When Diversification Fails: Decomposing Portfolio Risk-Forecast Errors into Asset-Volatility and Correlation Surprises, 1997–2026
When a portfolio risk forecast misses, is it because the assets became more volatile than expected or because their correlations changed? We split the errors of 20-day risk forecasts into an asset-risk part and a correlation part with an exact Shapley decomposition. Forecasts are made every five trading days from 1997 to 2026 for eleven U.S.-listed portfolios of stocks, bonds, gold, sectors, international equity, oil and the dollar, using three standard models. Asset-risk surprises account for 77% to 95% of the squared forecast error. Their share exceeds one half in every portfolio, model and sample after a Holm correction, in every year of the sample, and under alternative windows, horizons and decay factors. Forecasts lag the market: they overstate risk in calm periods and understate it when stress builds. Correlation surprises matter mainly for equity portfolios at the onset of stress, and most of them are consistent with a by-product of higher volatility. In random portfolios of up to 30 stocks, the correlation share of the error rises with the number of assets but stays below one half, largely tracking the sampling noise of short-window correlation estimates.
Authors
- Ismayilov Murad
- Ilyas Taghiyev
- Nijat Shikhaliyev
- Mikayil Ahmadbayli
Institutions
- Stanford University (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24425598
- Primary Topic
- Financial Risk and Volatility Modeling
- Type
- preprint