LG AI Research 424

Time series forecasting is a core technology that helps decision-making by predicting the future based on past data. At ICLR 2024, various methodologies were recently presented to improve the accuracy and efficiency of forecasting.

As a key research case, FITS (Frequency Interpolation Time Series Analysis Baseline) proposes an innovative methodology that leverages the frequency domain. By converting time-domain forecasting into a frequency-domain interpolation problem, it achieves performance surpassing existing SOTA models with fewer than 10,000 parameters. This enables efficient forecasting even in environments lacking high-performance computing resources.

Also, ModernTCN introduced a new Temporal Convolution Network model inspired by the Transformer architecture. By dividing time series data into multiple patches and embedding them, it effectively identifies temporal patterns, and demonstrates excellent performance not only in forecasting but also in various tasks such as anomaly detection, classification, and missing value imputation.