Optimal Execution Algorithms: TWAP, VWAP & Market Impact › Limitations and Practical Considerations

The theoretical frameworks presented here provide valuable intuition, but practitioners must address several challenges when implementing execution algorithms in production.

The Almgren-Chriss framework assumes linear market impact but empirical research consistently finds that impact is concave. Your first thousand shares have greater per-share price impact than your next thousand. Square-root impact models often fit data better:

$$\Delta P \propto \sigma \sqrt{\frac{Q}{V}}$$

where:

• $\Delta P$: price impact
• $\sigma$: volatility
• $Q$: order size
• $V$: market volume

However, these models do not yield closed-form optimal trajectories.

Similarly, the assumption of constant volatility fails during news events, earnings announcements, and market stress. Adaptive algorithms attempt to address this but accurate real-time volatility estimation remains challenging. The model also ignores the bid-ask spread and other fixed costs, which can dominate impact costs for smaller orders.

Optimal execution requires knowing impact coefficients $\gamma$ and $\eta$, volatility $\sigma$, and risk aversion $\lambda$. Permanent impact is notoriously difficult to measure. It requires separating the price movement caused by your trading from the price movement that would have occurred anyway. Temporary impact estimates are sensitive to the measurement window. These parameters likely vary by security, time of day, market conditions and order size. A model calibrated on historical data may perform poorly when your trading conditions change.

In practice, execution algorithms operate in an adversarial environment. High-frequency traders can detect algorithmic patterns and trade against them in a practice known as "algo sniffing." TWAP's predictability makes it particularly vulnerable. Even sophisticated algorithms can be reverse-engineered if their behavior is sufficiently regular. This creates pressure toward randomization and adaptation that goes beyond what theory suggests.

The difference between theoretical and realized performance often comes down to implementation details that academic models ignore. These include:

• Order routing decisions
• Exchange selection
• Handling of partial fills
• Behavior at the close
• Response to exchange outages

As we will discuss in the next chapter on trading systems infrastructure, these engineering challenges can easily overwhelm your theoretical gains from optimal scheduling.