# Double Descent in Portfolio Optimization: Dance between Theoretical Sharpe Ratio and Estimation Accuracy

**Authors:** Yonghe Lu, Yanrong Yang, Terry Zhang  
**arXiv:** [2411.18830](https://arxiv.org/abs/2411.18830)  
**Date:** 2024-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2411.18830v1/

## Key points

- Performance of low-dimensional models initially improves with complexity but eventually declines due to overfitting.
- Resurgence in performance with increasing complexity after reaching a certain threshold, leading to a double ascent Sharpe ratio curve.
- Delicate balance between embracing higher complexity under specific conditions.
- Navigating the trade-off between data availability and incorporating information flexibly in portfolio construction.
- Extending quantitative analysis to other portfolios like the minimum variance portfolio to uncover similar patterns of overparameterization and factor strength.
- Exploring the "double descent" phenomenon within AI-driven large-scale portfolio theory for potential revolution in portfolio management practices.
- Contribution to existing literature and theoretical advancements:
- Benefits of expanding model complexity into high-dimensional regimes.
- Superior performance in high-dimensional settings through specific inductive biases like ℓ2-norm based approaches.
- Intersection with asset pricing and artificial intelligence field:
- Building upon empirical success of machine learning models in prediction tasks and efficient portfolio construction.
- Developing a theory of complex portfolios grounded in individual asset modeling for valuable insights into evolving landscape.
- Understanding generalization behavior related to model complexity:
- Synthesizing insights from finance, artificial intelligence, and machine learning theory for innovative approaches towards optimizing complex portfolios while mitigating risks associated with overfitting.

## Summary

In the realm of mean-variance portfolio optimization, our study delves into the intricate relationship between model complexity and out-of-sample performance. By representing model complexity through the number of assets, we uncover a fascinating phenomenon where the performance of low-dimensional models initially improves with complexity but eventually declines due to overfitting. However, as model complexity reaches a certain threshold, there is a resurgence in performance with increasing complexity, leading to a double ascent Sharpe ratio curve reminiscent of the double descent observed in artificial intelligence. Our research sheds light on the delicate balance between embracing higher complexity under specific conditions for optimal results. This framework offers practitioners valuable insights into navigating the trade-off between data availability and incorporating information flexibly in portfolio construction. Looking ahead, there are promising avenues for further exploration. One direction involves extending our quantitative analysis to other types of portfolios such as the minimum variance portfolio to uncover similar patterns of overparameterization and factor strength influencing their performance. Furthermore, delving into the "double descent" phenomenon within the application of artificial intelligence to large-scale portfolio theory presents an exciting prospect. Understanding how AI-driven models can leverage complexity without succumbing to overfitting could revolutionize portfolio management practices. Expanding our study in these directions holds significant potential for both theoretical advancements and practical applications in financial portfolio optimization. Additionally, our work contributes to existing literature by highlighting the benefits of expanding model complexity into high-dimensional regimes. By showcasing that superior performance can manifest in high-dimensional settings through specific inductive biases like ℓ2-norm based approaches, we open up new avenues for exploring whether existing estimation methods can yield further gains when combined with high-dimensional models. Moreover, our paper intersects with the burgeoning field studying asset pricing in the era of artificial intelligence. We build upon previous research demonstrating machine learning models' empirical success in various prediction tasks and efficient portfolio construction. By developing a theory of complex portfolios grounded in individual asset modeling and elucidating the relationship between theoretical Sharpe ratios and observed Sharpe ratios, we contribute valuable insights to this evolving landscape. Lastly, drawing from machine learning theory on risk curves like "double descent," our study adds depth by providing a nuanced understanding of how model complexity impacts generalization behavior across different setups. By synthesizing insights from diverse fields like finance, artificial intelligence, and machine learning theory, our research paves the way for innovative approaches towards optimizing complex portfolios while mitigating risks associated with overfitting and underperformance.