End-to-End Neural Networks for Risk-Budgeted Portfolio Optimization

Summary

The paper combines return prediction and portfolio construction in a single neural network, aiming to reduce the decision errors that can arise when forecasts are optimized separately. It compares a model-free network, which learns allocations directly, with a model-based approach that predicts asset risk contributions and passes them to a differentiable risk-budgeting optimization layer. Risk parity is presented as the equal-risk-contribution case of risk budgeting, a framework that can diversify risk without relying on expected-return forecasts.

The reported out-of-sample tests for 2017–2021 give the model-based approach a Sharpe ratio of 1.16, compared with 0.79 for nominal risk parity and 0.83 for a fixed equal-weight portfolio. The authors also add stochastic gates to select assets, addressing the possibility that risk-based portfolios include low-volatility, low-return assets; the gated approach with filtering reports a Sharpe ratio of 1.24. These results are specific to the study’s data and test period. The document is an introduction rather than the full paper, so it does not provide enough detail to assess implementation choices or robustness.

Key ideas

• End-to-end training aligns portfolio decisions with a portfolio-level objective instead of optimizing forecasts in isolation.
• The model-based network predicts risk budgets and uses a differentiable optimization layer to translate them into allocations.
• Risk parity is a special case of risk budgeting in which assets receive equal risk contributions.
• Stochastic asset-selection gates are proposed to screen out assets that may contribute little return despite low volatility.
• The reported Sharpe ratios are historical out-of-sample results for a single stated period and do not establish future performance.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.