Gaussian Process-Based Corruption-resilience Forecasting Models

TL;DR — Enhancing time series forecasting by denoising Gaussian Process corruption via a corrupt-denoise-forecasting framework

Abstract

Time series forecasting is challenging due to complex temporal dependencies and unobserved external factors, which can lead to incorrect predictions by even the best forecasting models. Using more training data is one way to improve the accuracy, but this source is often limited. In contrast, we are building on successful denoising approaches for image generation. When a time series is corrupted by the common isotropic Gaussian noise, it yields unnaturally behaving time series. To avoid generating unnaturally behaving time series that do not represent the true error mode in modern forecasting models, we propose to employ Gaussian Processes to generate smoothly-correlated corrupted time series. However, instead of directly corrupting the training data, we propose a joint forecast-corrupt-denoise model to encourage the forecasting model to focus on accurately predicting coarse-grained behavior, while the denoising model focuses on capturing fine-grained behavior. All three parts are interacting via a corruption model which enforces the model to be resilient. Our extensive experiments demonstrate that our proposed corruption-resilient forecasting approach is able to improve the forecasting accuracy of several state-of-the-art forecasting models as well as several other denoising approaches