Using cGANs to Forecast Returns and Covariance for Asset Allocation

Summary

This research report examines conditional generative adversarial networks (cGANs) for forecasting asset returns and covariance matrices, then using those forecasts in portfolio allocation. The model conditions on recent multi-asset returns and random inputs to generate possible future return paths. Repeated generations are used to estimate expected returns and covariance. The authors compare these forecasts with historical-return and historical-volatility baselines across domestic, global, and Chinese equity-sector portfolios, using risk-budgeting and mean-variance models.

The report evaluates covariance forecasts against subsequent realized covariance and assesses portfolio weights and risk contributions. It reports that cGAN covariance estimates are closer on some measures and that cGAN-based portfolios generally compare favorably with traditional approaches in the tested settings. Return forecasts appear more useful for ranking assets than for predicting their absolute returns, and some relative-return results are unstable. The evidence is historical backtesting, so it does not establish future performance. The authors also flag sensitivity to random seeds and hyperparameters, possible changes in market patterns, and the limited range of input data and test scenarios.

Key ideas

• The model generates future multi-asset return paths conditioned on recent returns and random inputs.
• Repeated simulations provide estimates of both expected returns and covariance.
• The report compares cGAN forecasts with historical-return and historical-volatility baselines.
• Its tests cover risk-budgeting and mean-variance allocation in several asset universes.
• The authors identify sensitivity to model settings and limits to the tested scenarios.

Tags

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