Model-free prediction of time series: a nonparametric approach

We propose a novel approach for model-free time series forecasting. Unlike most existing methods, the proposed method does not rely on parametric error distributions nor assume parametric forms of the mean function, leading to broad applicability. We achieve such generality by establishing a simple but powerful representation of a time series {X-t; t is an element of Z} with sup(t) E|X-t| < infinity, that is, X-t has a solution which is a linear combination of infinite past values. Then using the obtained solution a prediction algorithm is presented, with large sample theoretical guarantees. Simulation studies show favourable performance of the proposed method compared with popular parametric and neural networks methods, and suggest its superiority when the sample size is small. An application to practical time series is discussed.

Automated summary

In this paper, a novel model-free approach for time series forecasting is proposed, which has broader applicability compared to existing methods. By establishing a simple and powerful representation of time series, a prediction algorithm is presented with theoretical guarantees. Simulation studies show that the proposed method outperforms popular parametric and neural networks methods, especially when the sample size is small. An application to practical time series is also discussed.