MetaSieve: Performance vs. Complexity Sieve for Time Series Forecasting

Motivated by the problem of finding optimal Performance vs. Complexity trade-off in the task of forecasting time series data, we propose a model-agnostic method MetaSieve that performs data dichotomy (i.e., in fact, sieves the data instances in a meta-learning manner) according to a chosen quality level while iterating over the model's complexity. The method is inspired by classical iterative numerical optimization ones but is applied to sets of time series. As a result, the method is significantly less time consuming than a traditional brute force-based meta-learning algorithm. It further turns out in the experiments that the MetaSieve quality results are rather comparable to those of the brute force-based one thus one has a noticeable reduction in time consumption in exchange for a slight decrease of forecasting quality. Additionally, we experimentally show a good performance of a MetaSieve-based classifier that provides the Performance vs. Complexity classes a priori, i.e. before the actual forecasting, on synthetic and real-world time series data.