Resampling and Estimation Error in Mean-Variance Optimization
Summary
The document asks whether mean-variance optimization research remains worthwhile when expected-return estimates are noisy. It cites a discussion of out-of-sample results in which more accurate volatility forecasts alone did not materially improve the Sharpe ratio, then asks whether other changes to the optimization process can help.
The accepted answer argues that mean-variance optimization still merits study under suitable assumptions and points to resampling variations of the efficient frontier as examples. It reports that Michaud and Ibbotson developed separate approaches, and cites an empirical study by Markowitz and Usmen that found resampled optimization outperformed the original model in its tests. The document gives no test design, performance figures, or detailed comparison, so the cited result should not be treated as proof that resampling works across markets or settings. The broader question of expected-return estimation error remains open in the exchange.
Key ideas
- The document frames expected-return estimation error as a major concern for mean-variance optimization.
- It reports that better volatility forecasts alone did not substantially improve out-of-sample Sharpe ratios in a cited example.
- Resampling the mean-variance frontier is presented as one avenue for improving the process.
- A cited empirical study found better performance for resampled optimization, but the document provides no details on its scope or robustness.
Tags
Full text
# Improving Portfolio Optimization on a Mean-Variance Basis # Improving Portfolio Optimization on a Mean-Variance Basis Is there a point to conduct research to improve mean-variance optimization (MVO)? Because I understand that most of the poor performance in MVO is a result of the estimation error in expected returns. Even in "Advanced Portfolio Management" by Paleologo, the author shows in Chapter 6.4 that a perfect forecast of volatility does not lead to substantial improvement in the Sharpe ratio when evaluating out-of-sample performance. Main Question: Is research into other aspects of improving the MVO process fruitless unless we fix the estimation error issue w.r.t. expected returns? ## Answer by phdstudent (score 3, accepted) https://quant.stackexchange.com/a/79704 Yes there is a big benefit in doing research on improving MVO. After all, the tangency portfolio is the best portfolio under several not super crazy assumptions. There has been a lot of work and patents on MVO (the most common example is the re-sampling of the MV frontier). Both Michaud and Ibbotson patented different variations of the process: a. Michaud’s patent - https://patents.google.com/patent/US6003018A/en (patent expired) b. Ibbotson’s patent - https://patents.google.com/patent/US20030195831A1/en (patent abandoned after Michaud’s patent expired) The patents have expired, but the filing documents describe the process in more detail as well if you do not have access to the book referenced above. Harry Markowitz and Nilufer Usmen published a paper in which they found that resampled mean variance optimization performed better than the original model in empirical tests, please refer to "Resampled Frontiers vs. Diffuse Bayes: An Experiment" Journal Of Investment Management Q4 2003" for more details.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.