sktime / skbase adapters

For the sktime forecasting ecosystem, tsbootstrap.adapters provides thin skbase.BaseObject estimator classes. Each adapter stores its parameters, builds the corresponding MethodSpec, and delegates to tsbootstrap.bootstrap().

These adapters are compatible with sktime’s check_estimator validation and can be passed anywhere the sktime API expects a bootstrap estimator.

Installation

The adapters require skbase (part of the sktime ecosystem):

uv add "tsbootstrap[sktime]"     # or: pip install "tsbootstrap[sktime]"

Using an adapter

from tsbootstrap.adapters import MovingBlockBootstrap
import numpy as np

x = np.random.default_rng(0).standard_normal(200)

est = MovingBlockBootstrap(block_length=10, n_bootstraps=100, random_state=0)

for sample in est.bootstrap(x):
    ...   # sample is an ndarray of shape (n,)

# With observation indices
for values, indices in est.bootstrap(x, return_indices=True):
    print(values.shape, indices)

All adapters accept n_bootstraps and random_state in their constructor. The bootstrap(X, y=None, return_indices=False) method yields samples.

Available adapters

Block method adapters

Class	Underlying spec	Key parameter(s)
`IIDBootstrap`	`IID`
`MovingBlockBootstrap`	`MovingBlock`	`block_length`
`CircularBlockBootstrap`	`CircularBlock`	`block_length`
`StationaryBlockBootstrap`	`StationaryBlock`	`avg_block_length`
`NonOverlappingBlockBootstrap`	`NonOverlappingBlock`	`block_length`
`TaperedBlockBootstrap`	`TaperedBlock`	`window`, `block_length`, `alpha`

Model method adapters

Class	Underlying spec	Key parameter(s)
`ARResidualBootstrap`	`ResidualBootstrap(model=AR(...))`	`order`
`ARIMAResidualBootstrap`	`ResidualBootstrap(model=ARIMA(...))`	`order`
`VARResidualBootstrap`	`ResidualBootstrap(model=VAR(...))`	`order`
`SieveBootstrap`	`SieveAR`	`min_lag`, `max_lag`, `criterion`

Validating with sktime

from sktime.utils import check_estimator
from tsbootstrap.adapters import MovingBlockBootstrap

check_estimator(MovingBlockBootstrap)

Relationship to the functional API

The adapters are strictly wrappers; there is no additional logic. Users who do not need sktime compatibility should use tsbootstrap.bootstrap() directly - it returns a richer BootstrapResult (with provenance metadata and OOB primitives) rather than a bare iterator of arrays.