autogluon.timeseries 1.2.1b20250224__py3-none-any.whl → 1.4.1b20251215__py3-none-any.whl

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

@@ -0,0 +1,185 @@
+from abc import ABC
+from typing import Any, Type
+
+from autogluon.timeseries.dataset import TimeSeriesDataFrame
+
+from .abstract import ArrayBasedTimeSeriesEnsembleModel
+from .regressor import (
+    EnsembleRegressor,
+    LinearStackerEnsembleRegressor,
+    MedianEnsembleRegressor,
+    PerQuantileTabularEnsembleRegressor,
+    TabularEnsembleRegressor,
+)
+
+
+class MedianEnsemble(ArrayBasedTimeSeriesEnsembleModel):
+    """Robust ensemble that computes predictions as the element-wise median of base model mean
+    and quantile forecasts, providing robustness to outlier predictions.
+
+    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 _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):
+    """Tabular ensemble that uses a single AutoGluon-Tabular model to learn ensemble combinations.
+
+    This ensemble trains a single tabular model (such as gradient boosting machines) to predict all
+    quantiles simultaneously from base model predictions. The tabular model learns complex non-linear
+    patterns in how base models should be combined, potentially capturing interactions and conditional
+    dependencies that simple weighted averages cannot represent.
+
+    Other Parameters
+    ----------------
+    model_name : str, default = "GBM"
+        Name of the AutoGluon-Tabular model to use for ensemble learning. Model name should be registered
+        in AutoGluon-Tabular model registry.
+    model_hyperparameters : dict, default = {}
+        Hyperparameters to pass to the underlying AutoGluon-Tabular model.
+    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.
+    """
+
+    ensemble_regressor_type = TabularEnsembleRegressor
+
+
+class PerQuantileTabularEnsemble(BaseTabularEnsemble):
+    """Tabular ensemble using separate AutoGluon-Tabular models for each quantile and mean forecast.
+
+    This ensemble trains dedicated tabular models for each quantile level plus a separate model
+    for the mean prediction. Each model specializes in learning optimal combinations for its
+    specific target, allowing for quantile-specific ensemble strategies that can capture different
+    model behaviors across the prediction distribution.
+
+    Other Parameters
+    ----------------
+    model_name : str, default = "GBM"
+        Name of the AutoGluon-Tabular model to use for ensemble learning. Model name should be registered
+        in AutoGluon-Tabular model registry.
+    model_hyperparameters : dict, default = {}
+        Hyperparameters to pass to the underlying AutoGluon-Tabular model.
+    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.
+    """
+
+    ensemble_regressor_type = PerQuantileTabularEnsembleRegressor
+
+
+class LinearStackerEnsemble(ArrayBasedTimeSeriesEnsembleModel):
+    """Linear stacking ensemble that learns optimal linear combination weights through gradient-based
+    optimization.
+
+    Weighted combinations can be per model or per model-quantile, model-horizon, model-quantile-horizon
+    combinations. These choices are controlled by the ``weights_per`` hyperparameter.
+
+    The optimization process uses gradient descent with configurable learning rates and convergence
+    criteria, allowing for flexible training dynamics. Weight pruning can be applied to remove
+    models with negligible contributions, resulting in sparse and interpretable ensembles.
+
+    Other Parameters
+    ----------------
+    weights_per : str, default = "m"
+        Granularity of weight learning.
+
+        - "m": single weight per model
+        - "mq": single weight for each model-quantile combination
+        - "mt": single weight for each model-time step where time steps run across the prediction horizon
+        - "mtq": single weight for each model-quantile-time step combination
+    lr : float, default = 0.1
+        Learning rate for PyTorch optimizer during weight training.
+    max_epochs : int, default = 10000
+        Maximum number of training epochs for weight optimization.
+    relative_tolerance : float, default = 1e-7
+        Relative tolerance for convergence detection during training.
+    prune_below : float, default = 0.0
+        Threshold below which weights are pruned to zero for sparsity. The weights are redistributed across
+        remaining models after pruning.
+    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 _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,
+                "prune_below": 0.0,
+            }
+        )
+        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"],
+            prune_below=hps["prune_below"],
+        )
+
+    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:
+        super()._fit(predictions_per_window, data_per_window, model_scores, time_limit)
+
+        assert isinstance(self.ensemble_regressor, LinearStackerEnsembleRegressor)
+
+        if self.ensemble_regressor.kept_indices is not None:
+            original_names = self._model_names
+            self._model_names = [original_names[i] for i in self.ensemble_regressor.kept_indices]

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 +1,186 @@
+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.
+    prune_below
+        Importance threshold for model sparsification. Models with importance below this
+        threshold are dropped after weight optimization. Set to 0.0 to disable sparsification.
+        Defaults to 0.0.
+    """
+
+    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,
+        prune_below: float = 0.0,
+    ):
+        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
+        self.prune_below = prune_below
+
+        self.weights: np.ndarray | None = None
+        self.kept_indices: list[int] | 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
+
+        with torch.no_grad():
+            self.weights = weighted_average.get_normalized_weights().detach().numpy()
+
+        assert self.weights is not None
+        if self.prune_below > 0.0:
+            importances = self.weights.mean(axis=tuple(range(self.weights.ndim - 1)))  # shape (num_models,)
+
+            mask = importances >= self.prune_below
+            if not mask.any():
+                mask[importances.argmax()] = True
+
+            if not mask.all():
+                self.kept_indices = np.where(mask)[0].tolist()
+                self.weights = self.weights[..., mask]
+                self.weights = self.weights / self.weights.sum(axis=-1, keepdims=True)
+
+        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")
+
+        all_predictions = np.concatenate([base_model_mean_predictions, base_model_quantile_predictions], axis=3)
+
+        if self.kept_indices is not None:
+            assert all_predictions.shape[-1] == len(self.kept_indices)
+
+        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

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

@@ -0,0 +1,107 @@
+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 .abstract import EnsembleRegressor
+
+logger = logging.getLogger(__name__)
+
+
+class TabularEnsembleRegressor(EnsembleRegressor):
+    """Ensemble regressor based on a single model from AutoGluon-Tabular that predicts all quantiles simultaneously."""
+
+    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.model = model_type(
+            problem_type="quantile",
+            hyperparameters=model_hyperparameters | {"ag.quantile_levels": quantile_levels},
+            path="",
+            name=model_name,
+        )
+
+    def fit(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+        labels: np.ndarray,
+        time_limit: float | None = None,
+    ) -> Self:
+        X = self._get_feature_df(base_model_mean_predictions, base_model_quantile_predictions)
+        num_windows, num_items, prediction_length = base_model_mean_predictions.shape[:3]
+        y = pd.Series(labels.reshape(num_windows * num_items * prediction_length))
+        self.model.fit(X=X, y=y, time_limit=time_limit)
+        return self
+
+    def predict(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        assert self.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 = self._get_feature_df(base_model_mean_predictions, base_model_quantile_predictions)
+
+        pred = self.model.predict(X)
+
+        # Reshape back to (num_windows, num_items, prediction_length, num_quantiles)
+        pred = pred.reshape(num_windows, num_items, prediction_length, len(self.quantile_levels))
+
+        # Use median quantile as mean prediction
+        median_idx = self._get_median_quantile_index()
+        mean_pred = pred[:, :, :, median_idx : median_idx + 1]
+        quantile_pred = pred
+
+        return mean_pred, quantile_pred
+
+    def _get_feature_df(
+        self,
+        base_model_mean_predictions: np.ndarray,
+        base_model_quantile_predictions: np.ndarray,
+    ) -> pd.DataFrame:
+        num_windows, num_items, prediction_length, _, num_models = base_model_mean_predictions.shape
+        num_tabular_items = num_windows * num_items * prediction_length
+        features_array = np.hstack(
+            [
+                base_model_mean_predictions.reshape(num_tabular_items, -1),
+                base_model_quantile_predictions.reshape(num_tabular_items, -1),
+            ]
+        )
+        return pd.DataFrame(features_array, columns=self._get_feature_names(num_models))
+
+    def _get_feature_names(self, num_models: int) -> list[str]:
+        feature_names = []
+        for mi in range(num_models):
+            feature_names.append(f"model_{mi}_mean")
+        for quantile in self.quantile_levels:
+            for mi in range(num_models):
+                feature_names.append(f"model_{mi}_q{quantile}")
+
+        return feature_names
+
+    def _get_median_quantile_index(self):
+        """Get quantile index closest to 0.5"""
+        quantile_array = np.array(self.quantile_levels)
+        median_idx = int(np.argmin(np.abs(quantile_array - 0.5)))
+        selected_quantile = quantile_array[median_idx]
+
+        if selected_quantile != 0.5:
+            logger.warning(
+                f"Selected quantile {selected_quantile} is not exactly 0.5. "
+                f"Using closest available quantile for median prediction."
+            )
+
+        return median_idx