bootstrap()

tsbootstrap.bootstrap(X: object, *, method: BaseMethodSpec, n_bootstraps: int = 999, random_state: int | Generator | SeedSequence | None = None, exog: object | None = None, dtype: Literal['float64', 'float32'] = 'float64', backend: Literal['numpy', 'compiled'] = 'numpy') → BootstrapResult[source]

Generate bootstrap replicates of a time series.

Parameters:

  • X (array-like) – Observations, shape (n,) or (n, d).

  • method (BaseMethodSpec) – A method spec (e.g. MovingBlock(block_length="auto")).

  • n_bootstraps (int, default 999) – Number of replicates.

  • random_state (int | numpy Generator | SeedSequence | None) – Reproducibility seed. Replicate i is bound to its own generator, so results are reproducible for a given seed and environment (OS, hardware, BLAS, NumPy), as with NumPy/scikit-learn.

  • exog (array-like or None) – Optional exogenous regressors, shape (n,) or (n, k), aligned with X, held fixed during regeneration. Supported for ResidualBootstrap with an AR (ARX), VAR (VARX), or ARIMA (ARIMAX) model, and for SieveAR. ARX/VARX require initial="fixed" and burn_in=0 (the exog must align with each step); ARIMAX has no such constraint (exog enters at the level after inverse-differencing).

  • dtype ({"float64", "float32"}, default "float64") – Precision of the returned replicate values. The model fit, autocovariance, and every reduction always run in float64; only the final simulation/path tensor is cast to dtype, halving its memory at float32 for large B. Lower precisions are reserved for a future GPU backend.

  • backend ({"numpy", "compiled"}, default "numpy") – "numpy" is the default reproducible path (one PCG64 stream per replicate). "compiled" selects an opt-in numba kernel that builds indices and gathers in one replicate-parallel pass, a large speed-up on the observation methods (IID and the block families). It uses a distinct counter-based RNG stream with its own reproducibility goldens, so its replicates are equal in distribution to the numpy path but not bit-identical, and it is never engaged unless requested. It does not support the recursive (model-based) methods and requires the [accel] extra.

Returns:

Sequence of BootstrapSample plus metadata.

Return type:

BootstrapResult