Cheng Lu, Ph.D. Candidate in Financial Engineering

Bio

Cheng Lu is a Ph.D. candidate in Financial Engineering at Stevens Institute of Technology, specializing in asset allocation, risk management and machine learning applications. He has published in the International Review of Financial Analysis and presented at major conferences, including FMA and INFORMS.

Skillset

In addition to excelling in teaching, academic writing and research presentations, his technical skills include Python, R, C++, SQL, Financial Modeling, Data Science (Machine Learning, Reinforcement Learning, Text Minning) and Databases (Bloomberg, CRSP).

Dissertation Summary

Advancement of Reinforcement Learning in Asset Allocation and Pricing

This dissertation explores the application of reinforcement learning (henceforth RL) to two important do- mains in finance: assetallocation and asset pricing. The first part of the dissertation focuses on enhancing the estimation of correlation matrices for asset allocation. The second part delves into the equity risk pre- mium, a key element of asset pricing. Both areas emphasize theadaptability and benefits of reinforcement learning in addressing complex financial challenges.

The first part, titled “Improved Estimation of the Correlation Matrix using Reinforcement Learning and Text-Based Networks”, proposes a data-driven methodology that applies RL to enhance the estimation of the correlation matrix. Leveraging a model-freeRL technology, the proposed method does not require any assumption of stock return. The approach, embedded within a shrinkageframework, is tailored to optimize a risk-averse agent’s portfolio while taking into consideration estimation risk. Additionally, theRL approach incorporates the text-based networks (TBN) proposed by Hoberg and Philips, which are similarity matri- cesconstructed from financial reports (10-K reports) using natural language processing (NLP) techniques. Analyzing a high-dimensional portfolio of over 400 assets spanning more than 20 years, empirical results indicate that our RL-based approachachieves lower out-of-sample volatility and superior Sharpe ratios and reduces downside risk after accounting for transaction costs.Robustness tests across varied portfolio dimensions and sample data sizes further validate the method’s reliability and effectiveness.

The second part, “Valuing Non-Myopic Views in Equity Premium Forecasting”, deploys RL to forecast the equity risk premiumusing reverse engineering. The proposed approach integrates forward-looking market views from equilibrium mechanisms, aligning with long-term investment objectives. Adopting RL’s dy- namic capabilities, it overcomes the limitations of the myopic view standard in traditional models. The procedure also focuses on the market risk premium and incorporates the equilibrium considerations into its predictive framework. This enhancement enables the model to account for market-wide risk factors and theirdynamic interplay with macroeconomic states. It ensures that our predictions reflect a forward-looking, broader market view andfundamental economic principles, ultimately leading to more accurate and robust forecasts of the equity risk premium.

In conclusion, this dissertation demonstrates the significant potential of reinforcement learning in address- ing complex financial challenges in asset allocation and asset pricing. By enhancing correlation matrix estimation through RL and text-based networks,we provide a robust methodology that optimizes portfolio performance while accounting for estimation risk. Furthermore, byincorporating non-myopic views in eq- uity premium forecasting using RL, we offer a dynamic and forward-looking approach thatbetter captures market dynamics and macroeconomic factors. These contributions not only advance the application of ma- chine learning techniques in finance but also open avenues for future research to further integrate RL into financial modeling and decision-making processes.