Reproduce “Monash Time Series Forecasting Archive” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rakshitha Godahewa, Christoph Bergmeir, Geoffrey I. Webb, Rob J. Hyndman, Pablo Montero-Manso
Monash University researchers introduce the Monash Time Series Forecasting Archive, the first comprehensive public repository comprising 20 related time series datasets tailored for benchmarking global forecasting models. The work also proposes a standardized `.tsf` data format and provides baseline performance evaluations for a range of forecasting models.
Key insights
- The feature analysis visually confirms that M-competition and Tourism yearly datasets have high trend and first-order autocorrelation, while datasets like Car Parts and Kaggle Web Traffic show higher entropy, often due to intermittency.
- Globally trained models, such as Pooled Regression, generally outperform univariate models on high-frequency datasets and intermittent time series, indicating their ability to leverage cross-series patterns.
- No single forecasting method is universally superior; model performance is highly dependent on dataset characteristics, including frequency, seasonality, and the presence of intermittency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.