Overview

Led a team of 6 students to develop a comprehensive market-making framework based on the seminal Avellaneda–Stoikov model, featuring volatility-adaptive spreads and sophisticated inventory risk controls.

Period: Jan 2024 – Jun 2024
Organization: EPF Graduate School of Engineering (Capstone Project)
Role: Team Leader & Quantitative Developer
Team Size: 6 members

Project Background

The Avellaneda–Stoikov framework is a foundational model in market microstructure that addresses the fundamental problem faced by market makers: how to set bid and ask quotes to maximize profit while managing inventory risk.

The Problem

Market makers must balance:

Model Foundation

Avellaneda–Stoikov Model (2008)

The optimal bid and ask quotes are given by:

\[p^{\text{bid}} = S - \frac{1}{\gamma}\ln(1 + \gamma/\kappa) + \frac{q \sigma^2 (T-t)}{\gamma}\] \[p^{\text{ask}} = S + \frac{1}{\gamma}\ln(1 + \gamma/\kappa) + \frac{q \sigma^2 (T-t)}{\gamma}\]

Where:

Key Features

Volatility-Adaptive Spreads

Inventory Risk Control

Adverse Selection Mitigation

Risk Management Features

Technical Implementation

Architecture

# Core Components:
- MarketDataHandler: Real-time order book processing
- VolatilityEstimator: Rolling volatility calculation
- InventoryManager: Position tracking and limits
- QuoteEngine: Bid/ask quote generation
- ExecutionEngine: Order placement and management
- RiskManager: Real-time risk monitoring
- Backtester: Event-time simulation framework

Data Processing

Testing & Validation

Results

Performance Metrics

Metric Improvement
Adverse Selection -9% reduction
PnL Variance -7% stabilization
Sharpe Ratio +15% increase
Fill Rate +12% improvement

Key Outcomes

Team Leadership

As team leader, I:

Challenges Overcome

  1. Event-Time Backtesting: Implemented accurate event-driven simulation (not time-sampled)
  2. Data Quality: Handled missing data, outliers, and exchange-specific quirks
  3. Parameter Optimization: Tuned risk aversion, volatility window, and inventory limits
  4. Team Coordination: Managed parallel development across multiple modules

Technical Stack

Key Learnings

  1. Model vs. Reality: Theoretical models require significant adaptation for real markets
  2. Data Quality Matters: Robust handling of market data anomalies is essential
  3. Parameter Sensitivity: Careful parameter tuning and validation is critical
  4. Team Dynamics: Effective communication and clear architecture enable parallel development

Extensions & Improvements

Beyond the base Avellaneda–Stoikov model, we implemented:

Academic Contribution

This project synthesized concepts from:

Future Work

Potential extensions include:

References