ABSTRACT

Another research topic in online portfolio selection (OLPS) is meta-learning, or meta-algorithms (MAs) (Das and Banerjee 2011), which is closely related to expert learning (Cesa-Bianchi and Lugosi 2006). This is directly applicable to the “fund of fund” (FOF),∗ which delegates portfolio capital to other funds. In general, MA defines several base experts, each of which is equipped with strategies from the same strategy class or different classes, or even MAs. Each expert outputs a portfolio vector, and MA combines these portfolios to form a final portfolio, which is used for rebalance. The whole system can achieve the best performance among the experts in hindsight, which thus is desired for some nonuniversal algorithms. MAs are similar to Cover’s UP algorithm in the follow the winner approach; however, they are proposed to handle different classes of experts, among which UP’s CRP becomes a special case. On the one hand, MAs can be used to smooth the final performance with respect to all experts, especially when base experts are sensitive to certain environments/parameters. On the other hand, combining universal algorithms and heuristic algorithms, which is not easy to obtain a theoretical regret bound, can provide the universality property for the whole system. Finally, MAs can be applied to all existing approaches and thus have much broader areas of application.