Best investment strategy selection using asymptotic meta learning

Meta learning is an advanced field of machine learning where automatic learning algorithms are applied to acquire meta-knowledge for a set of learning algorithms called base learners. One of meta-learning purposes is to select the best base learners for certain kind of data set to support future learning process. Comparing average out-of-sample predictability with data bootstrapping is one of popular meta-learning algorithms to measure the performance of each base learner for time series data. The challenge is that it is a very time-consuming for data analytics, such as quantitative investment time series modeling. In order to complete the meta-learning process on time before new time-series data arrive, we need to optimally allocate the limited computation budget to each learner. In this paper, we propose the Asymptotic Meta Learning (AML) to data bootstrapping process during Meta learning, which is one of asymptotic ordinal optimization algorithm for mean measure of designs with random outputs. The numerical experiments are conducted to illustrate its efficiency.