Profiles

What profile() finds in the data. ForecastingProfile is the object you pass to plan(), create_cv() and ask(): it holds summary statistics of the dataset (a DataProfile), the significant lags of each series and the recommended forecaster and estimator, never the observations themselves.

skforecast_ai.schemas.profiles

Classes:

Name	Description
SeriesLengthInfo	Per-series index range and observation count.
DataProfile	Profile of the input time series dataset.
SeriesPacf	PACF-significant lags for a single series.
ForecastingProfile	High-level profile of the forecasting problem.

Classes

SeriesLengthInfo

Bases: BaseModel

Per-series index range and observation count.

Attributes:

Name	Type	Description
start	str, default None	First timestamp of the series as a string (e.g. '2020-01-01'). None when the index is not datetime.
end	str, default None	Last timestamp of the series as a string. None when the index is not datetime.
length	int	Number of observations in the series.

Attributes

start class-attribute instance-attribute

start = None

end class-attribute instance-attribute

end = None

length instance-attribute

length

DataProfile

Bases: PickleDefaultsMixin, BaseModel

Profile of the input time series dataset.

Attributes:

Name	Type	Description
data_format	str, default 'single'	Layout of the dataset. One of 'single', 'wide', 'long'.
n_series	int	Number of individual time series.
series_lengths	dict	Mapping of series name to its SeriesLengthInfo (start, end, and length). Always populated, including single series (keyed by the target name). The task-aware effective number of observations is derived from this mapping (span for multi_series, common length for multivariate, single length otherwise).
span_index_length	int	Length of the union datetime index spanning every series, from the earliest start to the latest end at frequency. Falls back to the longest individual series when datetime bounds or frequency are unavailable. Computed automatically from series_lengths.
n_total_observations	int	Pooled total number of observations across all series (the sum of every series length). Computed automatically from series_lengths.
n_observations_display	int	Task-agnostic observation count for display and summaries: the series length for a single series, span_index_length otherwise.
time_zone	str, default None	Name of the time zone of the dates ('Europe/Madrid', 'UTC'), or None when they have none or it cannot be rebuilt from its name. start_date is then written as local time, without its UTC offset, and the positions and the dates of a cross-validation strategy are counted on the local times of the data, which skip or repeat an hour at a daylight saving change.
span_start_date	(str, None)	First date of the span of span_index_length: start_date, or in long format the earliest first date of the series.
target	(str, list)	Name(s) of the target column(s). A single string for single series and long format. A list of strings for wide format where each element is a series column.
target_dtype	str, default 'numeric'	Data type category of the target column. One of 'numeric', 'categorical', 'other'.
target_stats	dict	Mapping of target column (or series name) to a dict with keys 'min', 'max', 'mean', 'std' computed on non-NaN values. Empty dict if no valid observations exist.
missing_target	dict	Mapping of target column (or series name) to count of NaN values. Only entries with at least one missing value are included.
date_column	str, default None	Name of the column containing timestamps.
series_id_column	str, default None	Name of the column identifying individual series.
index_type	str	Type of the DataFrame index. One of 'datetime', 'range', 'other'.
frequency	str, default None	Inferred pandas frequency string (e.g. 'h', 'D', 'ME'). In long format, the frequency shared by every series.
frequency_is_set	bool, default False	Whether the index already has a frequency set (index.freq). False when the rows were not in date order.
index_is_monotonic	bool, default True	Whether the input was in ascending date order (within each series, for long format). Rows out of order are sorted before profiling, with a note in warnings, so the other fields describe the sorted data.
has_gaps	bool, default False	Whether the datetime index has missing timestamps within its range (in long format, whether any series has them within its own range).
has_duplicate_timestamps	bool, default False	Whether some timestamps appear in several identical rows, which the generated code drops. Timestamps repeated with different values raise a ValueError during profiling.
exog_columns	list	Names of exogenous predictor columns.
categorical_exog	list	Subset of exog_columns that are categorical.
missing_exog	dict	Mapping of exogenous column name to count of missing values. Only columns with at least one missing value are included.
unused_columns	list, default []	Columns of the data that are neither the target, the date, the series id nor an exogenous variable: left out with exog_columns of profile(), or columns of data passed with a saved profile that does not name them. The generated script does not read them.
data_path	str, default 'data.csv'	Path to the source CSV file. Derived automatically during profiling: if the input is a file path, this stores it; if the input is a DataFrame, defaults to 'data.csv'.
csv_sep	str, default ','	Separator of the fields of the CSV file of data_path: ',', ';' or a tabulator. Detected when the file is profiled, or given with sep; the generated script reads the file with it. ',' for data passed in memory.
csv_decimal	str, default '.'	Decimal mark of the CSV file of data_path: '.' or ',', the latter only when csv_sep is not ','. Detected or given as csv_sep.
warnings	list	Human-readable warnings generated during profiling.

Attributes

data_format class-attribute instance-attribute

data_format = 'single'

n_series instance-attribute

n_series

series_lengths instance-attribute

series_lengths

span_index_length class-attribute instance-attribute

span_index_length = 0

n_total_observations class-attribute instance-attribute

n_total_observations = 0

target instance-attribute

target

target_dtype class-attribute instance-attribute

target_dtype = 'numeric'

target_stats class-attribute instance-attribute

target_stats = Field(default_factory=dict)

missing_target class-attribute instance-attribute

missing_target = Field(default_factory=dict)

date_column class-attribute instance-attribute

date_column = None

series_id_column class-attribute instance-attribute

series_id_column = None

index_type instance-attribute

index_type

frequency class-attribute instance-attribute

frequency = None

frequency_is_set class-attribute instance-attribute

frequency_is_set = False

index_is_monotonic class-attribute instance-attribute

index_is_monotonic = True

has_gaps class-attribute instance-attribute

has_gaps = False

has_duplicate_timestamps class-attribute instance-attribute

has_duplicate_timestamps = False

exog_columns class-attribute instance-attribute

exog_columns = Field(default_factory=list)

categorical_exog class-attribute instance-attribute

categorical_exog = Field(default_factory=list)

missing_exog class-attribute instance-attribute

missing_exog = Field(default_factory=dict)

unused_columns class-attribute instance-attribute

unused_columns = Field(default_factory=list)

data_path class-attribute instance-attribute

data_path = 'data.csv'

csv_sep class-attribute instance-attribute

csv_sep = ','

csv_decimal class-attribute instance-attribute

csv_decimal = '.'

start_date class-attribute instance-attribute

start_date = None

time_zone class-attribute instance-attribute

time_zone = None

warnings class-attribute instance-attribute

warnings = Field(default_factory=list)

span_start_date property

span_start_date

First date of the span of the data, where span_index_length starts.

It is start_date, except in long format, where start_date is the latest first date of the series and the span starts at the earliest one (the union index of the series). The earliest is taken only when span_index_length observations at frequency run from it to the last date of the series; otherwise (a span counted as the longest series, dates that mix time zones) it is start_date, as before.

Returns:

Name	Type	Description
span_start_date	(str, None)	First date of the span, or None without dates.

n_observations_display property

n_observations_display

Task-agnostic observation count for display and summaries.

Returns the series length for single-series data and the union span (span_index_length) for multi-series data.

Returns:

Name	Type	Description
n_observations	int	Representative observation count for display.

SeriesPacf

Bases: BaseModel

PACF-significant lags for a single series.

Attributes:

Name	Type	Description
series_id	str	Name of the series (target column for single/wide, series id for long format).
n_observations	int	Number of non-NaN observations in the series (all NaNs, edge and interior, are excluded by the count). Used as the sample size for the PACF significance test. Not the raw column length.
lags	list of int	Significant lags retained by Benjamini-Hochberg FDR correction and the minimum effect-size floor, ordered by descending |PACF| (importance order, not ascending index).
pacf_abs	list of float	Absolute PACF magnitude aligned element-wise with lags (same order).

Attributes

series_id instance-attribute

series_id

n_observations instance-attribute

n_observations

lags class-attribute instance-attribute

lags = Field(default_factory=list)

pacf_abs class-attribute instance-attribute

pacf_abs = Field(default_factory=list)

ForecastingProfile

Bases: PickleDefaultsMixin, DisplayMixin, ExplainableResult, BaseModel

High-level profile of the forecasting problem.

Combines the dataset profile with the coarse modeling decisions: which forecaster family to use, which estimator to pair with it, and the alternative candidates the user could switch to. Detailed configuration (lags, metric, intervals, NaN handling, preprocessing) is left to ForecastPlan.

Attributes:

Name	Type	Description
data_profile	DataProfile	Profile of the input dataset (independent of the forecasting decisions).
task_type	str	Forecasting task category implied by the selected forecaster. One of 'single_series', 'multi_series', 'multivariate', 'statistical', 'foundation'.
forecaster	str	Selected skforecast forecaster class name.
forecaster_candidates	list	Ordered list of compatible forecaster class names. The first item is the preferred default.
estimator	str, default None	Selected estimator: a scikit-learn compatible estimator name, 'Arima' for statistical tasks, or the Hugging Face model ID of the foundation model for foundation tasks. None for the baseline, which has no estimator.
estimator_candidates	list	Ordered list of compatible estimator names. Empty for the baseline.
series_pacf	list of SeriesPacf	Per-series PACF-significant lags (the forecaster-invariant lag primitive). Empty for statistical and foundation tasks. The final lag set is derived in plan() by aggregating these primitives for the chosen forecaster.
window_features	list of dict, default None	Window feature configurations (dicts with keys 'stats' and 'window_size'). Computed eagerly at profile time as they are forecaster-invariant. None when the series is too short or the task is statistical/foundation.
calendar_features	list of str, default None	Recommended calendar feature names (a subset of those supported by skforecast.preprocessing.CalendarFeatures). Computed eagerly at profile time as they depend only on the index frequency and series length. None when the frequency is unknown, the series is too short, the frequency has no sub-period seasonality, or the task is statistical/foundation. The encoding is chosen later in plan() based on the resolved estimator.
explanation	str	Human-readable explanation of why this forecaster + estimator combination was chosen.

Attributes

data_profile instance-attribute

data_profile

task_type instance-attribute

task_type

forecaster instance-attribute

forecaster

forecaster_candidates class-attribute instance-attribute

forecaster_candidates = Field(default_factory=list)

estimator class-attribute instance-attribute

estimator = None

estimator_candidates class-attribute instance-attribute

estimator_candidates = Field(default_factory=list)

series_pacf class-attribute instance-attribute

series_pacf = Field(default_factory=list)

window_features class-attribute instance-attribute

window_features = None

calendar_features class-attribute instance-attribute

calendar_features = None

explanation instance-attribute

explanation

Functions: