Gaussian Processes for Local Polynomial Forecasting of Time Series

TL;DR

This paper introduces a Gaussian process-based forecasting method utilizing deep machine learning to accurately predict non-stationary, non-linear time series signals, especially when historical data for individual series is limited.

Contribution

The paper presents a novel Gaussian process regression approach that effectively forecasts multiple non-stationary time series with limited data, outperforming traditional generalized least-squares methods.

Findings

Successful forecasting of non-stationary signals with limited data

Outperforms generalized least-squares in complex scenarios

Employs deep learning techniques within Gaussian processes

Abstract

Non-stationary time series with non-linear trends are frequently encountered in applications. We consider here the feasibility of accurately forecasting the signals of multiple such time series considering jointly when the number of historic samples is inadequate for accurately forecasting the signal of each considered in isolation. We develop a new forecasting methodology based on Gaussian process regression that is successful in doing so in examples for which the method of generalized least-squares is not. The new method employs a form of deep machine learning.