Dynamic Trading with Predictable Returns & Transaction Costs

Quantitative Finance | Portfolio Optimization | Reinforcement Learning

Capstone Project — MS Financial Engineering
Stevens Institute of Technology

This project implements the dynamic trading framework proposed by Gârleanu & Pedersen (2013) to study portfolio allocation under predictable returns and transaction costs. The analytical framework is extended with a Reinforcement Learning (PPO) trading agent to explore adaptive portfolio strategies.

---

Research Objective

Evaluate whether dynamic portfolio rebalancing improves performance relative to classical static allocation strategies.

Key goals:

• Implement the Gârleanu–Pedersen dynamic trading framework
• Model signal persistence using VAR(1)
• Compare dynamic allocation against Markowitz and Equal-Weight portfolios
• Explore Reinforcement Learning as an adaptive trading strategy

---

Dataset

The analysis uses widely studied empirical asset pricing datasets.

• Fama-French factor dataset (including the Momentum factor)
• 10 Fama-French Industry Portfolios

Period: 1963 – 2024

---

Methodology

The modeling pipeline consists of:

1. Factor-based return prediction using Fama-French signals
2. Signal persistence modeling using VAR(1)
3. Covariance estimation using Ledoit-Wolf shrinkage
4. Dynamic portfolio optimization with transaction costs
5. Reinforcement Learning trading agent using PPO

---

Dynamic Trading Framework

The Gârleanu–Pedersen model determines how a portfolio should adjust over time when trading is costly.

Instead of instantaneously rebalancing to the frictionless Markowitz optimal portfolio, the strategy gradually moves toward a target allocation ("aim portfolio"), reducing excessive turnover and transaction costs.

The dynamic trading strategy follows the framework proposed by Gârleanu & Pedersen (2013).
Instead of rebalancing fully to the optimal Markowitz portfolio each period, the portfolio moves gradually toward a target allocation in order to control transaction costs.

The portfolio update rule can be expressed as:

x_{t+1} = (1 - κ) x_t + κ Aim_t

where:

• x_t = current portfolio weights
• Aim_t = target (Markowitz) portfolio based on expected returns
• κ = trading rate controlling how quickly the portfolio moves toward the target

This partial adjustment mechanism balances expected return against transaction costs and reduces excessive portfolio turnover.

The target portfolio ("aim") is the frictionless optimal allocation:

Aim_t = Σ⁻¹ μ_t

where:

• μ_t = expected return vector
• Σ = covariance matrix of asset returns

---

Key Techniques

• Dynamic Portfolio Optimization
• Transaction Cost Modeling
• Factor-Based Return Prediction
• VAR(1) Signal Dynamics
• Ledoit–Wolf Covariance Shrinkage
• Reinforcement Learning (PPO)
• Portfolio Backtesting

---

Strategy Performance (Risk Aversion γ = 3)

Strategy	Annual Return (%)	Volatility (%)	Sharpe Ratio
Dynamic Strategy	14.3	20.7	0.69
Markowitz Portfolio	10.4	13.0	0.80
RL Strategy (PPO)	14.1	21.7	0.65

---

Key Findings

• Dynamic trading produces higher cumulative returns than static Markowitz allocation.
• Transaction cost modeling leads to more realistic trading behavior.
• Reinforcement learning provides comparable performance in adaptive portfolio allocation.

---

Repository Structure

dynamic-trading-portfolio-optimization

data/
    dataset_fama_french_industry_returns.csv

report/
    dynamic_trading_capstone_report.pdf

presentation/
    dynamic_trading_capstone_presentation.pptx

---

Tech Stack

• Python
• NumPy
• Pandas
• CVXPY
• Gymnasium
• Stable-Baselines3
• Matplotlib

---

Reference

Gârleanu, N., & Pedersen, L. H. (2013).
Dynamic Trading with Predictable Returns and Transaction Costs
Journal of Finance.