Updating Views by Learning from the Others: Dynamically Combining Asset Allocation Strategies
54 Pages Posted: 15 Aug 2013
Date Written: August 13, 2013
The well-known difficulties in obtaining satisfactory results with Markowitz' intuitive portfolio theory have lead to an innumerable amount of proposed advancements by researchers and practitioners. As different as these approaches are, they typically appear to exhibit a satisfactory out-of-sample performance; however, at the same time, studies show that the equally weighted portfolio still cannot be dominated by them. The starting point of our study is therefore not an(other) entirely new idea, which is based on a new strategy we claim performs well, but instead the acknowledgement that the strategies proposed in earlier studies have specific advantages, which, though not consistently apparent, might prevail in specific and possible rare situations of dynamic markets. We therefore propose a strategy that "learns from" a population of already existing strategies and dynamically combines their respective characteristics, resulting in a strategy that is expected to perform best in light of the expected/predicted market situation. We show that our approach is successful by carrying out an empirical backtest study applied in a multi-asset setting for investor clienteles with mean-variance, mean-conditional value-at-risk, and maximum Omega utility functions. The improvements of our flexible approach, which include a higher mean return and lower volatility, stay (statistically) significant, even when we take into account transaction costs and improve the competing strategies by employing robust input parameter estimates.
Keywords: Asset allocation strategies, Statistical Learning, Relative Entropy, Black and Litterman
JEL Classification: G11, G32, G17
Suggested Citation: Suggested Citation