Implementation:Sktime Pytorch forecasting QuantileLoss

Knowledge Sources	pytorch-forecasting PyTorch Forecasting Docs
Domains	Time_Series, Loss_Functions, Probabilistic_Forecasting
Last Updated	2026-02-08 07:00 GMT

Overview

Concrete tool for computing multi-quantile pinball loss for probabilistic time series forecasting provided by the pytorch-forecasting library.

Description

The QuantileLoss class extends MultiHorizonMetric to compute pinball loss across multiple quantiles simultaneously. It computes the asymmetric loss for each quantile, multiplies by 2 (for MAE normalization), and concatenates results. It provides to_prediction() (extracts the median quantile as point forecast) and to_quantiles() (returns all quantile predictions). Default quantiles are [0.02, 0.1, 0.25, 0.5, 0.75, 0.9, 0.98].

Usage

Use as the loss function for TemporalFusionTransformer. It is the default loss when calling TemporalFusionTransformer.from_dataset(). The model's output_size should match the number of quantiles.

Code Reference

Source Location

• Repository: pytorch-forecasting
• File: pytorch_forecasting/metrics/quantile.py
• Lines: L10-68

Signature

class QuantileLoss(MultiHorizonMetric):
    def __init__(
        self,
        quantiles: list[float] | None = None,
        **kwargs,
    ):
        """
        Quantile loss.

        Args:
            quantiles: quantiles for metric.
                Default: [0.02, 0.1, 0.25, 0.5, 0.75, 0.9, 0.98]
        """

    def loss(self, y_pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
        """Compute quantile loss per quantile."""

    def to_prediction(self, y_pred: torch.Tensor) -> torch.Tensor:
        """Extract median (q=0.5) as point prediction."""

    def to_quantiles(self, y_pred: torch.Tensor) -> torch.Tensor:
        """Return all quantile predictions."""

Import

from pytorch_forecasting.metrics import QuantileLoss

I/O Contract

Inputs

Name	Type	Required	Description
quantiles	list[float]	No	Quantile levels (default: [0.02, 0.1, 0.25, 0.5, 0.75, 0.9, 0.98])
y_pred	torch.Tensor	Yes (to loss)	Model predictions (batch, horizon, n_quantiles)
target	torch.Tensor	Yes (to loss)	Actual target values (batch, horizon)

Outputs

Name	Type	Description
loss()	torch.Tensor	Per-sample quantile losses (batch, horizon, n_quantiles)
to_prediction()	torch.Tensor	Point forecast — median quantile (batch, horizon)
to_quantiles()	torch.Tensor	All quantile predictions (batch, horizon, n_quantiles)

Usage Examples

Default QuantileLoss for TFT

from pytorch_forecasting.metrics import QuantileLoss

# Default: 7 quantiles
loss = QuantileLoss()
print(f"Quantiles: {loss.quantiles}")
# [0.02, 0.1, 0.25, 0.5, 0.75, 0.9, 0.98]

# Use with TFT
tft = TemporalFusionTransformer.from_dataset(
    training,
    loss=loss,
    output_size=7,  # matches number of quantiles
)

Custom Quantiles

# Narrower prediction interval
loss = QuantileLoss(quantiles=[0.1, 0.5, 0.9])

tft = TemporalFusionTransformer.from_dataset(
    training,
    loss=loss,
    output_size=3,
)

Related Pages

Implements Principle

• Principle:Sktime_Pytorch_forecasting_Quantile_Loss

Requires Environment

• Environment:Sktime_Pytorch_forecasting_Core_Python_Dependencies