2.2.6 Advanced Extensions

Basic Prerequisites: 05 - Failure Modes & Practice.

---

1. Intuition & Practical Objective

Each failure of Almgren–Chriss (page 05) is the seed of an extension. This page is the launchpad: it shows the four extensions that turn the static linear model into something a modern desk can run - nonlinear impact (Almgren 2003), resilient/transient impact (Obizhaeva–Wang 2013; Gatheral, Schied & Slynko 2012), dark pools / alternative venues, and adaptive stochastic control (Cartea–Jaimungal–Penalva 2015) - and it hands off to the dedicated topic-folders for the microstructure and market-making views.

Why in this order? Nonlinear impact fixes the cost function; resilience fixes the market’s memory (the book refills, so the optimal schedule gains discrete trades at both ends); dark pools add a venue choice with its own fill risk; adaptive control fixes the dynamics by re-solving as information arrives. Together they are the minimal set that makes the AC idea survive contact with a real order.

The one-sentence essence. “Relax one assumption at a time: concave impact bends the trajectory, a resilient book makes it block-continuous-block, a dark venue trades price improvement for fill risk, and stochastic control turns the static curve into a feedback law.”

---

2. Mathematical Ground Truth & Derivations

2.1 Nonlinear (power-law) impact - Almgren (2003). With $h(v)=\eta v^{\alpha }$, $\alpha \in (0,1]$, the cost functional becomes ${\int }_0^T[\eta (-\dot{x}{)}^{1+\alpha }+\lambda {\sigma }^2{x}^2]dt$, and the Euler-Lagrange equation is nonlinear. Almgren’s key results:

• a characteristic time $T_{\star }=(\alpha \eta X^{\alpha -1}/(\lambda {\sigma }^2){)}^{1/(\alpha +1)}$, which depends on portfolio size;
• in the regime $\alpha \gg 1$ the trajectory reaches $x=0$ with $\dot{x}=0$ at a finite $T_{\max }=\frac{\alpha +1}{\alpha -1}{T}_{\star }$;
• closed forms per regime: for $0<\alpha <1$ a stretched-exponential; for $\alpha =1$ the exponential (AC); for $\alpha >1$ a polynomial tail.

2.2 Resilient / transient impact - Obizhaeva–Wang (2013). Model the book’s state: the transient impact $a_t$ created by trading decays with a resilience rate $\rho$, and the execution price is depressed by the accumulated, still-unreplenished impact:

$${\text{cost}}_k=n_k(\frac{s}{2}+a_k+\frac{n_k}{2q}),a_{k+1}=(1-\rho )(a_k+\frac{n_k}{q}),q=\text{book depth}.$$

The optimal schedule is not smooth: it is a discrete block at $t=0$, a continuous/small-trade middle, and a discrete block at $T$ - “block-continuous-block”. In the risk-neutral limit the two blocks are exactly equal ($n_1=n_N$), the Obizhaeva–Wang symmetry. This is the discrete time-value of finite resilience: trading before the book refreshes is expensive, so you wait at the ends and work only when the book has refilled.

2.3 Transient impact in continuous time - Gatheral (2010, 2013); Gatheral-Schied-Slynko (2012). Generalize the temporary impact to a decay kernel $G(\cdot )$: the price impact of a trade at $u<t$ contributes $G(t-u)d{X}_u$, and the mid-price is

$$S_t=S_0+\sigma W_t+{\int }_0^{t}G(t-u)d{X}_u.$$

The no-dynamic-arbitrage condition (Huberman–Stanzl 2004; Gatheral 2010) requires $G$ to be nonincreasing and convex - this rules out many “natural” models and is the consistency constraint every impact model must satisfy. The optimal execution problem becomes a Fredholm equation; for power-law kernels $G(t)\propto t^{-\gamma }$ the optimal strategy has a characteristic oscillating/decaying profile (Gatheral’s Figure 22.2).

2.4 Dark pools and venue choice (Cartea–Jaimungal–Penalva 2015, Ch 8-9). A dark venue executes at the midpoint (saving the half-spread) but fills only with probability $p$; unfilled shares must be worked later in the lit book with an adverse-selection penalty. The per-share expected cost of routing to dark is

$$c_{\text{dark}}(p)=(1-p)(c_{\text{lit}}+\delta ),$$

so dark beats lit iff $p>p^{\star }=\delta /(c_{\text{lit}}+\delta )$ - a threshold rule on the fill probability. In practice $p$ is estimated per venue per order size, and the strategy becomes a venue-allocation control.

2.5 Adaptive / stochastic control (Cartea–Jaimungal–Penalva 2015). Replace the static trajectory with a feedback law ${\nu }^{\star }(t,x,\text{state})$ from an HJB equation that includes inventory, transient impact, and possibly alpha/drift. Then parameters are re-estimated as fills arrive and the schedule updates - this is the production form of AC.

---

3. Computational Implementation - two extensions, verified

A. Resilient book ⇒ block-continuous-block (Obizhaeva–Wang). Build the exact quadratic objective for the decaying-impact model and solve the equality-constrained programme. The risk-neutral limit must reproduce the OW symmetry $n_1=n_N$. numpy + stdlib.

Read the first and last columns: for low resilience ($\rho =0.05$ - the book barely refills) the schedule is nearly symmetric with a large terminal block ($n_N=61,309$, and the last quarter carries 25% of the volume); for high resilience ($\rho =0.90$ - the book refills fast) the schedule is front-loaded and the terminal block vanishes ($n_N=518$). The Obizhaeva–Wang limit is visible in the ratio $n_1/n_N$: it falls toward $1$ as resilience vanishes - $1.41$ at $\rho =0.05$, $5.6$ at $\rho =0.40$, $36.6$ at $\rho =0.90$ (exactly $1$ only in the joint limit $\lambda \to 0,\rho \to 0$).

B. Dark-pool allocation. The break-even fill probability on the example numbers.

A venue with fill probability above 33.2% is worth routing to on these numbers; below it, the unfilled tail (and the adverse selection of trading it late) dominates the saved half-spread.

---

4. Failure Modes & First-Principles Breakdowns

1. Power-law exponents are estimated, not known. Almgren’s $T_{\star }\propto X^{(\alpha -1)/(\alpha +1)}$ is sensitive to $\alpha$; a mis-fit exponent mis-paces large orders systematically (the 2005 Almgren-Thum-Hauptmann-Li estimation paper exists precisely because this matters).
2. A decay kernel must satisfy no-dynamic-arbitrage. Gatheral (2010) / Huberman–Stanzl (2004): $G$ must be nonincreasing and convex. Ad-hoc exponential kernels with the wrong parameters admit price manipulation - a model that creates free money is wrong regardless of its fit.
3. Resilience is unobservable and regime-dependent. $\rho$ changes with volatility and with the presence of other large traders; assuming it constant under-states the terminal-block risk that OW highlights.
4. Dark fill probability is adversely selected. $p$ estimated ex-ante overstates the realized fill rate conditional on your trade being informed/urgent (you fill when the market is about to move against you). The threshold rule $p^{\star }$ uses an average $p$, which is the wrong conditional quantity.
5. Adaptive control inherits estimation lag. Re-solving as fills arrive helps only if the re-estimation is faster than the regime change; otherwise you chase a stale parameter (and re-solving too often adds turnover, incurring extra impact).
6. The extensions do not compose for free. Nonlinear impact + transient impact + dark pools + control is a high-dimensional, poorly-identified system; in practice desks pick two or three effects, calibrate conservatively, and leave the rest to robust/randomized execution.

---

5. References

• Almgren, Robert - “Optimal execution with nonlinear impact functions and trading-enhanced risk,” Applied Mathematical Finance 10(1), 1-18 (2003). Power-law impact, $T_{\star }\propto X^{(\alpha -1)/(\alpha +1)}$, trading-enhanced risk.
• Obizhaeva, Anna; Wang, Jiang - “Optimal trading strategy and supply/demand dynamics,” Journal of Financial Markets 16(1), 1-32 (2013). Resilient book, block-continuous-block, $n_1=n_N$ symmetry.
• Gatheral, Jim - “No-dynamic-arbitrage and market impact,” Quantitative Finance 10(7), 749-759 (2010). The admissibility constraint on impact/decay models.
• Gatheral, Jim; Schied, Alexander; Slynko, Alla - “Transient linear price impact and Fredholm integral equations,” Mathematical Finance 22(3), 445-474 (2012). Optimal execution for general decay kernels (power-law kernels, oscillating strategies).
• Gatheral, Jim - “Dynamical models of market impact and algorithms for order execution,” in Handbook on Systemic Risk (2013)
• Cont, Rama; Kukanov, Arseniy; Stoikov, Sasha - “The price impact of order book events,” Journal of Financial Econometrics 12(1), 47-88 (2014). *OFI and the concave/square-root impact relation
• Cartea, A.; Jaimungal, S.; Penalva, J. - Algorithmic and High-Frequency Trading (2015)
• Gueant, Olivier - The Financial Mathematics of Market Liquidity (2016)
• Huberman, Gur; Stanzl, Werner - “Price manipulation and quasi-arbitrage,” Econometrica 72(4), 1247-1275 (2004). The no-manipulation condition behind Gatheral’s constraint.

---

6. Connected Graph Bridges

• Back: 05 - Failure Modes & Practice · Index Hub
• Impact-law depth: Market Impact & Depth (Kyle $\lambda$, temporary vs permanent, square-root law, transient impact)
• Market-making twin: Avellaneda–Stoikov & Optimal Quoting (inventory control; the same HJB family)
• Execution heuristics: Execution Algorithms: VWAP, TWAP, POV · Market Microstructure & Order Types (dark/hidden orders)
• Risk & portfolio: Liquidity Risk & Funding · Constraints & Transaction Costs