Portfolio Weights Under Correlation Estimate Uncertainty

Summary

This document replaces handpicked correlation “candidate matrices” in a three-asset portfolio method with weights averaged across plausible correlation estimates. It uses Fisher’s transformation to form a sampling distribution for each pairwise correlation from its estimate and sample size, then optimizes weights at selected points in those distributions. The examples show how uncertainty in one pair’s correlation can shift allocations even when the central estimates are all zero.

The author also compares correlation lookback periods by testing whether estimated correlations predict future correlations, and considers how lookback choice affects weight robustness. The reported analysis suggests that very short histories forecast poorly, while very long histories may modestly improve forecasts but make resulting weights less robust; shorter histories can also cause assets to change groups. These are empirical observations within the author’s data and setup, not universal thresholds. The method focuses on correlation uncertainty and assumes volatility estimates are precise enough to ignore their uncertainty; it is demonstrated for three assets and relies on an approximate parameter distribution.

Key ideas

• Use a sampling distribution for each estimated pairwise correlation instead of matching a portfolio to a fixed candidate matrix.
• Optimize portfolio weights at multiple points in the correlation distributions and average the resulting allocations.
• Fisher’s transformation makes it possible to estimate correlation uncertainty from the observed correlation and sample size.
• Lookback length affects both correlation forecasting and the robustness of the resulting weights.
• The analysis ignores uncertainty in volatility estimates and demonstrates the approach on three-asset portfolios.

Tags

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