6.2.6 Advanced Extensions

Basic Prerequisites: 03 · The AS Model and 05 · Failure Modes.

---

1. Intuition & Practical Objective

Every defect of the baseline AS model motivates an extension. This page is the launchpad: it presents the canonical upgrades that stay close to the AS family - inventory constraints (Guéant–Lehalle–Fernandez-Tapia), adverse selection and alpha (Cartea–Jaimungal–Penalva), alternative risk measures (Cartea–Jaimungal), and multi-asset / portfolio quoting - and links each to its dedicated in-pillar page.

Why these first? Inventory limits are the minimal fix (the original AS quotes are only valid for moderate $|q|$). Adverse selection is the most important fix (it is what actually kills desks). Multi-asset is the scaling fix (real makers quote thousands of correlated instruments and cannot treat each in isolation). Everything farther out - hidden Markov mid-prices, deep learning quoters - links from here.

---

2. Mathematical Ground Truth & Derivations

2.1 Guéant–Lehalle–Fernandez-Tapia: rigorous HJB with inventory limits

Guéant, Lehalle & Fernandez-Tapia (2013) re-solve the market-making problem with the constraints the original paper left open. Under intensities ${\lambda }^a(\delta )=A{e}^{-k\delta }$ and ${\lambda }^b(\delta )=A{e}^{-k\delta }$, a change of variables turns the HJB system into a system of linear ODEs for the functions $v_q(t)$:

$${\dot{v}}_q(t)=\alpha q^2{v}_q(t)-\eta (v_{q-1}(t)+v_{q+1}(t)),\alpha =\frac{k}{2}\gamma {\sigma }^2,\eta =A{\left(1+\frac{\gamma }{k}\right)}^{-(1+k/\gamma )},$$

(boundary states drop the out-of-range neighbour: ${\dot{v}}_Q=\alpha Q^2{v}_Q-\eta v_{Q-1}$, ${\dot{v}}_{-Q}=\alpha Q^2{v}_{-Q}-\eta v_{-Q+1}$.)

on the $2Q+1$ inventory states $\{-Q,\dots ,Q\}$ (with $v_q(T)=1$), from which $u(t,x,q,s)=-\exp (-\gamma (x+qs))v_q(t{)}^{-\gamma /k}$ recovers the value function. Two consequences matter in practice:

• Inventory limits make the problem well-posed. The system is solved under $|q|\le Q$; quotes on the binding side are effectively pulled. This supplies a verification theorem absent from the original AS paper (the admissibility of the raw AS quotes is, in the authors’ words, “an open problem”).
• Closed-form asymptotics. In the large-$T$ / moderate-$q$ regime the optimal distances converge to a stationary form ${\delta }_{\infty }^{b\ast }(q)\simeq \frac{1}{\gamma }\ln (1+\frac{\gamma }{k})+\frac{1}{2k}\sqrt{\frac{\alpha }{\eta }}(2q+1)$, ${\delta }_{\infty }^{a\ast }(q)\simeq \frac{1}{\gamma }\ln (1+\frac{\gamma }{k})-\frac{1}{2k}\sqrt{\frac{\alpha }{\eta }}(2q-1)$, whose spread reproduces the AS $\frac{2}{\gamma }\ln (1+\gamma /k)$ plus an inventory-dependent correction. This is the form most production engines implement.

2.2 Cartea–Jaimungal–Penalva: adverse selection and alpha

The modern textbook adds three ingredients AS omits: (i) a drift / alpha ${\mu }_t$ in the mid-price, (ii) adverse selection, via a fraction of “informed” market orders whose arrival correlates with future price moves, and (iii) market impact of the dealer’s own quotes. The reservation price generalizes to include a forecast of the price over the holding horizon, and the optimal spread gains a term proportional to the expected adverse move per fill. This is the structural fix for the 05 failure; the practical proxy is toxicity-gated quoting (VPIN).

2.3 Cartea–Jaimungal: alternative risk measures

Exponential utility fixes the skew to a single parameter $\gamma$. Cartea & Jaimungal (2015) replace CARA with general dynamic risk measures (mean–variance, expected shortfall, drawdown-based) and derive the corresponding reservation prices and spreads - showing how the shape of the skew depends on the chosen risk criterion, not just its scale. If a desk’s mandate is drawdown-limited rather than variance-averse, this is the paper to start from.

2.4 Multi-asset / portfolio quoting

In reality a maker quotes many correlated instruments sharing capital and margin. The natural extension is a vector inventory $\mathbf{q}$ with a covariance matrix $\Sigma$; the reservation price becomes $r_i=s_i-\gamma {\mathbf{e}}_i^{\top }\Sigma \mathbf{q}(T-t)$ (inventory of other assets cross-hedges or infects asset $i$), and the spread gains a term from the cross-impact of fills across instruments. The skew is no longer per-asset but portfolio-wide: a long position in a positively-correlated neighbour is treated as (partial) excess inventory in the target asset.

2.5 Trend / drift extension

Guéant et al. also give the solution for a mid-price with a drift $d{S}_t=\mu dt+\sigma d{W}_t$ (their Prop. 4): the same linear-ODE machinery applies, and the reservation price acquires a $\mu (T-t)$-type term. Ignoring a real drift, as the baseline does, is the $(\text{A2})$ failure of 05.

---

3. Computational Implementation - the effect of a hard inventory cap

The cheapest, most robust extension: enforce $|q|\le Q$ by pulling the inventory-growing quote at the cap. We reuse the 04 simulator and compare no-cap, $Q=5$, $Q=3$.

The cap does exactly what theory says: it hard-bounds inventory (max$|q|$ becomes $5$, then $3$) and shrinks the terminal-inventory dispersion ($3.17\to 2.50\to 1.77$) at a tiny cost in mean P&L ($56.99\to 55.57$). This is the production-grade behaviour Guéant et al. formalize: the skew encourages flattening; the cap guarantees it.

---

4. Failure Modes & First-Principles Breakdowns

1. A cap trades edge for safety. Bounding $|q|$ removes the tail of inventory risk but also removes the fills that would have earned the widest spreads; the mean P&L falls. Choose $Q$ from a risk budget, not by taste.
2. Adverse selection survives every inventory fix. Capping inventory does nothing about informed flow - the fix must add an information term (Cartea–Jaimungal–Penalva) or an external toxicity gate (VPIN). Do not mistake inventory control for safety against being picked off.
3. Closed-form asymptotics have validity ranges. The Guéant approximation is excellent for small $|q|$ and degrades for large $|q|$ (it approximates ratios $f_{q\pm 1}^0/f_q^0$); use the exact linear-ODE solution when the inventory limit is tight or $q$ large.
4. Multi-asset coupling is ignored at your peril. Treating $N$ correlated instruments independently double-counts risk and quotes $N$ times the intended inventory; the correlation matrix must enter the reservation price or the desk is systematically over- or under-hedged.
5. Drift is easy to add and easy to overfit. A real $\mu$ belongs in the reservation price (Guéant Prop. 4), but an estimated $\mu$ is noisy - a spurious drift shifts quotes and can turn a market-neutral maker into a directional bet.

---

5. References

• Guéant, Lehalle & Fernandez-Tapia (2013), Dealing with the inventory risk, Math. & Financial Econ. 7(4), 477–507
• Cartea, Jaimungal & Penalva (2015), Algorithmic and High-Frequency Trading, Cambridge UP
• Cartea & Jaimungal (2015), Risk metrics and fine tuning of high-frequency trading strategies, Mathematical Finance 25(3), 576–611
• Avellaneda & Stoikov (2008), Quantitative Finance 8(3)

---

6. Connected Graph Bridges

• Back: 05 · Failure Modes · Index Hub
• In-pillar forward: Inventory Management & Quote Skewing · Adverse Selection & Glosten–Milgrom · Toxic Order Flow & VPIN
• Cross-pillar: Optimal Execution & Almgren–Chriss · POV