2.3.4 Scheduling and Volume Profiles

Basic Prerequisites: 02 - TWAP-VWAP-POV and basic statistics.

---

1. Intuition & Practical Objective

Every VWAP algorithm stands or falls on one number it does not directly observe: the intraday volume profile - the fraction of the day’s volume that will arrive in each time bucket. Trading volume is famously U-shaped: heavy at the open (price discovery, institutional rebalancing), thin at midday (lunch lull), heavy again at the close (index-fund rebalancing). A VWAP schedule is literally the parent order poured along that curve.

The practical objective of this page: (1) learn how the profile is estimated from history as a rolling average of per-bucket volume fractions; (2) see how a forecast profile is converted into an exact child-order schedule (with anti-gaming randomization and exact-renormalization); (3) measure what happens when the forecast is wrong - the tracking error that defines the algorithm’s real-world slippage.

The one-sentence essence. “A VWAP schedule is the parent order weighted by the forecast volume profile; its job is to make your trading intensity proportional to the market’s - and its only real asset (and its only real failure) is that forecast.”

---

2. Mathematical Ground Truth & Derivations

Volume profile estimation. Over $D$ trailing days, the forecast fraction of volume in bucket $t$ is the average of per-day realized shares:

$${\varphi }_t=\frac{1}{D}\sum _{d=1}^D\frac{V_{d,t}}{V_{d,\text{total}}},\sum _{t=1}^B{\varphi }_t=\mathrm{1.0.}$$

This is the workhorse: it is just normalize-then-average. It is what a “20-day average profile” on a broker’s engine means.

VWAP child sizing. Given parent size $V$ and forecast $\varphi$, the base child in bucket $t$ is

$$q_t=V{\varphi }_t,$$

then add bounded anti-gaming randomization $q_t\to q_t(1+u_t)$, $u_t\in [-\delta ,\delta ]$, and renormalize to preserve the parent total exactly so ${\sum }_t{q}_t=V$.

Tracking error as the score of the forecast. The schedule’s realized quality is the deviation of its allocation from the realized profile ${\varphi }_t^{\text{real}}$:

$$\text{TE}(\text{realized})=\sqrt{\sum _{t=1}^B{\left(\frac{q_t}{V}-{\varphi }_t^{\text{real}}\right)}^2}.$$

When ${\varphi }^{\text{real}}={\varphi }^{\text{forecast}}$ (a clean day), TE is tiny. When a news shock reshuffles volume, TE explodes - and with it the slippage vs the realized VWAP. The TE-vs-profile gap is the whole game.

Why the U-shape exists (connects to microstructure). Volume is heaviest when information and inventory-rebalancing incentives peak - the open and close. This is the empirical shadow of the slowly-decaying temporary impact that makes “U-shaped strategies” optimal in Hasbrouck Ch 15 and AC.

---

3. Computational Implementation - forecast, schedule, and misestimation

Estimate the profile from a history of noisy U-shaped days, build a randomized-but-exact VWAP schedule, then shock the realized profile (an afternoon news event) and measure the tracking error blow-up. Stdlib only; reproduced exactly.

Read the numbers. The schedule is exact: children trace the estimated U (5,284 at the open bucket … 11,247 at the close) and sum to the parent $100,000$ after renormalization. On a clean day its tracking error is a negligible $0.0080$. But when news spikes the last three buckets to $1.6\times$, the forecast-based schedule now weights them too lightly - TE jumps to $0.0741$, nine times worse - while TWAP (volume-blind) is even worse at $0.1238$. The engine is only as good as the profile; this is the documented failure on macro-announcement days.

---

4. Failure Modes & First-Principles Breakdowns

1. The clean-day illusion. The $0.0080$ forecast TE is earned only on days whose volume follows history. On a rebalancing or event day the profile is a different shape and the schedule is mis-weighted by construction.
2. Rounding and renormalization drift. If your engine does not renormalize, rounding errors compound and the final liquidation is forced into the last bucket - the exact “late-day dump” symptom of misestimated profiles.
3. Overfitting the profile. Averaging too few days (small $D$) fits noise; too many (large $D$) smooths away genuine day-of-week effects (e.g., Monday vs Friday profiles differ). The “U” is not a constant; it is day- and name-specific.
4. Same-tool-as-opponent. A static profile is itself a signal competitors can reverse-engineer; the randomization ($\delta$) is not decoration but protection (page 05).

---

5. References

• Kissell, Glantz & Malamut - Optimal Trading Strategies (2003); Kissell - The Science of Algorithmic Trading (2014). The desk standard on volume profiles and building/slicing schedules around them.
• Hasbrouck, Joel - Empirical Market Microstructure (2007)
• Johnson, Barry - Algorithmic Trading & DMA (2010). Volume-profile estimation and VWAP/TWAP construction in practice.
• Foucault, Pagano, Roëll - Market Liquidity (2013)

---

6. Connected Graph Bridges

• Continue: 05 · Failure Modes & Practice
• Optimal scheduling: Optimal Execution & Almgren–Chriss · AC Advanced · U-shaped strategies
• Microstructure base: Market Microstructure & Order Types