Third Annual Risk Conference 2025
What it is about
Alejandro Rodríguez Domínguez, director of Quantitative Analysis and Artificial Intelligence at Miraltabank, presented at the Third Annual Risk Conference (September 2025) a methodology for building portfolios that incorporates the causality and dynamics of assets, using automatic differentiation in neural networks. It is a technical presentation.
The 23 slides start from the limitations of modern portfolio theory and factor investing (spurious correlations, non-linear effects, instability over time), set out a model of asset dynamics with differential equations and Reichenbach's common cause principle, and present the Hierarchical Sensitivity Parity method, with numerical results against other strategies.
Objectives
- Explain the problems of traditional portfolio optimization and of factor investing.
- Present a framework that selects common causal factors and models the dynamics of diversification.
- Show the Hierarchical Sensitivity Parity method and its performance against alternatives.
Key takeaways
- Sample estimates in modern portfolio theory perform poorly out of sample because of uncertainty in the parameters.
- Factor investing faces spurious correlations, non-linear effects, temporal instability and multicollinearity.
- The proposal uses asset and portfolio dynamics with respect to factors that are optimal in causality and persistence, based on Reichenbach's common cause principle.
- The portfolio's causal sensitivities are estimated with neural networks and automatic differentiation, and the Hierarchical Sensitivity Parity method groups them through hierarchical clustering.
- The results are compared with strategies such as minimum volatility, 1/N and hierarchical risk parity; the material is a fund's presentation and does not constitute investment advice.
Connect with our networks