Calibrating Portfolio Optimizers with Random Data and Out-of-Sample Tests
Summary
The document proposes evaluating portfolio optimization methods on simulated returns before comparing them on real assets. Its experiment uses nine assets with randomly assigned Sharpe ratios and correlations, then generates many multivariate Gaussian return histories. It varies the in-sample estimation window and assesses five-year out-of-sample Sharpe ratios across repeated histories. The setup assumes equal asset volatility, long-only weights, and a Sharpe-maximizing investor.
The methods include Monte Carlo and bootstrap optimization, plus shrinkage variants that pull estimated Sharpe ratios or correlations toward chosen reference values. Equal weighting, naive mean-variance, maximum diversification, and a correlation-shrinkage method appear as special cases or competitors. Evaluation considers both median outcomes and a cautious lower-tail percentile, with computation time and convergence also relevant. The author reports that some mean shrinkage can be optimal and that simulation-based methods serve as a performance benchmark but are slow. Random data offer a controlled calibration baseline, not a faithful model of markets; the author expects real data may require more shrinkage and flags a future real-data comparison.
Key ideas
- Random Gaussian histories provide a controlled setting for comparing portfolio optimization methods.
- The experiment evaluates out-of-sample Sharpe ratios across varying estimation lengths and repeated samples.
- Shrinkage can be applied to estimated Sharpe ratios, correlations, or both.
- Lower-tail performance matters because extreme weights can worsen downside outcomes.
- Simulation-based optimizers are treated as a benchmark, but their computational cost is a drawback.
- Results from well-behaved synthetic data may overstate robustness on real markets.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.