<no title> — darts documentation

darts.utils.data.tabularization.add_static_covariates_to_lagged_data(features, target_series, uses_static_covariates=True, last_shape=None)[source]¶

Add static covariates to the features’ table for RegressionModels. If uses_static_covariates=True, all target series used in fit() and predict() must have static covariates with identical dimensionality. Otherwise, will not consider static covariates.

The static covariates are added to the right of the lagged features following the convention: with a 2 component series, and 2 static covariates per component -> scov_1_comp_1 | scov_1_comp_2 | scov_2_comp_1 | scov_2_comp_2

Parameters

features (Union[ndarray, Sequence[ndarray]]) – The features’ numpy array(s) to which the static covariates will be added. Can either be a lone feature matrix or a Sequence of feature matrices; in the latter case, static covariates will be appended to each feature matrix in this Sequence.
target_series (Union[TimeSeries, Sequence[TimeSeries]]) – The target series from which to read the static covariates.
uses_static_covariates (bool) – Whether the model uses/expects static covariates. If True, it enforces that static covariates must have identical shapes across all of target series.
last_shape (Optional[Tuple[int, int]]) – Optionally, the last observed shape of the static covariates. This is None before fitting, or when uses_static_covariates is False.

Returns

The features’ array(s) with appended static covariates columns. If the features input was passed as a Sequence of np.array`s, then a `Sequence is also returned; if features was passed as an np.array, a np.array is returned. last_shape is the shape of the static covariates.

Return type

(features, last_shape)

darts.utils.data.tabularization.create_lagged_component_names(target_series=None, past_covariates=None, future_covariates=None, lags=None, lags_past_covariates=None, lags_future_covariates=None, output_chunk_length=1, concatenate=True, use_static_covariates=False)[source]¶

Helper function called to retrieve the name of the features and labels arrays created with create_lagged_data(). The order of the features is the following:

Along the n_lagged_features axis, X has the following structure:: lagged_target | lagged_past_covariates | lagged_future_covariates | static covariates

For *_lags=[-2,-1] and *_series.n_components = 2 (lags shared across all the components), each lagged_* has the following structure (grouped by lags):

comp0_*_lag-2 | comp1_*_lag-2 | comp0_*_lag_-1 | comp1_*_lag-1

For *_lags={‘comp0’:[-3, -1], ‘comp1’:[-5, -3]} and *_series.n_components = 2 (component- specific lags), each lagged_* has the following structure (sorted by lags, then by components):

comp1_*_lag-5 | comp0_*_lag-3 | comp1_*_lag_-3 | comp0_*_lag-1

and for static covariates (2 static covariates acting on 2 target components):: cov0_*_target_comp0 | cov0_*_target_comp1 | cov1_*_target_comp0 | cov1_*_target_comp1

Along the n_lagged_labels axis, y has the following structure (for output_chunk_length=4 and target_series.n_components=2):

comp0_target_lag0 | comp1_target_lag0 | … | comp0_target_lag3 | comp1_target_lag3

Note : will only use the component names of the first series from target_series, past_covariates, future_covariates, and static_covariates.

The naming convention for target, past and future covariates lags is: "{name}_{type}_lag{i}", where:

{name} the component name of the (first) series

{type} is the feature type, one of “target”, “pastcov”, and “futcov”

{i} is the lag value

The naming convention for static covariates is: "{name}_statcov_target_{comp}", where:

{name} the static covariate name of the (first) series

{comp} the target component name of the (first) that the static covariate act on. If the static
covariate acts globally on a multivariate target series, will show “global”.

The naming convention for labels is: "{name}_target_hrz{i}", where:

{name} the component name of the (first) series

{i} is the step in the forecast horizon

Return type

Tuple[List[List[str]], List[List[str]]]

Returns

features_cols_name – The names of the lagged features in the X array generated by create_lagged_data() as a List[str]. If concatenate=True, also contains the columns names for the y array (on the right).
labels_cols_name –

The names of the lagged features in the y array generated by create_lagged_data()
as a List[str].

See also

tabularization.create_lagged_data: generate the lagged features and labels as (list of) Arrays.

darts.utils.data.tabularization.create_lagged_data(target_series=None, past_covariates=None, future_covariates=None, lags=None, lags_past_covariates=None, lags_future_covariates=None, output_chunk_length=1, output_chunk_shift=0, uses_static_covariates=True, last_static_covariates_shape=None, max_samples_per_ts=None, multi_models=True, check_inputs=True, use_moving_windows=True, is_training=True, concatenate=True)[source]¶

Creates the features array X and labels array y to train a lagged-variables regression model (e.g. an sklearn model) when is_training = True; alternatively, creates the features array X to produce a series of prediction from an already-trained regression model when is_training = False. In both cases, a list of time indices corresponding to each generated observation is also returned.

Notes

Instead of calling create_lagged_data directly, it is instead recommended that:

create_lagged_training_data be called if one wishes to create the X and y arrays

to train a regression model. - create_lagged_prediction_data be called if one wishes to create the X array required to generate a prediction from an already-trained regression model.

This is because even though both of these functions are merely wrappers around create_lagged_data, their call signatures are more easily interpreted than create_lagged_data. For example, create_lagged_prediction_data does not accept output_chunk_length nor multi_models as inputs, since these inputs are not used when constructing prediction data. Similarly, create_lagged_prediction_data returns only X and times as outputs, as opposed to returning y as None along with X and times.

The X array is constructed from the lagged values of up to three separate timeseries:: 1. The target_series, which contains the values we’re trying to predict. A regression model that uses previous values of the target its predicting is referred to as autoregressive; please refer to [1] for further details about autoregressive timeseries models. 2. The past covariates series, which contains values that are not known into the future. Unlike the target series, however, past covariates are not to be predicted by the regression model. 3. The future covariates (AKA ‘exogenous’ covariates) series, which contains values that are known into the future, even beyond the data in target_series and past_covariates.

See [2] for a more detailed discussion about target, past, and future covariates. Conversely, y is comprised only of the lagged values of target_series.

The shape of X is:: X.shape = (n_observations, n_lagged_features, n_samples),

where n_observations equals either the number of time points shared between all specified series, or max_samples_per_ts, whichever is smallest. The shape of y is:

y.shape = (n_observations, output_chunk_length, n_samples),

if multi_models = True, otherwise:: y.shape = (n_observations, 1, n_samples).

Along the n_lagged_features axis, X has the following structure (for *_lags=[-2,-1] and *_series.n_components = 2):

lagged_target | lagged_past_covariates | lagged_future_covariates

where each lagged_* has the following structure:: lag_-2_comp_1_* | lag_-2_comp_2_* | lag_-1_comp_1_* | lag_-1_comp_2_*

Along the n_lagged_labels axis, y has the following structure (for output_chunk_length=4 and target_series.n_components=2):

lag_+0_comp_1_target | lag_+0_comp_2_target | … | lag_+3_comp_1_target | lag_+3_comp_2_target

The lags and lags_past_covariates must contain only values less than or equal to -1. In other words, one cannot use the value of either of these series at time t to predict the value of the target series at the same time t; this is because the values of target_series and past_covariates at time t aren’t available at prediction time, by definition. Conversely, since the values of future_covariates are known into the future, lags_future_covariates can contain negative, positive, and/or zero lag values (i.e. we can use the values of future_covariates at time t or beyond to predict the value of target_series at time t).

The exact method used to construct X and y depends on whether all specified timeseries are of the same frequency or not:

If all specified timeseries are of the same frequency, strided_moving_window is used to extract

contiguous time blocks from each timeseries; the lagged variables are then extracted from each window. - If all specified timeseries are not of the same frequency, then find_shared_times is first used to find those times common to all three timeseries, after which the lagged features are extracted by offsetting the time indices of these common times by the requested lags.

In cases where it can be validly applied, the ‘moving window’ method is expected to be faster than the ‘intersecting time’ method. However, in exceptional cases where only a small number of lags are being extracted, but the difference between the lag values is large (e.g. lags = [-1, -1000]), the ‘moving window’ method is expected to consume significantly more memory, since it extracts all series values between the maximum and minimum lags as ‘windows’, before actually extracting the specific requested lag values.

In order for the lagged features of a series to be added to X, both that series and the corresponding lags must be specified; if a series is specified without the corresponding lags, that series will be ignored and not added to X. X and y arrays are constructed independently over the samples dimension (i.e. the second axis) of each series.

If the provided series are stochastic (i.e. series.n_components > 1), then an X and y array will be constructed for each sample; the arrays corresponding to each sample are concatenated togather along the 2`nd axis of `X and y. In other words, create_lagged_data is vectorised over the sample axis of the target_series, past_covariates, and future_covariates inputs. Importantly, if stochastic series are provided, each series must have the same number of samples, otherwise an error will be thrown.

Each series input (i.e. target_series, past_covariates, and future_covariates) can be specified either as a single TimeSeries, or as a Sequence of TimeSeries; the specified series must all be of the same type, however (i.e. either all TimeSeries or all Sequence[TimeSeries]). If Sequence[TimeSeries] are specified, then a feature matrix X and labels array y will be constructed using the corresponding TimeSeries in each Sequence (i.e. the first TimeSeries in each Sequence are used to create an X and y, then the second TimeSeries in each Sequence are used to create an X and y, etc.). If concatenate = True, these X’s and y’s will be concatenated along the 0`th axis; otherwise, a list of `X and y array will be returned. Note that times is always returned as a Sequence[pd.Index], however, even when concatenate = True.

Parameters

target_series (Union[TimeSeries, Sequence[TimeSeries], None]) – Optionally, the series for the regression model to predict. Must be specified if is_training = True. Can be specified as either a TimeSeries or as a Sequence[TimeSeries].
past_covariates (Union[TimeSeries, Sequence[TimeSeries], None]) – Optionally, the past covariates series that the regression model will use as inputs. Unlike the target_series, past_covariates are not to be predicted by the regression model. Can be specified as either a TimeSeries or as a Sequence[TimeSeries].
future_covariates (Union[TimeSeries, Sequence[TimeSeries], None]) – Optionally, the future covariates (i.e. exogenous covariates) series that the regression model will use as inputs. Can be specified as either a TimeSeries or as a Sequence[TimeSeries].
lags (Union[Sequence[int], Dict[str, List[int]], None]) – Optionally, the lags of the target series to be used as (autoregressive) features. If not specified, autoregressive features will not be added to X. Each lag value is assumed to be negative (e.g. lags = [-3, -1] will extract target_series values which are 3 time steps and 1 time step away from the current value). If the lags are provided as a dictionary, the lags values are specific to each component in the target series.
lags_past_covariates (Union[Sequence[int], Dict[str, List[int]], None]) – Optionally, the lags of past_covariates to be used as features. Like lags, each lag value is assumed to be less than or equal to -1. If the lags are provided as a dictionary, the lags values are specific to each component in the past covariates series.
lags_future_covariates (Union[Sequence[int], Dict[str, List[int]], None]) – Optionally, the lags of future_covariates to be used as features. Unlike lags and lags_past_covariates, lags_future_covariates values can be positive (i.e. use values after time t to predict target at time t), zero (i.e. use values at time t to predict target at time t), and/or negative (i.e. use values before time t to predict target at time t). If output_chunk_shift > 0, the lags are relative to the first time step of the shifted output chunk. If the lags are provided as a dictionary, the lags values are specific to each component in the future covariates series.
output_chunk_length (int) – Optionally, the number of time steps ahead into the future the regression model is to predict. Must best specified if is_training = True.
output_chunk_shift (int) – Optionally, the number of time steps to shift the output chunk ahead into the future.
uses_static_covariates (bool) – Whether the model uses/expects static covariates. If True, it enforces that static covariates must have identical shapes across all target series.
last_static_covariates_shape (Optional[Tuple[int, int]]) – Optionally, the last observed shape of the static covariates. This is None before fitting, or when uses_static_covariates is False.
max_samples_per_ts (Optional[int]) – Optionally, the maximum number of samples to be drawn for training/validation; only the most recent samples are kept. In theory, specifying a smaller max_samples_per_ts should reduce computation time, especially in cases where many observations could be generated.
multi_models (bool) – Optionally, specifies whether the regression model predicts multiple time steps into the future. If True, then the regression model is assumed to predict all time steps from time t to t+output_chunk_length. If False, then the regression model is assumed to predict only the time step at t+output_chunk_length. This input is ignored if is_training = False.
check_inputs (bool) – Optionally, specifies that the lags_* and series_* inputs should be checked for validity. Should be set to False if inputs have already been checked for validity (e.g. inside the __init__ of a class), otherwise should be set to True.
use_moving_windows (bool) – Optionally, specifies that the ‘moving window’ method should be used to construct X and y if all provided series are of the same frequency. If use_moving_windows = False, the ‘time intersection’ method will always be used, even when all provided series are of the same frequency. In general, setting to True results in faster tabularization at the potential cost of higher memory usage. See Notes for further details.
is_training (bool) – Optionally, specifies whether the constructed lagged data are to be used for training a regression model (i.e. is_training = True), or for generating predictions from an already-trained regression model (i.e. is_training = False). If is_training = True, target_series and output_chunk_length must be specified, the multi_models input is utilised, and a label array y is returned. Conversely, if is_training = False, then target_series and output_chunk_length do not need to be specified, the multi_models input is ignored, and the returned y value is None.
concatenate (bool) – Optionally, specifies that X and y should both be returned as single np.ndarray`s, instead of as a `Sequence[np.ndarray]. If each series input is specified as a Sequence[TimeSeries] and concatenate = False, X and y will be lists whose i`th element corresponds to the feature matrix or label array formed by the `i`th `TimeSeries in each Sequence[TimeSeries] input. Conversely, if concatenate = True when Sequence[TimeSeries] are provided, then X and y will be arrays created by concatenating all feature/label arrays formed by each TimeSeries along the 0`th axis. Note that `times is still returned as Sequence[pd.Index], even when concatenate = True.

Return type

Tuple[Union[ndarray, Sequence[ndarray]], Union[None, ndarray, Sequence[ndarray]], Sequence[Index], Optional[Tuple[int, int]]]

Returns

X – The constructed features array(s), with shape (n_observations, n_lagged_features, n_samples). If the series inputs were specified as Sequence[TimeSeries] and concatenate = False, then X is returned as a Sequence[np.array]; otherwise, X is returned as a single np.array.
y – The constructed labels array. If multi_models = True, then y is a (n_observations, output_chunk_length, n_samples)-shaped array; conversely, if multi_models = False, then y is a (n_observations, 1, n_samples)-shaped array. If the series inputs were specified as Sequence[TimeSeries] and concatenate = False, then y is returned as a Sequence[np.array]; otherwise, y is returned as a single np.array.
times – The time_index of each observation in X and y, returned as a Sequence of pd.Index`es. If the series inputs were specified as `Sequence[TimeSeries], then the i`th list element gives the times of those observations formed using the `i`th `TimeSeries object in each Sequence. Otherwise, if the series inputs were specified as TimeSeries, the only element is the times of those observations formed from the lone TimeSeries inputs.
last_static_covariates_shape – The last observed shape of the static covariates. This is None when uses_static_covariates is False.

Raises

ValueError – If the specified time series do not share any times for which features (and labels if is_training = True) can be constructed.
ValueError – If no lags are specified, or if any of the specified lag values are non-negative.
ValueError – If any of the series are too short to create features and/or labels for the requested lags and output_chunk_length values.
ValueError – If target_series and/or output_chunk_length are not specified when is_training = True.
ValueError – If the provided series do not share the same type of time_index (e.g. target_series uses a pd.RangeIndex, but future_covariates uses a pd.DatetimeIndex).

References

1: https://otexts.com/fpp2/AR.html#AR
2: https://unit8.com/resources/time-series-forecasting-using-past-and-future-external-data-with-darts/