the impact cost ($\lambda$ per unit traded).

Connection to AMMs: Convex vs. Linear Price Impact

In Kyle’s model, price impact is linear: $\Delta P=\lambda \cdot \Delta y$. In a constant-product AMM, price impact is convex: the marginal price moves faster as trade size increases relative to pool reserves.

For a pool with reserves $(R_x,R_y)$ and constant product $k=R_x\cdot R_y$, the effective price impact of buying $\Delta x$ units is:

$P_{\text{effective}}=\frac{R_y}{R_x-\Delta x}>\frac{R_y}{R_x}=P_{\text{marginal}}$

This convexity means:

• Small trades experience approximately linear impact (resembling Kyle)
• Large trades face rapidly increasing impact — a natural defense against informed trading, but also a cost for legitimate large liquidity needs
• Unlike Kyle’s maker, the AMM cannot modulate $\lambda$ — it is hardcoded into the bonding curve

The implication: AMMs are too generous to small informed trades (low impact) and too punitive to large uninformed trades (high impact), compared to what a Kyle-rational market maker would offer.

---

Companion notebook: notebooks/market-microstructure/05-kyle-lambda.py — simulate the Kyle equilibrium, estimate $\lambda$ from order flow data, compare linear vs. convex price impact schedules.

Questions to sit with:

1. Kyle assumes the insider knows $V$ with certainty. What changes if the insider has a noisy signal instead? Does $\lambda$ increase or decrease?
2. In crypto markets, on-chain order flow is visible before execution (mempool). How does this change the insider’s optimal strategy?
3. If you were designing an AMM, how would you introduce an adaptive $\lambda$ that responds to estimated informed trading fraction?