Abstract
We examine a variety of graphical models to construct optimal portfolios. Graphical models such as PCA-KMeans, autoencoders, dynamic clustering, and structural learning can capture the time varying patterns in the covariance matrix and allow the creation of an optimal and robust portfolio. We compared the resulting portfolios from the different models with baseline methods. In many cases our graphical strategies generated steadily increasing returns with low risk and outgrew the S&P 500 index. This work suggests that graphical models can effectively learn the temporal dependencies in time series data and are proved useful in asset management.
Introduction
Modern portfolio theory, rooted in Markowitz’s work, relies heavily on the historical covariance matrix to diversify risk. However, this approach faces two major hurdles: instability in high-dimensional spaces and non-stationarity in financial markets. A standard covariance matrix assumes a fully connected graph where every stock is linked to every other stock, which is often computationally burdensome and physically unrealistic. This research explores how graphical models—including Variational Autoencoders (VAEs) and dynamic clustering—can outperform traditional benchmarks by capturing sparse, evolving dependencies between assets.
Methodology
The researchers implemented a multi-stage pipeline using daily closing price data for US equities within the S&P 500 index (specifically Tech, Financials, and Energy sectors) from 2012 to 2020. The methodology consists of:
- Dimensionality Reduction: PCA and Variational Autoencoders (VAEs) map high-dimensional returns into a lower-dimensional latent space. A Cauchy VAE was utilized to better account for the "heavy tails" (extreme events) common in financial data.
- Dynamic Clustering: Stock clusters were updated quarterly to reflect market evolution. The study compared Agglomerative Clustering with Affinity Propagation, the latter of which discovers the optimal number of clusters based on the data.
- Structural Learning: Graphical LASSO was employed to infer sparse edge connectivities, identifying true dependencies beyond market noise.
- Portfolio Selection: Strategies were tested based on selecting stocks that were either "easy to model" or "outliers" based on their reconstruction error in the latent space.
Results and Discussion
A key insight from the study is the emergence of sector intelligence. The Cauchy VAE successfully clustered stocks by their actual industry sectors using only daily return data, without being provided any sector labels.
Quantitative performance metrics show a clear advantage for graphical methods. The Affinity Propagation dynamic clusters strategy was the standout performer, significantly outgrowing the S&P 500 over the 2012–2018 period. By using sparse, partially connected graphs, the models filtered out market noise, leading to more robust portfolios.
Furthermore, using an "Efficient Frontier" weighting strategy with the Autoencoder yielded a massive 309% return in 2016. Many graphical strategies also maintained higher Sharpe ratios than the S&P 500, indicating superior risk-adjusted performance.
Conclusion
Graphical models effectively learn temporal dependencies in time series data, providing a powerful tool for asset management. While the results are promising, future work should address weight concentration in portfolio optimization and incorporate regime-switching components to better handle bear markets. Expanding features to include trading volume or news sentiment could further enhance the predictive power of these graphical representations.
Summary updated