Skip to main content

​
PanelIndex

Inversion metadata for a (n_series, T) panel.

​
Attributes

series_ids : tuple Series identifiers in panel-row order (axis 0). Sorted ascending. start_offsets : NDArray[np.integer], shape (n_series,) Column index of each series’ first non-NaN observation. lengths : NDArray[np.integer], shape (n_series,) Per-series effective (non-NaN) observation count. timestamps : tuple or None Length-T time labels for the columns, or None when the panel was built positionally.

​
PanelIndex.lengths

​
PanelIndex.series_ids

​
PanelIndex.start_offsets

​
PanelIndex.timestamps

​
long_to_panel

Pivot long-format observations into a left-NaN-padded panel.

​
Parameters

values : sequence of float, shape (n_obs,) Observation values. series_ids : sequence, shape (n_obs,) Series identifier for each observation. time_index : sequence, shape (n_obs,) Time label for each observation. Labels must be mutually comparable (e.g. all timestamps, or all ints). dtype : numpy float dtype, default float64 Output panel dtype.

​
Returns

panel : NDArray, shape (n_series, T) Left-NaN-padded panel; T is the number of distinct timestamps. index : PanelIndex Inversion metadata.

​
Raises

ValueError If the input lengths mismatch, or a series has an interior gap (a missing interior timestamp, which would produce an interior NaN).

​
panel_to_long

Flatten a panel back to long-format non-NaN rows.

​
Parameters

panel : NDArray, shape (n_series, T) index : PanelIndex Metadata from :func:long_to_panel. When index.timestamps is None, column positions 0..T-1 are used as time labels.

​
Returns

row_series_ids : list, shape (n_rows,) row_timestamps : list, shape (n_rows,) row_values : NDArray, shape (n_rows,) Non-NaN cells in (series, time) order.

​
cv_frames_to_arrays

Assemble calendar-cutoff CV cubes into per-series or aligned arrays.

​
Parameters

cubes : sequence of NDArray, each shape (n_series, n_windows, horizon) CV outputs sharing a NaN pattern (e.g. [predictions, truths]). A (series, window) cell is valid iff it is non-NaN across the whole horizon in every cube. aligned : bool True returns the cubes unchanged (NaN cells preserved for the ragged calibration path). False returns dense per-series arrays: each series keeps only its valid windows, right-aligned, truncated to the minimum valid-window count across series.

​
Returns

list of NDArray One array per input cube, in the same order.

​
Raises

ValueError If cubes is empty, shapes disagree, or (aligned=False) any series has zero valid windows.

​
exog_to_long

Flatten a 4-D exogenous array to long-format rows.

​
Parameters

exog : NDArray, shape (n_series, n_samples, horizon, n_features) series_ids : sequence, shape (n_series,)

​
Returns

row_series_ids : list, shape (n_rows,) row_sample_idx : list of int, shape (n_rows,) row_step_idx : list of int, shape (n_rows,) Zero-based horizon step. features : NDArray, shape (n_rows, n_features) Rows in (series, sample, step) order.

​
Raises

ValueError If exog is not 4-D or its leading axis mismatches series_ids.

​
level_to_quantiles

Convert a symmetric Nixtla interval level to a (lo, hi) quantile pair.

​
Parameters

level : int Interval level in (0, 100), e.g. 80.

​
Returns

tuple of float ((100 - level) / 200, 1 - (100 - level) / 200) — e.g. (0.1, 0.9).

​
Raises

ValueError If level is not in (0, 100).

​
quantiles_to_level

Convert a symmetric (lo, hi) quantile pair to a Nixtla interval level.

​
Parameters

q_lo, q_hi : float Quantiles in (0, 1) symmetric about 0.5 (q_lo + q_hi == 1).

​
Returns

int The interval level, e.g. (0.1, 0.9) -> 80.

​
Raises

ValueError If the pair is out of range or not symmetric about 0.5.