Executive Summary
An algorithmic trading firm was losing 12bps to slippage, eroding 30% of strategy alpha. Building a slippage prediction model with dynamic order routing reduced slippage to 5.4bps, adding $12M annual PnL.
Key Outcomes
- ▹ 12bps → 5.4bps slippage (55% reduction)
- ▹ $12M additional annual PnL
- ▹ 66% of trades executed within 1 tick of arrival price
Client Situation
The firm's mid-frequency strategies traded 500k orders daily. Slippage was the largest cost after commissions, especially during volatile periods.
Key Challenges
- ⚠ Slippage spiking to 30bps during news events
- ⚠ No differentiation between liquidity-taking vs providing venues
- ⚠ Static order sizing ignoring market impact
Existing Architecture
Simple market/limit order logic with fixed participation rates. No slippage forecasting or adaptive routing.
- Market orders taking liquidity paying spread + slippage
- No prediction of adverse selection
- Inability to detect toxic order flow
Solution Design
Slippage prediction model using market microstructure features, with dynamic order type and venue selection.
Key Decisions
- ✓ Predict slippage using order book imbalance, spread, volatility
- ✓ Dynamic limit order pricing with queue position estimate
- ✓ Toxicity detection to avoid adverse selection
Implementation
Shadow mode for 1 month comparing model predictions against actual slippage before live deployment.
Phase 1: Phase 1: Feature Engineering
Created 50 microstructure features from order book and trade data.
Phase 2: Phase 2: Model Training
XGBoost predicting slippage with 0.82 R² on validation data.
Phase 3: Phase 3: Live Integration
Model-driven order routing for 25% of volume, expanded to 100% after 2 weeks.
Technical Challenges
- Slippage definition consistency
Impact: Different benchmarks giving conflicting signals
Resolution: Standardized on mid-price arrival slippage for all strategies
- Model inference latency
Impact: 10ms prediction time unacceptable for fast strategies
Resolution: Precomputed features with 1ms cached inference
Results
- Average slippage
- Before12bpsAfter5.4bpsImprovement55% reduction
- Orders filled within 1 tick
- Before48%After66%Improvement38% increase
- Adverse selection events
- Before15% of ordersAfter6% of ordersImprovement60% reduction
Lessons Learned
- 📘 Slippage prediction model saved 0.5bps on liquidity-taking trades
- 📘 Toxicity detection avoided 60% of adverse selection
- 📘 Dynamic order sizing based on urgency improved execution
What We Would Do Differently
- 💡 Include dark pool routing earlier
- 💡 Implement reinforcement learning for order aggressiveness
Role Relevance
Quant traders with microstructure expertise designed the feature set and validated the model's economic impact, not just statistical fit.
Critical Skills Demonstrated
Related Roles
Frequently Asked Questions
- Which features most predicted slippage?
- Order book imbalance, quote size ratio, and recent trade direction.
- How did you validate the model?
- Out-of-sample backtest on 6 months of data with 0.78 R².