Execution quality compounds. Small deviations between model and realized prices add up and can erase signal edge. We use adaptive, online models for slippage and market impact so our routing and scheduling improve as conditions change intraday and across regimes.
Feature set
- Liquidity and depth: spread, depth imbalance, queue position.
- Volatility and momentum: short‑horizon volatility, micro‑trend.
- Order context: parent order size, urgency, participation cap.
- Venue microstructure: fill probabilities, maker/taker fees, toxicity.
Online learning
We fit light‑weight models (regularized GLMs and incremental trees) with partial‑fit updates. Exogenous breaks trigger learning‑rate bumps and temporary constraints on aggression. The objective is expected implementation shortfall under risk and participation constraints.
Backtest pitfalls
- Simulate queue dynamics and avoid last‑look bias.
- Include venue fees/rebates and adverse selection.
- Respect market impact cross‑section when trading baskets.
Implementation notes
- Throttle updates and guardrails to avoid over‑reaction.
- Embed forecasts into a scheduler (TWAP/VWAP/POV hybrids).
- Continuously reconcile forecast vs. realized costs.