2.2.4 The Efficient Frontier of Execution and the Optimal Trajectory

Basic Prerequisites: 03 - The Almgren–Chriss Model.

---

1. Intuition & Practical Objective

There is no single “best” execution schedule - only a frontier. For every level of risk you are willing to accept, there is a cheapest way to trade; for every level of cost, there is a lowest-risk way. Plot those best-possible pairs and you have the efficient frontier of execution (Almgren–Chriss 2000, §2). Picking $\lambda$ is then a client decision, and the trajectory follows: high $\lambda$ (impatient) sells fast and pays impact; low $\lambda$ (patient) sells slowly and carries risk.

This page has two deliverables: (a) how to construct the frontier and read off a strategy for a given $\lambda$, and (b) how to describe/compare trajectories using the half-life $\theta$ and the shape of $x_t$. The one operational rule is the tangency condition: choosing $\lambda$ selects the frontier point where the curve’s slope equals $-\lambda$.

The one-sentence essence. “The frontier is the lower envelope of attainable $(V,E)$ pairs - convex, with the risk-neutral TWAP point at its cost-minimal end - and minimizing $E+\lambda V$ is exactly drawing a line of slope $-\lambda$ tangent to that curve.”

---

2. Mathematical Ground Truth & Derivations

The frontier (AC §2.1, §2.4). For fixed $X,T,\sigma ,\eta$ the family $\{x^{\star }(\lambda ):\lambda \ge 0\}$ traces a curve $(V(\lambda ),E(\lambda ))$ in the (variance, cost) plane. Its structure:

• Cost-minimal end ($\lambda =0$): the “naive” strategy, $x_t=X(1-t/T)$, minimizing $E$ while taking maximum variance. This is TWAP.
• Variance-minimal end ($\lambda \to \infty$): liquidate at $t=0$; $V\to 0$, $E\to \infty$ (in the continuous limit).
• Convexity: $d^{2}E/d{V}^2>0$ - the frontier is strictly convex, so each $\lambda$ selects a unique point, and the selection is the tangency condition

$$\frac{dE}{dV}{|}_{\lambda }=-\lambda .$$

The Lagrange multiplier $\lambda$ is literally (minus) the slope of the frontier at the chosen point. AC §3.1 identify it with the Arrow-Pratt absolute risk aversion ${\lambda }_u=-u^{\prime \prime }(w)/u^{\prime }(w)$ of a quadratic utility, giving $U_{\text{util}}(x)={\lambda }_{u}V(x)+E(x)$ - so “pick a point” and “maximize utility” are the same act.

Trajectory invariants. Three numbers summarize any AC trajectory:

Symbol	Meaning	Formula	This example
$\kappa$	urgency (inverse half-life)	$\sqrt{\lambda {\sigma }^2/\tilde{\eta }}$	$0.6011/\text{day}$
$\theta =1/\kappa$	trade half-life ($e$-folding)	$\sqrt{\tilde{\eta }/(\lambda {\sigma }^2)}$	$1.664$ days
$\kappa T$	constraint regime	$T/\theta$	$3.006$

Reading $\kappa T$ (AC §2.3). $\kappa T\ll 1$: the horizon binds, temporary impact dominates, the curve approaches TWAP. $\kappa T\gg 1$: risk dominates, the curve approaches the block and most liquidation occurs in the first fraction of $T$. There is no dependence on $X$: with linear impact, all basket sizes of the same stock trade on the same time scale - a direct consequence of impact and variance both scaling as $X^2$.

Closed-form frontier (AC eq 20). The pair $(E(\lambda ),V(\lambda ))$ along the frontier is available in closed form from the $\sinh$ trajectory; the practical route (used below) is simply to evaluate $E$ and $V$ numerically on the trajectory for a grid of $\lambda$ - no approximation, complete agreement with AC’s Fig. 1.

---

3. Computational Implementation - building and reading the frontier

Sweep $\lambda$, compute each optimal trajectory, evaluate $(E,V)$, verify convexity and the tangency condition. numpy + stdlib.

Three verified facts from AC §2: (i) the frontier is convex; (ii) each $\lambda$ selects the tangent point exactly; (iii) $\lambda =0$ reproduces TWAP ($E=$ $644,500, sd = $1,222,765). Note how cheap the first risk reduction is - moving from $\lambda =0$ to $\lambda =10^{-7}$ cuts the standard deviation by $67k for only $7.7k of extra expected cost. This is AC’s point that the naïve (TWAP) strategy is never efficient: a small cost buys a large variance reduction.

---

4. Failure Modes & First-Principles Breakdowns

1. The frontier is a model output, not a market fact. Every point assumes $\sigma ,\eta ,\gamma$ are known and constant. Change $\eta$ by 2x and the whole curve shifts; the “efficient” trajectory you chose may be dominated under the true parameters.
2. $\lambda$ is not directly observable. It is a preference, and it is easy to under- or over-state. Because $\theta \propto 1/\sqrt{\lambda }$, a 4x $\lambda$ error is a 2x half-life error - a first-order mistake in how the order is worked.
3. Variance is the wrong risk measure for tail events. The frontier assumes Gaussian shortfall. On a jump or liquidity-crisis path the relevant risk is a quantile/expected-shortfall, not the second moment - AC themselves add the L-VaR (quantile) frontier for exactly this reason (AC §3.2).
4. Convexity is an artifact of convex costs. With concave or non-convex impact (capacity limits, block discounts) the frontier can become non-convex, breaking the unique-tangency selection and admitting randomized (“mixed”) strategies.
5. A convex frontier does not mean a stable trajectory. The whole construction is ex-ante; the realized path wanders (that is what $V$ measures). Do not read the smooth curve as a prediction of realized cost.

---

5. References

• Almgren, Robert; Chriss, Neil - “Optimal execution of portfolio transactions,” Journal of Risk 3(2), 5-40 (2000)
• Gueant, Olivier - The Financial Mathematics of Market Liquidity (2016)
• Cartea, A.; Jaimungal, S.; Penalva, J. - Algorithmic and High-Frequency Trading (2015)
• Hasbrouck, Joel - Empirical Market Microstructure (2007)

---

6. Connected Graph Bridges

• Back: 03 - The Almgren–Chriss Model · Index Hub
• Forward: 05 - Failure Modes & Practice → 06 - Advanced Extensions
• Portfolio analogue: Modern Portfolio Theory & Mean–Variance (the same mean-variance frontier, one level up: allocations instead of schedules)
• Risk: VaR & Expected Shortfall (why the L-VaR frontier exists)