Chronos-2 Model for zero-shot forecasting. This is an implementation of Amazon's Chronos-2 model [1]_, [2]_, ported from `amazon-scie

An implementation of the NLinear model, as presented in [1]_. This implementation is improved by allowing the optional use of past covariates

LGBM Model Parameters ---------- lags Lagged target `series` values used to predict the next time step/s.

Model-specific adapter between Darts forecasting models and the SHAP library. Parameters ---------- model A fitte

Adds transformed static covariates back to original `TimeSeries`. The categorical component mapping is used to correctly name categor

Creates a validation set and returns a new set of kwargs passed to `self.model.fit()` including the validation set. This method can be overrid

Shrink the beginning of the historical forecasts time index based on the value of `retrain`, `train_length` and `val_length`.

Returns `True` if all specified (i.e. non-`None`) `series` have the same frequency.

Applies the calibrated interval to the predicted quantiles. Returns an array with `len(quantiles)` conformalized quantile predictions (lower q

Checks if the series is stochastic (number of samples is larger than one).

This function fills the missing values in the TimeSeries `series`, using the `pandas.Dataframe.interpolate()` method. Parameters ---

Computes the standard error of `r` at order `m` with respect to `length` according to Bartlett's formula. Parameters ---------- r

Minimum prediction horizon for base models to satisfy the ensemble (regression) model's future covariate requirements during predict.

Builds the feature names for the SHAP explanations based on the input features used by the forecasting model. See above for the namin

Build the mixer blocks for the model.

Computes the lower and upper calibrated forecast intervals based on residuals. Parameters ---------- residuals Th

Check that the positional parameters used to instantiate the new model loading the weights match those of the saved model, to return

Ensures that the dataset exists and its MD5 checksum matches the expected hash. Raises ------- DatasetLoadingExcepti

Throws `ValueError` if any `lag` values aren't negative OR if no lags have been specified.

By default, historical forecasts cannot be optimized

Throws `ValueError` if `series` is too short for specified `lags` and, when `is_training`, `output_chunk_length`.

Checks that the input is valid

Checks if the parameter window is less or equal than the length of the given series

Handles the classification metrics input parameters and checks.

Returns a cleaned instance of the model. Has no effect for local forecasting models.

Returns a cleaned model, keeping only the necessary attributes for prediction.

Return a `pandas.Index` with unique string component / column names

Compares pd.Timestamp instances on attributes. Compares two timestamps according two a given set of attributes (such as minute, hour, day, etc.).

Concatenate the hierarchies of multiple series, when concatenating series along axis 1 (components). This simply merges the hierarchy dictionaries

Concatenate static covariates along the component axis (rows of static covariates). Use this for stacking or concatenating time series along compo

Fills the missing values of `series` with only the value provided (default zeroes). Parameters ---------- series The TimeSer

Converts generated anomaly score from `np.ndarray` into a sequence of series. For efficiency reasons, the anomaly scores were computed in one

Returns mapping from names of untransformed categorical static covariates names and names of transformed categorical static covariate

Returns a boolean array indicating which components of the TRANSFORMED `stat_covs` are numerical and a boolean array indicating which

Extract lagged data from target, past covariates and future covariates for auto-regression with SKLearnModels.

Helper function called by `_create_lagged_data` that computes `X`, `y`, and `times` by first finding the time points in each series that *cou

Helper function called by `create_lagged_data` that computes `X`, `y`, and `times` by extracting 'moving windows' from each series using the

This method has to be implemented by all children. It is in charge of instantiating the actual torch model, based on examples input/o

Crops TimeSeries instance to contain full periods. Crops a given TimeSeries `series` that will be used as a training set in such a way that i

Test the callable with sample `pd.DatetimeIndex` and `pd.RangeIndex` to determine the number of output components.

Downloads the dataset in the root_path directory Raises ------- DatasetLoadingException if downloading o

Eager attention implementation using manual matmul. Args: query_states: [batch, n_heads, seq_len, kv_proj_dim] key_st

Computes the model output. Parameters ---------- model The Darts `SKLearnModel`. x T

Computes the model output. When this method is called, it is guaranteed that either `series` is None, or that the model supp

Extends each target `series` to the end of the last historical forecast for that series. Fills the values all missing dates with `np.nan`. As

Returns the automatically added seasonalities by Prophet's base model based on kwargs of model creation

Extract, concatenate and reorder component-wise lags to obtain a feature order identical to tabularization.

Extract a deterministic series from `series` (quantile=0.5 if `series` is probabilistic).

extracts immutable model creation parameters from `ModelMeta` and deletes reference.

Extract params from `config` to set up the given `module_class`.

Extracts all static covariates from a `TimeSeries`, and then extracts the numerical and categorical components to transform from thes

Return the time index and values with missing dates inserted. This requires either a provided `freq` or the possibility to infer a unique fre

Fits/trains the model on the provided series. DualCovariatesModels must implement the fit logic in this method.

Abstract method to implement the model and scorer training.

Fits the aggregator, assuming the input is in the correct shape. Parameters ---------- anomalies The (sequence of

CatBoost currently only supports categorical features as int. If categorical features are specified, the samples are converted into a

Original forward pass of the TimesFM 2.5 model. Parameters ---------- inputs Input tensor of shape (batch_size, n

Original forward pass of the Chronos-2 model. Parameters ---------- context Input tensor of shape (batch_size, co

Private method to implement the forward method in the subclasses.

Creates a cost matrix from a window. Depending on the density of the active cells in the window, will select either a dense o

Based on the type of the argument and the nature of the covariates, perform some sanity checks before converting the lags to a list o

Generates batch predictions. Feeds `PLForecastingModule` with past, future, and static features to forecast the next ``n`` target values

Get parameter key : default_value pairs for the estimator

Return the default SHAP method.

Returns `self._fitted_params` if `calling_fit = False`, otherwise returns an empty tuple. If `calling_fit = False`, also checks that

Computes the mode (value with the highest frequency) for all time steps and components. Parameters ---------- vals A numpy array

Computes the class label with highest probability for all time steps and components. Parameters ---------- vals A numpy array wit

Returns a list of lagged features for the target, past, future, and static covariates. The returned lagged features are in the same order as

Computes and returns: - the lagged feature indices for extraction from windows - the reordered indices to apply after the window extraction (

Computes the maximum historical forecastable time index for training or prediction mode. Only accounts for `is_training`, `forecast_horizon`, `ov

Gets the median prediction per component extracted from the model output.

Overrides `_get_params` of `BaseDataTransformer`. Creates generator of dictionaries containing both the fixed parameter values (i.e.

Performs checks common to ForecastingModel and SKLearnModel backtest() methods Parameters ---------- model The forecasting m