The Multiverse Across Asset Classes: Design Uncertainty in Asset Allocations
Résumé
This paper documents the performance sensitivity of asset allocation methods with respect to design choices in the backtests. Endowed with five asset classes, we document the variations in Sharpe ratio of strategies with alternative (i) utility functions, (ii) signal-generating algorithms, (iii) sample periods, (iv) rebalancing frequency and (v) leeway with respect to a given benchmark, i.e, tracking error constraints. Our results show that while risk aversion does not impact risk-adjusted performance much (risk and return vary together), all other options can either significantly boost or deteriorate Sharpe ratios, especially signal source and inception date. Standard machine learning predictions nevertheless appear to deliver superior performance in a large majority of empirical designs.
