Reproduce “Enhancing Portfolio Optimization with Deep Learning Insights” on OpenResearch

About this paper

Brandon Luo, Jim Skufca

Researchers from Georgia Institute of Technology explored deep learning applications in portfolio optimization, demonstrating that while an existing LSTM-based method replicates initial promising results, it exhibits diminished performance during market volatility or with new asset classes. They developed a Transformer model with pre-training and expanded features, which achieved a Sharpe ratio of 2.7, outperforming the original LSTM (1.828) and traditional benchmarks.

Key insights

• The original LSTM-based deep learning approach, while initially performing well and replicating the Oxford study, showed significant limitations in robustness when exposed to extended periods of high market volatility or different asset compositions.
• Simply expanding the feature space with additional financial and economic indicators did not adequately enhance the original LSTM architecture's risk-adjusted performance during challenging market conditions.
• Advanced deep learning architectures like the Transformer, combined with pre-training on extensive historical data and integration of diverse state variables, offer superior adaptability and robustness for portfolio optimization in dynamic and unforeseen market conditions.

Summary by alphaXiv.