Periodicity Makes Perfect : Using Fourier Inspired Periodicity to Improve Long Horizon Time Series Forecasting

Abstract Long-horizon Time Series Forecasting (LTSF) is a critical area of research that allows us to plan for the long term goals specially in energy, finance, healthcare, and climate science sectors. Transformers, and their derivatives form the leading edge of Artificial Intelligence (AI) research in almost all domains. However, in LTSF, resource heavy transformer based models and light weight linear models have performed similarly across the benchmark datasets. One critical difference between time series data and other data fields is the periodicity. Thus, performance of LTSF models can be enhanced by focusing on periodicity of data. However, many famous time series models do not prioritize periodicity, during their calculations. In this study, we improve performance of three foundational time series models including 2 linear (NLinear and DLinear) as well as one based on transformer backbone (iTransformer). We evaluate the performance of models across five benchmark datasets including ECL, ETT h1 \& h2, Traffic and Weather. We demonstrate an average improvement of 3.86% and 9.51% over our base models (iTransformer and NLinear / DLinear respectively) which demonstrates that simply adding periodicity in time series models can improve performance of formative models and bring them at par with the State-Of-The-Art (SOTA) models.