2.3.5 Failure Modes and Practice

Basic Prerequisites: 02 - TWAP-VWAP-POV and 04 - Scheduling & Volume Profiles.

---

1. Intuition & Practical Objective

This page asks the question a textbook never does: how does an execution algorithm actually get exploited in the wild? Every schedule is a pattern, and the market is populated by fast players whose entire business is detecting patterns. The three structural enemies are:

1. Gaming - a predictable schedule (fixed TWAP heartbeat, or a VWAP engine that leaks its profile) becomes a front-running key. Snipers buy just ahead of each child and sell it to you at a lifted price.
2. Volume-profile misestimation - the U-shaped forecast (page 04) fails exactly when it matters: on a macro-announcement day, when volume reshuffles and the engine under-trades the open, forcing a late-day dump.
3. Adverse selection - schedule-driven participation that is slow to react trades on the wrong side of informed moves: it keeps buying as price ladders up (or sells into a falling knife) because the schedule says “keep participating,” not “the market moved.”

The practical objective: make each failure quantitative, then state the desk-level defenses. Every number below is re-executed and reproduced.

The one-sentence essence. “An execution algorithm that is predictable or volume-blind is a standing offer to faster traders - gaming, misestimation and adverse selection are just the three ways that offer gets taken.”

---

2. Mathematical Ground Truth & Derivations

Gaming: the fixed-schedule concession. Let a TWAP engine ship a fixed child $X/B$ at deterministic 60-second intervals. A sniper who has reverse-engineered the heartbeat buys $h$ shares just before each child, pushing the quote up by $\Delta$; the engine’s child then fills $\Delta$ worse. Per child the engine pays an extra $\frac{X}{B}\Delta$, so over the parent:

$$\text{extra cost}=\sum _{t=1}^B\frac{X}{B}\Delta =X\Delta .$$

The fix is randomization (Foucault: VWAP “can be gamed by slow trickling”) - child size, timing, and venue all get bounded noise, so the schedule is not a reproducible linear prediction. This is why the anti-gaming ${\epsilon }_t$ in page 01/02 is not decoration.

Misestimation: tracking-error magnification. From page 04, a VWAP schedule with weights $w=q/V$ faces tracking error

$$\text{TE}=\sqrt{{\sum }_t{\left(w_t-{\varphi }_t^{\text{real}}\right)}^2},$$

which jumps from $O(0.01)$ on a clean day to $O(0.07)$ on a news day. The consequence is slippage: because the engine under-weights the early heavy buckets and over-weights the close, its average price disconnects from the realized VWAP and it is forced to liquidate remaining size into the thinnest, least-predictable tape of the day.

Adverse selection: the price ladder. If the true value moves in a trend while an informed buyer accumulates, the average price paid by a slow, schedule-driven buyer rises monotonically. The cost of being late by one ladder step $\delta$ on the unfilled remainder is $(1-\kappa )q\delta$ in IS (from page 03’s opportunity-cost term) - a purely informational penalty that no amount of impact-taming avoids.

---

3. Computational Implementation - quantify all three failures

Simulate (1) the forced concession of a deterministic TWAP, (2) the tracking-error blow-up on a CPI day, and (3) the extra cost of trading after (rather than before) an informed move. Stdlib only; reproduced exactly.

Read the numbers. (1) A sniper forcing just $0.02 of extra concession per share (2 cents!) on a deterministic TWAP costs the parent a clean $2,000 (2 bps) - pure transfer to the adversary, avoided by randomization. (2) On an open-announcement day the schedule puts only $5.26\%$ of size at the open when $12.57\%$ of the volume landed there - a tracking-error blow-up to $0.0762$ that forces a close dump. (3) Being slow on an informed buy costs $148,000$, or 370 bps - an order of magnitude worse than any impact the schedule was built to avoid. The schedule won the impact battle and lost the information war.

---

4. Failure Modes & First-Principles Breakdowns

1. Gaming is a zero-sum transfer (Harris 2003; Hasbrouck Ch 14). When one trader’s VWAP slippage is another’s edge against the same benchmark, “low measured cost” can be pure selection. A good-looking execution report on a gameable benchmark proves nothing.
2. VWAP is order-size-dependent. If you account for a large share of the day’s volume, your average price approaches the VWAP regardless of how well you handled it (Hasbrouck Ch 14) - the benchmark auto-flatters big orders.
3. Fast laggard = adverse selection (Glosten–Milgrom connection). The microstructure root of failure (3) is precisely the Glosten–Milgrom information asymmetry: an informed flow arrives, and anyone who posts or participates passively gets picked off.
4. The practice stack is defense-in-depth: randomize schedules to defeat gaming; use real-time volume (POV, page 06) to defeat profile misestimation; and let the algorithm react to price/variance (adaptive IS, page 06) to cut adverse selection. No single static TWAP/VWAP defeats all three.

---

5. References

• Hasbrouck, Joel - Empirical Market Microstructure (2007)
• Foucault, Pagano, Roëll - Market Liquidity (2013)
• Harris, Larry - Trading and Exchanges (2003). The classic statement of benchmark gamability and execution ambiguity.
• Kissell - The Science of Algorithmic Trading (2014). Desk practice: randomization, real-time participation, TCA of execution quality.

---

6. Connected Graph Bridges

• Continue: 06 · Advanced Extensions
• Root cause: Adverse Selection & Glosten–Milgrom · Market Impact & Depth
• Optimal defense: Optimal Execution & Almgren–Chriss · Queue Position & Fill Probability