autogluon.timeseries 1.0.1b20240304__py3-none-any.whl → 1.4.1b20251210__py3-none-any.whl

autogluon/timeseries/models/ensemble/array_based/abstract.py

@@ -0,0 +1,236 @@
+from abc import ABC, abstractmethod
+from typing import Any, Sequence
+
+import numpy as np
+
+from autogluon.timeseries.dataset import TimeSeriesDataFrame
+from autogluon.timeseries.metrics.abstract import TimeSeriesScorer
+from autogluon.timeseries.utils.features import CovariateMetadata
+
+from ..abstract import AbstractTimeSeriesEnsembleModel
+from .regressor import EnsembleRegressor
+
+
+class ArrayBasedTimeSeriesEnsembleModel(AbstractTimeSeriesEnsembleModel, ABC):
+    """Abstract base class for time series ensemble models which operate on arrays of base model
+    predictions for training and inference.
+
+    Other Parameters
+    ----------------
+    isotonization: str, default = "sort"
+        The isotonization method to use (i.e. the algorithm to prevent quantile non-crossing).
+        Currently only "sort" is supported.
+    detect_and_ignore_failures: bool, default = True
+        Whether to detect and ignore "failed models", defined as models which have a loss that is larger
+        than 10x the median loss of all the models. This can be very important for the regression-based
+        ensembles, as moving the weight from such a "failed model" to zero can require a long training
+        time.
+    """
+
+    def __init__(
+        self,
+        path: str | None = None,
+        name: str | None = None,
+        hyperparameters: dict[str, Any] | None = None,
+        freq: str | None = None,
+        prediction_length: int = 1,
+        covariate_metadata: CovariateMetadata | None = None,
+        target: str = "target",
+        quantile_levels: Sequence[float] = (0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9),
+        eval_metric: str | TimeSeriesScorer | None = None,
+    ):
+        super().__init__(
+            path=path,
+            name=name,
+            hyperparameters=hyperparameters,
+            freq=freq,
+            prediction_length=prediction_length,
+            covariate_metadata=covariate_metadata,
+            target=target,
+            quantile_levels=quantile_levels,
+            eval_metric=eval_metric,
+        )
+        self.ensemble_regressor: EnsembleRegressor | None = None
+        self._model_names: list[str] = []
+
+    def _get_default_hyperparameters(self) -> dict[str, Any]:
+        return {
+            "isotonization": "sort",
+            "detect_and_ignore_failures": True,
+        }
+
+    @staticmethod
+    def to_array(df: TimeSeriesDataFrame) -> np.ndarray:
+        """Given a TimeSeriesDataFrame object, return a single array composing the values contained
+        in the data frame.
+
+        Parameters
+        ----------
+        df
+            TimeSeriesDataFrame to convert to an array. Must contain exactly `prediction_length`
+            values for each item. The columns of `df` can correspond to ground truth values
+            or predictions (in which case, these will be the mean or quantile forecasts).
+
+        Returns
+        -------
+        array
+            of shape (num_items, prediction_length, num_outputs).
+        """
+        assert df.index.is_monotonic_increasing
+        array = df.to_numpy()
+        num_items = df.num_items
+        shape = (
+            num_items,
+            df.shape[0] // num_items,  # timesteps per item
+            df.shape[1],  # num_outputs
+        )
+        return array.reshape(shape)
+
+    def _get_base_model_predictions(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]] | dict[str, TimeSeriesDataFrame],
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """Given a mapping from model names to a list of data frames representing
+        their predictions per window, return a multidimensional array representation.
+
+        Parameters
+        ----------
+        predictions_per_window
+            A dictionary with list[TimeSeriesDataFrame] values, where each TimeSeriesDataFrame
+            contains predictions for the window in question. If the dictionary values are
+            TimeSeriesDataFrame, they will be treated like a single window.
+
+        Returns
+        -------
+        base_model_mean_predictions
+            Array of shape (num_windows, num_items, prediction_length, 1, num_models)
+        base_model_quantile_predictions
+            Array of shape (num_windows, num_items, prediction_length, num_quantiles, num_models)
+        """
+
+        if not predictions_per_window:
+            raise ValueError("No base model predictions are provided.")
+
+        first_prediction = list(predictions_per_window.values())[0]
+        if isinstance(first_prediction, TimeSeriesDataFrame):
+            predictions_per_window = {k: [v] for k, v in predictions_per_window.items()}  # type: ignore
+
+        predictions = {
+            model_name: [self.to_array(window) for window in windows]  # type: ignore
+            for model_name, windows in predictions_per_window.items()
+        }
+        base_model_predictions = np.stack([x for x in predictions.values()], axis=-1)
+
+        return base_model_predictions[:, :, :, :1, :], base_model_predictions[:, :, :, 1:, :]
+
+    def _isotonize(self, prediction_array: np.ndarray) -> np.ndarray:
+        """Apply isotonization to ensure quantile non-crossing.
+
+        Parameters
+        ----------
+        prediction_array
+            Array of shape (num_windows, num_items, prediction_length, num_quantiles)
+
+        Returns
+        -------
+        isotonized_array
+            Array with same shape but quantiles sorted along last dimension
+        """
+        isotonization = self.get_hyperparameter("isotonization")
+        if isotonization == "sort":
+            return np.sort(prediction_array, axis=-1)
+        return prediction_array
+
+    def _fit(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        data_per_window: list[TimeSeriesDataFrame],
+        model_scores: dict[str, float] | None = None,
+        time_limit: float | None = None,
+    ) -> None:
+        # process inputs
+        filtered_predictions = self._filter_failed_models(predictions_per_window, model_scores)
+        base_model_mean_predictions, base_model_quantile_predictions = self._get_base_model_predictions(
+            filtered_predictions
+        )
+
+        # process labels
+        ground_truth_per_window = [y.slice_by_timestep(-self.prediction_length, None) for y in data_per_window]
+        labels = np.stack(
+            [self.to_array(gt) for gt in ground_truth_per_window], axis=0
+        )  # (num_windows, num_items, prediction_length, 1)
+
+        self._model_names = list(filtered_predictions.keys())
+        self.ensemble_regressor = self._get_ensemble_regressor()
+        self.ensemble_regressor.fit(
+            base_model_mean_predictions=base_model_mean_predictions,
+            base_model_quantile_predictions=base_model_quantile_predictions,
+            labels=labels,
+            time_limit=time_limit,
+        )
+
+    @abstractmethod
+    def _get_ensemble_regressor(self) -> EnsembleRegressor:
+        pass
+
+    def _predict(self, data: dict[str, TimeSeriesDataFrame], **kwargs) -> TimeSeriesDataFrame:
+        if self.ensemble_regressor is None:
+            if not self._model_names:
+                raise ValueError("Ensemble model has not been fitted yet.")
+            # Try to recreate the regressor (for loaded models)
+            self.ensemble_regressor = self._get_ensemble_regressor()
+
+        input_data = {}
+        for m in self.model_names:
+            assert m in data, f"Predictions for model {m} not provided during ensemble prediction."
+            input_data[m] = data[m]
+
+        base_model_mean_predictions, base_model_quantile_predictions = self._get_base_model_predictions(input_data)
+
+        mean_predictions, quantile_predictions = self.ensemble_regressor.predict(
+            base_model_mean_predictions=base_model_mean_predictions,
+            base_model_quantile_predictions=base_model_quantile_predictions,
+        )
+
+        quantile_predictions = self._isotonize(quantile_predictions)
+        prediction_array = np.concatenate([mean_predictions, quantile_predictions], axis=-1)
+
+        output = list(input_data.values())[0].copy()
+        num_folds, num_items, num_timesteps, num_outputs = prediction_array.shape
+        assert (num_folds, num_timesteps) == (1, self.prediction_length)
+        assert len(output.columns) == num_outputs
+
+        output[output.columns] = prediction_array.reshape((num_items * num_timesteps, num_outputs))
+
+        return output
+
+    @property
+    def model_names(self) -> list[str]:
+        return self._model_names
+
+    def remap_base_models(self, model_refit_map: dict[str, str]) -> None:
+        """Update names of the base models based on the mapping in model_refit_map."""
+        self._model_names = [model_refit_map.get(name, name) for name in self._model_names]
+
+    def _filter_failed_models(
+        self,
+        predictions_per_window: dict[str, list[TimeSeriesDataFrame]],
+        model_scores: dict[str, float] | None,
+    ) -> dict[str, list[TimeSeriesDataFrame]]:
+        """Filter out failed models based on detect_and_ignore_failures setting."""
+        if not self.get_hyperparameter("detect_and_ignore_failures"):
+            return predictions_per_window
+
+        if model_scores is None or len(model_scores) == 0:
+            return predictions_per_window
+
+        valid_scores = {k: v for k, v in model_scores.items() if np.isfinite(v)}
+        if len(valid_scores) == 0:
+            raise ValueError("All models have NaN scores. At least one model must run successfully to fit an ensemble")
+
+        losses = {k: -v for k, v in valid_scores.items()}
+        median_loss = np.nanmedian(list(losses.values()))
+        threshold = 10 * median_loss
+        good_models = {k for k, loss in losses.items() if loss <= threshold}
+
+        return {k: v for k, v in predictions_per_window.items() if k in good_models}

autogluon/timeseries/models/ensemble/array_based/models.py

@@ -0,0 +1,73 @@
+from abc import ABC
+from typing import Any, Type
+
+from .abstract import ArrayBasedTimeSeriesEnsembleModel
+from .regressor import (
+    EnsembleRegressor,
+    LinearStackerEnsembleRegressor,
+    MedianEnsembleRegressor,
+    PerQuantileTabularEnsembleRegressor,
+    TabularEnsembleRegressor,
+)
+
+
+class MedianEnsemble(ArrayBasedTimeSeriesEnsembleModel):
+    def _get_ensemble_regressor(self) -> MedianEnsembleRegressor:
+        return MedianEnsembleRegressor()
+
+
+class BaseTabularEnsemble(ArrayBasedTimeSeriesEnsembleModel, ABC):
+    ensemble_regressor_type: Type[EnsembleRegressor]
+
+    def _get_default_hyperparameters(self) -> dict[str, Any]:
+        default_hps = super()._get_default_hyperparameters()
+        default_hps.update({"model_name": "GBM", "model_hyperparameters": {}})
+        return default_hps
+
+    def _get_ensemble_regressor(self):
+        hyperparameters = self.get_hyperparameters()
+        return self.ensemble_regressor_type(
+            quantile_levels=list(self.quantile_levels),
+            model_name=hyperparameters["model_name"],
+            model_hyperparameters=hyperparameters["model_hyperparameters"],
+        )
+
+
+class TabularEnsemble(BaseTabularEnsemble):
+    """Time series ensemble model using a single AutoGluon-Tabular model for all quantiles."""
+
+    ensemble_regressor_type = TabularEnsembleRegressor
+
+
+class PerQuantileTabularEnsemble(BaseTabularEnsemble):
+    """Time series ensemble model using separate AutoGluon-Tabular models for each quantile in
+    addition to a dedicated model for the mean (point) forecast.
+    """
+
+    ensemble_regressor_type = PerQuantileTabularEnsembleRegressor
+
+
+class LinearStackerEnsemble(ArrayBasedTimeSeriesEnsembleModel):
+    """Time series ensemble model using linear stacker with PyTorch optimization."""
+
+    def _get_default_hyperparameters(self) -> dict[str, Any]:
+        default_hps = super()._get_default_hyperparameters()
+        default_hps.update(
+            {
+                "weights_per": "m",
+                "lr": 0.1,
+                "max_epochs": 10000,
+                "relative_tolerance": 1e-7,
+            }
+        )
+        return default_hps
+
+    def _get_ensemble_regressor(self) -> LinearStackerEnsembleRegressor:
+        hps = self.get_hyperparameters()
+        return LinearStackerEnsembleRegressor(
+            quantile_levels=list(self.quantile_levels),
+            weights_per=hps["weights_per"],
+            lr=hps["lr"],
+            max_epochs=hps["max_epochs"],
+            relative_tolerance=hps["relative_tolerance"],
+        )

autogluon/timeseries/models/ensemble/array_based/regressor/__init__.py

@@ -0,0 +1,12 @@
+from .abstract import EnsembleRegressor, MedianEnsembleRegressor
+from .linear_stacker import LinearStackerEnsembleRegressor
+from .per_quantile_tabular import PerQuantileTabularEnsembleRegressor
+from .tabular import TabularEnsembleRegressor
+
+__all__ = [
+    "EnsembleRegressor",
+    "LinearStackerEnsembleRegressor",
+    "MedianEnsembleRegressor",
+    "PerQuantileTabularEnsembleRegressor",
+    "TabularEnsembleRegressor",
+]

autogluon/timeseries/models/ensemble/array_based/regressor/abstract.py

@@ -0,0 +1,88 @@
+from abc import ABC, abstractmethod
+
+import numpy as np
+from typing_extensions import Self
+
+
+class EnsembleRegressor(ABC):
+    def __init__(self, *args, **kwargs):
+        pass
+
+    @abstractmethod
+    def fit(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+        labels: np.ndarray,
+        time_limit: float | None = None,
+    ) -> Self:
+        """
+        Parameters
+        ----------
+        base_model_mean_predictions
+            Mean (point) predictions of base models. Array of shape
+            (num_windows, num_items, prediction_length, 1, num_models)
+
+        base_model_quantile_predictions
+            Quantile predictions of base models. Array of shape
+            (num_windows, num_items, prediction_length, num_quantiles, num_models)
+
+        labels
+            Ground truth array of shape
+            (num_windows, num_items, prediction_length, 1)
+
+        time_limit
+            Approximately how long `fit` will run (wall-clock time in seconds). If
+            not specified, training time will not be limited.
+        """
+        pass
+
+    @abstractmethod
+    def predict(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """Predict with the fitted ensemble regressor for a single window.
+        The items do not have to refer to the same item indices used when fitting
+        the model.
+
+        Parameters
+        ----------
+        base_model_mean_predictions
+            Mean (point) predictions of base models. Array of shape
+            (1, num_items, prediction_length, 1, num_models)
+
+        base_model_quantile_predictions
+            Quantile predictions of base models. Array of shape
+            (1, num_items, prediction_length, num_quantiles, num_models)
+
+        Returns
+        -------
+        ensemble_mean_predictions
+            Array of shape (1, num_items, prediction_length, 1)
+        ensemble_quantile_predictions
+            Array of shape (1, num_items, prediction_length, num_quantiles)
+        """
+        pass
+
+
+class MedianEnsembleRegressor(EnsembleRegressor):
+    def fit(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+        labels: np.ndarray,
+        time_limit: float | None = None,
+    ) -> Self:
+        return self
+
+    def predict(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        return (
+            np.nanmedian(base_model_mean_predictions, axis=-1),
+            np.nanmedian(base_model_quantile_predictions, axis=-1),
+        )

autogluon/timeseries/models/ensemble/array_based/regressor/linear_stacker.py

@@ -0,0 +1,167 @@
+from typing import Literal
+
+import numpy as np
+from typing_extensions import Self
+
+from autogluon.timeseries.utils.timer import Timer
+
+from .abstract import EnsembleRegressor
+
+
+class LinearStackerEnsembleRegressor(EnsembleRegressor):
+    """Linear stacker ensemble regressor using PyTorch optimization with softmax weights.
+
+    Implements weighted averaging of base model predictions with learnable weights optimized
+    via gradient descent. Uses PyTorch during training for optimization, then stores weights
+    as numpy arrays for efficient prediction.
+
+    Parameters
+    ----------
+    quantile_levels
+        List of quantile levels for quantile predictions (e.g., [0.1, 0.5, 0.9]).
+    weights_per
+        Weight configuration specifying which dimensions to learn weights for:
+        - "m": Per-model weights (shape: num_models), defaults to "m"
+        - "mt": Per-model and per-time weights (shape: prediction_length, num_models)
+        - "mq": Per-model and per-model-output (quantiles and mean) weights
+          (shape: num_quantiles+1, num_models)
+        - "mtq": Per-model, per-time, and per-quantile weights
+          (shape: prediction_length, num_quantiles+1, num_models)
+    lr
+        Learning rate for Adam optimizer. Defaults to 0.1.
+    max_epochs
+        Maximum number of training epochs. Defaults to 10000.
+    relative_tolerance
+        Convergence tolerance for relative loss change between epochs. Defaults to 1e-7.
+    """
+
+    def __init__(
+        self,
+        quantile_levels: list[float],
+        weights_per: Literal["m", "mt", "mq", "mtq"] = "m",
+        lr: float = 0.1,
+        max_epochs: int = 10_000,
+        relative_tolerance: float = 1e-7,
+    ):
+        super().__init__()
+        self.quantile_levels = quantile_levels
+        self.weights_per = weights_per
+        self.lr = lr
+        self.max_epochs = max_epochs
+        self.relative_tolerance = relative_tolerance
+
+        # Learned weights (stored as numpy arrays)
+        self.weights: np.ndarray | None = None
+
+    def _compute_weight_shape(self, base_model_predictions_shape: tuple) -> tuple:
+        """Compute weight tensor shape based on weights_per configuration."""
+        _, _, prediction_length, num_outputs, num_models = base_model_predictions_shape
+
+        shapes = {
+            "m": (1, 1, num_models),
+            "mt": (prediction_length, 1, num_models),
+            "mq": (1, num_outputs, num_models),
+            "mtq": (prediction_length, num_outputs, num_models),
+        }
+        try:
+            return (1, 1) + shapes[self.weights_per]
+        except KeyError:
+            raise ValueError(f"Unsupported weights_per: {self.weights_per}")
+
+    def make_weighted_average_module(self, base_model_predictions_shape: tuple):
+        import torch
+
+        class WeightedAverage(torch.nn.Module):
+            def __init__(self, shape):
+                super().__init__()
+                self.raw_weights = torch.nn.Parameter(torch.zeros(*shape, dtype=torch.float32))
+
+            def get_normalized_weights(self):
+                return torch.softmax(self.raw_weights, dim=-1)  # softmax over models
+
+            def forward(self, base_model_predictions: torch.Tensor):
+                return torch.sum(self.get_normalized_weights() * base_model_predictions, dim=-1)
+
+        return WeightedAverage(self._compute_weight_shape(base_model_predictions_shape))
+
+    def fit(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+        labels: np.ndarray,
+        time_limit: float | None = None,
+    ) -> Self:
+        import torch
+
+        def _ql(
+            labels_tensor: torch.Tensor,
+            ensemble_predictions: torch.Tensor,
+        ) -> torch.Tensor:
+            """Compute the weighted quantile loss on predictions and ground truth (labels).
+            Considering that the first dimension of predictions is the mean, we treat
+            mean predictions on the same footing as median (0.5) predictions as contribution
+            to the overall weighted quantile loss.
+            """
+            quantile_levels = torch.tensor([0.5] + self.quantile_levels, dtype=torch.float32)
+            error = labels_tensor - ensemble_predictions  # (num_windows, num_items, num_time, num_outputs)
+            quantile_loss = torch.maximum(quantile_levels * error, (quantile_levels - 1) * error)
+            return torch.mean(quantile_loss)
+
+        timer = Timer(time_limit).start()
+
+        base_model_predictions = torch.tensor(
+            np.concatenate(
+                [base_model_mean_predictions, base_model_quantile_predictions],
+                axis=3,
+            ),
+            dtype=torch.float32,
+        )
+        labels_tensor = torch.tensor(labels, dtype=torch.float32)
+
+        weighted_average = self.make_weighted_average_module(base_model_predictions.shape)
+
+        optimizer = torch.optim.Adam(weighted_average.parameters(), lr=self.lr)
+
+        prev_loss = float("inf")
+        for _ in range(self.max_epochs):
+            optimizer.zero_grad()
+
+            ensemble_predictions = weighted_average(base_model_predictions)
+
+            loss = _ql(labels_tensor, ensemble_predictions)
+            loss.backward()
+            optimizer.step()
+
+            loss_change = abs(prev_loss - loss.item()) / (loss.item() + 1e-8)
+            if loss_change < self.relative_tolerance:
+                break
+            prev_loss = loss.item()
+
+            if timer.timed_out():
+                break
+
+        # store final weights as numpy array
+        # TODO: add sparsification to ensure negligible weights are dropped
+        with torch.no_grad():
+            self.weights = weighted_average.get_normalized_weights().detach().numpy()
+
+        return self
+
+    def predict(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        if self.weights is None:
+            raise ValueError("Model must be fitted before prediction")
+
+        # combine base model predictions
+        all_predictions = np.concatenate([base_model_mean_predictions, base_model_quantile_predictions], axis=3)
+
+        # predict
+        ensemble_pred = np.sum(self.weights * all_predictions, axis=-1)
+
+        mean_predictions = ensemble_pred[:, :, :, :1]
+        quantile_predictions = ensemble_pred[:, :, :, 1:]
+
+        return mean_predictions, quantile_predictions

autogluon/timeseries/models/ensemble/array_based/regressor/per_quantile_tabular.py

@@ -0,0 +1,94 @@
+import logging
+
+import numpy as np
+import pandas as pd
+from typing_extensions import Self
+
+from autogluon.tabular.registry import ag_model_registry as tabular_ag_model_registry
+from autogluon.timeseries.utils.timer import SplitTimer
+
+from .abstract import EnsembleRegressor
+
+logger = logging.getLogger(__name__)
+
+
+class PerQuantileTabularEnsembleRegressor(EnsembleRegressor):
+    """Ensemble regressor using separate models per quantile plus dedicated mean model."""
+
+    def __init__(
+        self,
+        quantile_levels: list[float],
+        model_name: str,
+        model_hyperparameters: dict | None = None,
+    ):
+        super().__init__()
+        self.quantile_levels = quantile_levels
+        model_type = tabular_ag_model_registry.key_to_cls(model_name)
+        model_hyperparameters = model_hyperparameters or {}
+        self.mean_model = model_type(
+            problem_type="regression",
+            hyperparameters=model_hyperparameters,
+            path="",
+            name=f"{model_name}_mean",
+        )
+        self.quantile_models = [
+            model_type(
+                problem_type="quantile",
+                hyperparameters=model_hyperparameters | {"ag.quantile_levels": [quantile]},
+                path="",
+                name=f"{model_name}_q{quantile}",
+            )
+            for quantile in quantile_levels
+        ]
+
+    def fit(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+        labels: np.ndarray,
+        time_limit: float | None = None,
+    ) -> Self:
+        num_windows, num_items, prediction_length = base_model_mean_predictions.shape[:3]
+        y = pd.Series(labels.reshape(num_windows * num_items * prediction_length))
+
+        total_rounds = 1 + len(self.quantile_levels)
+        timer = SplitTimer(time_limit, rounds=total_rounds).start()
+
+        # Fit mean model
+        X_mean = self._get_feature_df(base_model_mean_predictions, 0)
+        self.mean_model.fit(X=X_mean, y=y, time_limit=timer.round_time_remaining())
+        timer.next_round()
+
+        # Fit quantile models
+        for i, model in enumerate(self.quantile_models):
+            X_q = self._get_feature_df(base_model_quantile_predictions, i)
+            model.fit(X=X_q, y=y, time_limit=timer.round_time_remaining())
+            timer.next_round()
+
+        return self
+
+    def _get_feature_df(self, predictions: np.ndarray, index: int) -> pd.DataFrame:
+        num_windows, num_items, prediction_length, _, num_models = predictions.shape
+        num_tabular_items = num_windows * num_items * prediction_length
+        return pd.DataFrame(
+            predictions[:, :, :, index].reshape(num_tabular_items, num_models),
+            columns=[f"model_{mi}" for mi in range(num_models)],
+        )
+
+    def predict(
+        self, base_model_mean_predictions: np.ndarray, base_model_quantile_predictions: np.ndarray
+    ) -> tuple[np.ndarray, np.ndarray]:
+        assert self.mean_model.is_fit()
+        num_windows, num_items, prediction_length = base_model_mean_predictions.shape[:3]
+        assert num_windows == 1, "Prediction expects a single window to be provided"
+
+        X_mean = self._get_feature_df(base_model_mean_predictions, 0)
+        mean_predictions = self.mean_model.predict(X_mean).reshape(num_windows, num_items, prediction_length, 1)
+
+        quantile_predictions_list = []
+        for i, model in enumerate(self.quantile_models):
+            X_q = self._get_feature_df(base_model_quantile_predictions, i)
+            quantile_predictions_list.append(model.predict(X_q).reshape(num_windows, num_items, prediction_length))
+        quantile_predictions = np.stack(quantile_predictions_list, axis=-1)
+
+        return mean_predictions, quantile_predictions