Estimation risk is the loss of investor utility from building a portfolio on sample estimates rather than true, unknown parameters. Classical mean-variance optimization treats the sample mean as if it were the true expected return, which ignores this uncertainty. In "Bayes-Stein Estimation for Portfolio Analysis," Jorion (1986) simulated returns for seven countries and a World index (T = 60, January 1977–December 1981). The Bayes-Stein estimator's gain over the diffuse prior ranged from 8 percent per annum (T=25) to 0.2 percent per annum (T=200).
What the Study Found
From T = 25 to T = 200, Bayes-Stein had lower empirical risk than Certainty Equivalence and Bayes Diffuse Prior. At T = 25, empirical risk was 1.5606 for Certainty Equivalence, 0.3452 for Bayes Diffuse Prior, 0.1337 for Minimum Variance, and 0.1815 for Bayes-Stein. In risk-free equivalent return, Bayes-Stein's gain over the diffuse prior fell from 8 percent (T=25) to 2 percent (T=50) to 0.2 percent (T=200) per annum. When the highest true mean rose from 22 percent to 44 percent per annum, shrinkage gains were cut in half on average. Bayes-Stein still uniformly dominated the sample mean even with the highest true mean at 44 percent per annum.
Methodology
Jorion calibrated "true" means and a covariance matrix to monthly dollar returns for seven countries and a World index. The calibration sample contained T = 60 monthly observations. The sample period ran from January 1977 to December 1981. Returns were simulated via the IMSL GGNSM subroutine for T = 25 to 200, comparing four estimators over K = 1,000 draws.
Key Statistics
| Metric | Finding | Context |
|---|---|---|
| Bayes-Stein risk-free equivalent gain over Bayes Diffuse Prior | 8% (T=25) to 2% (T=50) to 0.2% (T=200) per annum | Simulation analysis across sample sizes |
| Empirical risk, Bayes-Stein vs. Certainty Equivalence | 0.1815 vs. 1.5606 (T=25); 0.0205 vs. 0.0253 (T=200) | Table 2, empirical risk functions |
| Shrinkage factor (mean, std. dev.) | 0.5883, 0.1564 (T=25); 0.3164, 0.0906 (T=200) | Table 2, shrinkage toward grand mean |
| Absolute risk aversion | 1/(52.2% p.a.) | Negative exponential utility calibration used in the simulation |
| James-Stein shrinkage coefficient | ŵ = min(1, [(N−2)/T] / [(Y−Y0·1)′Σ⁻¹(Y−Y0·1)]) | Eq. 11, defines the shrinkage weight applied to sample means |
| Empirical Bayes shrinkage coefficient | ŵ = (N+2) / [(N+2) + (Y−Y0·1)′TΣ⁻¹(Y−Y0·1)] | Eq. 17, shrinkage coefficient with λ estimated from the data |
Why This Matters
Portfolio construction built on raw sample means treats noisy historical averages as if they were certain. Shrinking each asset's return estimate toward a common grand mean counteracts this overfitting without a fully specified subjective prior. Because the shrinkage factor comes from the data itself, the correction adapts to how much dispersion the sample actually shows. Portfolio managers gain a data-driven alternative to ad hoc adjustments when return estimates are unreliable.