Real-Data Instability in Sharpe Ratio and Correlation Forecasts

Summary

This post compares how well trading-strategy Sharpe ratios and return correlations can be forecast from real data versus simulated returns drawn from a fixed distribution. The described experiment samples strategy components, measures one-year-ahead estimation error, and compares real-data RMSE with RMSE from simulated returns matched to observed Sharpe ratios and correlations. Across thousands of sampled components, the median error ratio is reported as 1.06 for Sharpe ratios and about 5.6 for correlations.

The author interprets this gap as evidence that strategy correlations are much less stable in actual markets than under a stationary model, while Sharpe ratio forecasting is already difficult even in the simulated setting. The post connects the result to correlation shrinkage in portfolio optimisation, despite earlier simulation work suggesting no shrinkage was optimal. The figures are specific to the author's sample and setup; the extract gives no detailed data description or uncertainty estimates, and the broader claim depends on the chosen strategies and assumptions.

Key ideas

• The experiment compares forecast RMSE on real strategy returns with RMSE on simulated returns from a fixed distribution.
• The reported median real-to-simulated RMSE ratio is 1.06 for Sharpe ratios and about 5.6 for correlations.
• The author finds strategy return correlations substantially less stable in real data than in the simulation.
• The findings support considering correlation shrinkage when estimating portfolio inputs from real observations.
• The reported comparisons depend on the sampling design and do not establish universal forecast-error levels.

Tags

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