Forecaster

CiscoTSMForecaster

Zero-shot univariate forecaster using the Cisco Time Series Model (CTSM).

CTSM 1.0 is a 250M-parameter, decoder-only transformer foundation model developed by Cisco (Splunk) for univariate zero-shot time series forecasting [1]. It uses a multiresolution architecture: internally it derives a coarse-resolution context (60 sparser than the input) and a fine-resolution context from a single time series, then predicts up to 128 steps ahead. Long-horizon forecasting beyond 128 steps is supported via autoregressive rolling.

Calling fit only stores the training context and loads the model weights. No model training or fine-tuning is performed.

Quickstart

python

from sktime.forecasting.cisco_tsm import CiscoTSMForecaster

estimator = CiscoTSMForecaster(model_path: str='cisco-ai/cisco-time-series-model-1.0', num_layers: int=25, backend: str='cpu', context_length=None, quantiles=None, ignore_deps: bool=False)

Parameters(6)

model_pathstr, default=”cisco-ai/cisco-time-series-model-1.0”

HuggingFace repository ID for the CTSM checkpoint. Use "cisco-ai/cisco-time-series-model-1.0-preview" for the earlier, larger 500M-parameter preview checkpoint (requires num_layers=50).

num_layersint, default=25

Number of transformer layers. Use 25 for CTSM 1.0 and 50 for 1.0-preview.

backendstr, default=”cpu”

Hardware backend: "cpu" or "gpu". When set to "gpu", the model is placed on the first available CUDA device. If no GPU is found, the package falls back to CPU automatically.

context_lengthint or None, default=None

Maximum number of most-recent observations to pass as context. If None, the full training series (up to the model’s internal maximum of 30 720 points) is used. Shorter contexts reduce memory usage but may degrade forecast quality.

quantileslist of float or None, default=None

Quantile levels pre-computed by the model. If None, the official 15-quantile set is used: ``[0.01, 0.05, 0.1, 0.2, 0.25, 0.3, 0.4, 0.5,

0.6, 0.7, 0.75, 0.8, 0.9, 0.95, 0.99]``.

These levels are available via predict_quantiles / predict_interval. Arbitrary alpha values not in this set are handled by linear interpolation over the available levels.

ignore_depsbool, default=False

If True, soft-dependency checks for cisco-tsm and torch are skipped. Useful for testing the sktime adapter contract without the optional packages installed.

Examples

>>> from sktime.datasets import load_airline
>>> from sktime.forecasting.cisco_tsm import CiscoTSMForecaster
>>> y = load_airline ()
>>> forecaster = CiscoTSMForecaster ()
>>> forecaster. fit (y) CiscoTSMForecaster(
... )
>>> y_pred = forecaster. predict (fh = [1, 2, 3 ])
>>> pred_int = forecaster. predict_interval (fh = [1, 2, 3 ], coverage = 0.9)

References

[1]

Liang Gou et al. “Cisco Time Series Model Technical Report.” arXiv:2511.19841, 2025. https://arxiv.org/abs/2511.19841