Data-Centric Benchmarking of Neural Network Architectures for the Univariate Time Series Forecasting Task

Time series forecasting has witnessed a rapid proliferation of novel neural network approaches in recent times. However, performances in terms of benchmarking results are generally not consistent, and it is complicated to determine in which cases one approach fits better than another. Therefore, we propose adopting a data-centric perspective for benchmarking neural network architectures on time series forecasting by generating ad hoc synthetic datasets. In particular, we combine sinusoidal functions to synthesize univariate time series data for multi-input-multi-output prediction tasks. We compare the most popular architectures for time series, namely long short-term memory (LSTM) networks, convolutional neural networks (CNNs), and transformers, and directly connect their performance with different controlled data characteristics, such as the sequence length, noise and frequency, and delay length. Our findings suggest that transformers are the best architecture for dealing with different delay lengths. In contrast, for different noise and frequency levels and different sequence lengths, LSTM is the best-performing architecture by a significant amount. Based on our insights, we derive recommendations which allow machine learning (ML) practitioners to decide which architecture to apply, given the dataset's characteristics.