Time-Series Decomposition Methods in DeTime

Start with the four core maintained paths. Move to wrappers only when the data or comparison need points there.

Core method cards

Decision flow

Wrapper use cases

Method	Typical scenario	First parameter to check
STL	classical seasonal-trend baseline for one known period	period
MSTL	one series with multiple seasonal periods	periods
ROBUST_STL	STL-style baseline with outliers	period
EMD	adaptive IMF inspection for nonlinear signals	n_imfs
CEEMDAN	noise-assisted EMD when mode stability matters	trials
VMD	band-limited modes with a chosen mode count	K, alpha
WAVELET	multiscale signal inspection	wavelet, level
MVMD	optional multivariate VMD through PySDKit	K, alpha
MEMD	optional multivariate EMD through PySDKit	primary_period

Animated method explanations

Explore all 31 registered method entries through a four-step GIF, an English explanation, parameter guidance, and the numerical arrays behind the plots. Open a method and choose Play animation. Each page also provides a static image and direct GIF download. Examples use synthetic signals chosen to expose the mechanism.

Aliases and experimental blocks are identified below. Their pages explain the operator implemented by DeTime and its relationship to the source architecture.

Smoothing, seasonal structure, and subspaces

Method	Mechanism to watch	Scope
MA_BASELINE	Slide a uniform window → Subtract the smooth trend → Average seasonal phases → Center and form residual	stable
MSSA	Stack aligned windows → Compute shared singular modes → Reconstruct each channel → Group trend and season	maintained
MSTL	Specify the seasonal periods → Estimate seasonal terms → Add back and refine each term → Aggregate the common output	stable
ROBUST_STL	Fit initial smooths → Compute robustness weights → Refit and update the weights → Retain a residual	stable
SSA	Embed overlapping windows → Factor the trajectory matrix → Average each anti-diagonal → Group trend and season	maintained
STD	Partition and center cycles → Separate dispersion and shape → Restore each cycle's shape → Compute the common result	maintained
STDR	Partition and center cycles → Separate dispersion and shape → Average shapes across cycles → Compute the common result	maintained
STL	Form seasonal-phase subseries → Use local polynomial smoothing → Alternate seasonal and trend → Form the additive result	stable

Adaptive modes, wavelets, and time-frequency structure

Method	Mechanism to watch	Scope
CEEMDAN	Generate perturbations → Average first IMFs → Update adaptively → Retain the residue	stable
EMD	Locate extrema → Build envelopes → Sift → Remove the IMF	stable
GABOR_CLUSTER	Window and transform → Build atom features → Assign the codebook → Mask and reconstruct	experimental
MEMD	Project the data → Construct vector envelopes → Sift jointly → Extract aligned modes	optional-backend
MVMD	Align the channels → Update channel spectra → Pool spectral energy → Recover mode stacks	optional-backend
VMD	Prepare the spectrum → Update a mode → Update its center → Reconstruct and group	stable
WAVELET	Choose the filter bank → Build the pyramid → Select coefficient blocks → Assign outputs	stable

Experimental decomposition blocks

Method	Mechanism to watch	Scope
AMD_BLOCK	Compute several smooth trends → Average equal seasonal phases → Repeat the phase template → Inspect the derived residual	experimental
AUTOFORMER_BLOCK	Specify the filter → Move along the series → Subtract the trend → Inspect the output contract	Alias of MOVING_AVERAGE_DECOMPOSITION_BLOCK
DELELSTM_BLOCK	Update level and slope → Average equal seasonal phases → Repeat the phase template → Inspect the derived residual	experimental
DLINEAR_BLOCK	Specify the filter → Move along the series → Subtract the trend → Inspect the output contract	Alias of MOVING_AVERAGE_DECOMPOSITION_BLOCK
FREQMOE_BLOCK	Separate trend and selection input → Select discrete spectral peaks → Fit and average candidate seasons → Inspect actual output streams	experimental
INPARFORMER_BLOCK	Smooth the prefix → Average equal seasonal phases → Repeat the phase template → Inspect the derived residual	experimental
LEDDAM_BLOCK	Specify the filter → Move along the series → Subtract the trend → Inspect the output contract	experimental
MOVING_AVERAGE_DECOMPOSITION_BLOCK	Specify the filter → Move along the series → Subtract the trend → Inspect the output contract	experimental
NBEATS_INTERPRETABLE	Construct fixed basis functions → Separate backcast and output weights → Pass a remainder to seasonal blocks → Distinguish the two residuals	experimental
PARSIMONY_BLOCK	Remove a smoothed trend → Build the harmonic design → Add fitted basis terms → Inspect the decomposition	experimental
ST_MTM_BLOCK	Smooth the prefix → Average equal seasonal phases → Smooth the repeated template → Inspect the derived residual	experimental
TIMEKAN_BLOCK	Smooth and inspect prefix frequencies → Fit two kinds of seasonal estimate → Apply the implemented fixed blend → Inspect the checked output	experimental
TIMES2D_BLOCK	Separate trend and selection input → Select discrete spectral peaks → Fit and average candidate seasons → Inspect actual output streams	experimental
WAVEFORM_BLOCK	Analyze with fixed wavelet filters → Reconstruct the approximation → Reconstruct selected detail bands → Inspect the remaining detail	experimental
WAVELETMIXER_BLOCK	Analyze with fixed wavelet filters → Reconstruct the approximation → Reconstruct selected detail bands → Inspect the remaining detail	experimental
XPATCH_BLOCK	Mix observation and state → Advance the recurrence → Smooth the difference → Inspect all three streams	experimental

Reproduce or audit the animations.

More detail

• Method Matrix gives a compact table across input mode, maturity, dependencies, parameters, outputs, and recommended use.
• Config Reference documents exact DecompositionConfig fields and method-specific parameters.
• Method References collects primary literature and official package links.
• Notebook Gallery shows runnable plots for the retained method surface.