Skip to main content
SynAugment.mbb and SynAugment.dba transform observed series directly. They do not fit a parametric generator and deliberately bypass the mean, standard deviation, and lag-1 autocorrelation pinning used by SynAugment.augment. Use only training observations as input. Augmenting from validation or test values leaks future information into model training.

​
Moving-block bootstrap

MBB separates a series into classical moving-average trend, seasonal, and remainder components. It resamples overlapping remainder blocks and combines them with the original trend and seasonal components.
When seasonal_period=None, SynForecast attempts to detect a period. When block_size=None, it uses twice the detected period, or approximately the square root of the series length without seasonality. The effective block size is capped at max(2, len(series) // 3) so multiple blocks contribute even for short inputs. MBB is a good fit when local residual dependence matters and the observed trend and seasonal shape should remain anchored. Each input series needs at least four observations and one finite value. Missing values are interpolated before decomposition.

​
DTW barycenter averaging

DBA creates a synthetic series around each usable reference series. It finds nearest neighbours with banded dynamic time warping (DTW), randomizes their weights, and iteratively computes a weighted barycenter.
With scale="reference", series are z-normalized for alignment and each barycenter is mapped back to its reference mean and standard deviation. Use scale="none" only when the panel already shares a meaningful common scale. The output length and timestamps follow the reference series, including when its neighbours have different lengths. DBA needs at least two usable series. DTW time and memory grow with series length times band width, so reduce window_fraction, n_neighbors, or n_iterations for long panels. Pairwise panel distances are computed once per call in bounded parallel chunks in native Rust. Only the requested nearest neighbors are retained, using O(number of series × n_neighbors) storage; distance computation still examines every pair. Series without a finite observation are skipped with a logged warning; mbb also skips series shorter than four observations. MBB computes its moving-average trend in linear time and reuses each source’s decomposition across all requested copies.

​
Output contract

Both methods:
  • return the same dataframe backend as the input unless engine is set;
  • copy all non-target columns from the source/reference series unchanged;
  • generate collision-safe string IDs such as store_1_mbb_0 or store_1_dba_0;
  • are deterministic for a fixed SynAugment seed; and
  • run decomposition, block sampling, DTW, and barycenter updates in native Rust.
include_original=False returns only synthetic rows. Always compare augmented training against an observed-only baseline over temporal holdouts; neither method guarantees better forecast accuracy. See SynAugment: data augmentation for fit-and-simulate and TSMixup workflows, and the API reference for every parameter.