Decomposition Guide: English Sharing Kit
Prepared 29 October 2026. These are reusable drafts; their presence in this repository does not indicate publication on an external platform.
Each draft links to the full comparison, which contains the official sources and the scope of the comparison. Use the article URL after the documentation deployment is live:
https://systems-mechanobiology.github.io/DeTime/research/python-decomposition-tools/
LinkedIn or a research-community post
STL, SSA, wavelets, EMD, or VMD: what are you actually choosing?
A decomposition comparison needs more than a list of methods.
Which seasonal periods are supplied? How are modes grouped? Are outputs coefficients or sample-aligned components? Does preprocessing use observations that would be unavailable when the forecast is issued?
We put together a practical guide to statsmodels, PyWavelets, PyEMD, emd, pyts, SSALib, sktime, PySDKit and DeTime. It covers the documented interfaces, reproducibility choices and evaluation pitfalls, with links to the official sources.
One useful check: if the residual is defined as input minus trend minus seasonality, exact reconstruction is expected. Evidence that the estimated components represent the process needs a separate test.
We maintain DeTime, a Python and CLI toolkit for running decomposition families through a shared configuration and result interface. The guide also explains when a specialist package is the more useful starting point.
Read the guide: https://systems-mechanobiology.github.io/DeTime/research/python-decomposition-tools/
Short post
STL, SSA, wavelets, EMD, or VMD? We compared Python decomposition tools by assumptions, outputs, reproducibility and forecasting workflow. From the DeTime maintainers: https://systems-mechanobiology.github.io/DeTime/research/python-decomposition-tools/
A five-post thread
1
STL, SSA, wavelets, EMD, or VMD? Start with the structure you want to estimate, the observations available at prediction time, and the output your next step needs. We maintain DeTime and wrote a practical comparison of the Python tools.
2
STL uses a seasonal period; MSTL supports several. Wavelets expose scale coefficients; SSA needs a window and grouping rule; EMD-family methods and VMD produce modes. Record the choices that turn these outputs into your features.
3
If residual = input - trend - season, reconstruction is exact by construction. That checks arithmetic consistency. Testing component identification needs known components, independent measurements, or another justified reference.
4
For a forecast at time t, decompose only the history available at t. Computing components on the full record and slicing afterward can retain future information. Reproduce the information available at deployment.
5
The guide covers statsmodels, PyWavelets, PyEMD, emd, pyts, SSALib, sktime, PySDKit and DeTime, with official sources and the scope of each comparison: https://systems-mechanobiology.github.io/DeTime/research/python-decomposition-tools/
Newsletter blurb
A practical map of Python decomposition tools. This guide compares STL, MSTL, SSA, wavelets, EMD-family methods and VMD by their assumptions, outputs and reproducibility controls. It also explains why reconstruction error alone cannot establish component recovery, and how to keep future observations out of forecasting preprocessing. Written by the DeTime maintainers, with links to official documentation and a distinction between implemented software, proposed evaluations and measured evidence.
Repository description for profiles
DeTime provides Python and CLI workflows for time-series decomposition, with a shared configuration, trend/season/residual results, backend metadata and machine-readable schemas. The project includes method references, tutorials and a source-linked guide to choosing decomposition tools.