Title Comparative analysis of time series forecasting models in m competitions /
Translation of Title Laiko eilučių prognozavimo palyginamoji analizė naudojant M konkursų duomenis.
Authors Bagdonas, Mantas
Full Text Download
Pages 34
Keywords [eng] DLinear, FITS, time series, M5 competition, MOFC, demand forecasting, linear regression, frequency domain
Abstract [eng] This thesis examines modern time series forecasting methods using the complex M5 competition dataset, which includes over 30,000 hierarchical time series of daily Walmart product demand. Lightweight models like DLinear and FITS are compared with traditional methods such as ARIMAX and machine learning models like LightGBM. The results demonstrate that simple, linear models can produce forecasts comparable to resource-intensive methods while maintaining efficiency. It was also found, that the most prominent traditional time series models proposed for the M5 competition as benchmarks were outperformed by DLinear and FITS with wide margins. The study supports the proposition of DLinear as a new baseline for time series forecasting due to its low complexity and competitive accuracy, offering practical recommendations for practitioners to advance forecasting methods effectively.
Dissertation Institution Vilniaus universitetas.
Type Master thesis
Language English
Publication date 2025