Foundational Time Series Model Multi Series Cross Validation

The frequency of the data represented as a string. 'D' for daily, 'M' for monthly, 'H' for hourly, and 'W' for weekly frequencies are available.

Number of windows to evaluate.

The forecasting horizon. This represents the number of time steps into the future that the forecast should predict.

Forecast across the entire series history (the add_history use case). The horizon and number of windows are derived server-side (any supplied h / n_windows are ignored), and the exogenous model is refit a bounded number of times to keep whole-history requests fast. Has no effect without exogenous features.

Model to use as a string. Common options are (but not restricted to) timegpt-1 and timegpt-1-long-horizon. Full options vary by different users. Contact support@nixtla.io for more information. We recommend using timegpt-1-long-horizon for forecasting if you want to predict more than one seasonal period given the frequency of your data.

A boolean flag that indicates whether the API should preprocess (clean) the exogenous signal before applying the large time model. If True, the exogenous signal is cleaned; if False, the exogenous variables are applied after the large time model.

The number of tuning steps used to train the large time model on the data. Set this value to 0 for zero-shot inference, i.e., to make predictions without any further model tuning.

The loss used to train the large time model on the data. Select from ['default', 'mae', 'mse', 'rmse', 'mape', 'smape', 'poisson']. It will only be used if finetune_steps larger than 0. Default is a robust loss function that is less sensitive to outliers.

The depth of the finetuning. Uses a scale from 1 to 5, where 1 means little finetuning, and 5 means that the entire model is finetuned. Note that this parameter is only effective for timegpt-1 and timegpt-1-long-horizon models, meanwhile it has no effect on the other models. By default, the value is set to 1.

ID of previously finetuned model

Step size between each cross validation window. If None it will be equal to the forecasting horizon.

Fine-tune the model in each window. If False, only fine-tunes on the first window. Only used if finetune_steps > 0.

Compute multivariate predictions across a batch of multiple time series. Requires all time series with overlapping dates. Note that this is only effective for timegpt-2.1 model and it has no effect on the other models. By default, the value is set to False.

Compute the exogenous features contributions to the forecast.

Method used to compute feature contributions. Options are: 'shapley' (default), 'intervention', 'granger', 'transfer_entropy'. The methods differ in semantics: 'shapley' returns per-timestep contributions that sum to the forecast (last row = per-timestep base prediction); 'intervention' returns each feature's counterfactual effect (forecast minus forecast with that feature held at its baseline) and does NOT sum to the forecast; 'granger'/'transfer_entropy' allocate each series' forecast deviation from its mean proportionally to model-agnostic historical importance weights — the rows sum to the forecast by construction, but they are a proportional allocation describing relationships in the data, not a per-feature attribution of this specific forecast. Use the /v2/explain endpoint for standalone historical importance weights.