2.3.2 TWAP, VWAP and POV

Basic Prerequisites: 01 - From Zero and basic weighted averages.

---

1. Intuition & Practical Objective

Once you accept that the parent order must be sliced (page 01), the next question is along what? Three answers dominate the desk because they are simple to state, cheap to run, and easy to explain to a client:

• TWAP (Time-Weighted Average Price) - send the same child size every bucket, on the clock. No volume data, no forecast, no model. The natural default when you know nothing.
• VWAP (Volume-Weighted Average Price) - send a child proportional to the expected volume of each bucket. This is strictly smarter than TWAP when you have a volume forecast, because it aims to match the market’s own pace and therefore trade where the day’s liquidity lives.
• POV (Percentage-of-Volume / participation rate) - send a child equal to a fixed fraction $\rho$ of the live volume arriving each moment. POV never over-participates; if the market stops trading, POV stops trading.

The practical objective of this page: build all three schedules from one volume profile, then simulate them against a realized price path and measure which one actually tracks the VWAP benchmark - the acceptance test every execution report runs.

The one-sentence essence. “TWAP splits by the clock, VWAP splits by the forecast distribution of volume, POV splits by the realized distribution - the smarter the volume awareness, the closer the average price to the day’s VWAP.”

---

2. Mathematical Ground Truth & Derivations

The market VWAP (Foucault eq 2.7). Over $K$ trades the day’s benchmark is the volume-weighted mean,

$$\text{VWAP}=\sum _{k=1}^K{w}_k{p}_k,w_k=\frac{v_k}{\sum v_k}.$$

Over $B$ buckets with realized volume shares ${\varphi }_t$, this is $\text{VWAP}={\sum }_t{\varphi }_t{p}_t$. Any schedule whose child weights $w_t^{\text{exec}}$ equal the realized shares ${\varphi }_t$ executes exactly at the VWAP - that is the definition of a “perfect VWAP engine.”

TWAP schedule. Equal child sizes,

$$q_t=\frac{V}{B}+{\epsilon }_t,$$

where ${\epsilon }_t$ is bounded anti-gaming noise. Its weights are flat, $w_t^{\text{exec}}=1/B$, so it executes at the time-weighted average price, ${\bar{p}}_{\text{TWAP}}=\frac{1}{B}{\sum }_t{p}_t$, which deviates from VWAP by the covariance between price and volume,

$$\text{VWAP}-{\bar{p}}_{\text{TWAP}}=\underset{\text{covariance of price with volume share}}{\underbrace{{\text{Cov}}_{\varphi }\left(p,\varphi \right)}},$$

VWAP schedule. Proportional to the forecast profile ${\varphi }_t$:

$$q_t=V{\varphi }_t+{\epsilon }_t.$$

If the forecast equals realized volume, the engine’s average price is the VWAP to within $\epsilon$. Its tracking error - the RMS deviation of its allocation from the realized profile - is the quantity every execution report prints:

$$\text{TE}=\sqrt{{\sum }_{t=1}^B{\left(w_t^{\text{exec}}-{\varphi }_t\right)}^2}.$$

POV schedule. Track live volume, don’t forecast it:

$$q_t=\rho v_t^{\text{live}},v_t^{\text{live}}=\text{arriving market volume}.$$

POV has by construction $w_t^{\text{exec}}={\varphi }_t^{\text{realized}}$ (you participate in exactly a constant slice of every unit traded), so its tracking error w.r.t. realized volume is $0$ by definition - its downside is that it is fully exposed to whatever volume arrives (see page 05).

---

3. Computational Implementation - build and simulate all three

Build the VWAP and TWAP schedules from the canonical 13-bucket U-shaped profile, then simulate a rising price path (volume = forecast, so the VWAP engine is exact) and measure which engine tracks the VWAP benchmark. Stdlib only; reproduced exactly.

Read the numbers. With volume = forecast, the VWAP engine hits the benchmark to four decimals ($+0.00$ bps) and its tracking error is $0.0000$ by construction. TWAP, spreading equally across a U-shaped day, over-trades the thin midday and under-trades the heavy open/close - a tracking error of $0.0607$ and a $\approx 6$ bps gap vs the VWAP benchmark. The VWAP schedule’s children (5,263 … 11,579) trace the smile; TWAP’s are flat. The VWAP engine is not smarter - it is matched to the day’s volume; TWAP is deliberately volume-blind.

---

4. Failure Modes & First-Principles Breakdowns

1. “Beat VWAP” is meaningless as a discovery - it is a benchmarking artifact. Because VWAP depends on your own realized volume, an engine can “beat” the day’s VWAP merely by over-trading the heavy buckets (or a broker by dominating volume). Hasbrouck Ch 14 and Harris (2003) both flag this: VWAP slippage is gameable, so it is a reporting benchmark, not a decision benchmark.
2. TWAP’s volume blindness is a feature and a bug. No forecast needed (good on an illiquid, unpredictable name) but zero volume awareness (bad on any day with a U-shaped profile, where it systematically mis-paces).
3. VWAP’s accuracy is hostage to the forecast. The $0.0000$ above was a free gift: realized volume exactly equaled forecast. The moment a CPI print moves volume, the VWAP engine’s tracking error explodes (page 05).
4. POV’s self-pacing is exposure. POV cannot be front-run by sharing size, but it must participate in whatever volume arrives - on an informed down-trend it is forced to sell into the knife (adverse selection).

---

5. References

• Foucault, Pagano, Roëll - Market Liquidity (2013)
• Hasbrouck, Joel - Empirical Market Microstructure (2007)
• Kissell, Glantz & Malamut - Optimal Trading Strategies (2003); Johnson - Algorithmic Trading & DMA (2010). The practitioner treatment of building TWAP/VWAP/POV and their benchmarks.

---

6. Connected Graph Bridges

• Continue: 03 · Implementation Shortfall · 04 · Scheduling & Volume Profiles
• Benchmark-deep-dive: Optimal Execution & Almgren–Chriss
• Microstructure base: Market Microstructure & Order Types