[ICLR 2024] Advances in Time Series Forecasting - LG AI Research Blog

As innovation in the time series forecasting field has recently accelerated, ICLR 2024 presented various methodologies to improve forecasting accuracy and efficiency. The key research models are as follows.

• FITS (Frequency Interpolation Time Series Analysis Baseline): Leveraging the frequency domain, it achieves excellent performance with as few as about 10,000 parameters. It exploits the equivalence between forecasting in the time domain and interpolation in the frequency domain, and its computational efficiency is very high, making it useful even in low-spec environments.

• ModernTCN: An innovative Convolutional Network architecture inspired by the Transformer structure. By dividing time series data into patches and using a new Convolution structure, it shows high performance and fast training speed not only for forecasting but also for various tasks such as anomaly detection, classification, and imputation.

• TimeMixer: Captures Trend and Seasonality of time series through sampling at various scales. It adopts a top-down (Coarse-to-fine) Feature Mixing strategy for Trend and a bottom-up (Fine-to-coarse) Feature Mixing strategy for Seasonality, securing forecasting performance that surpasses existing models.