Add QuantileOutput support to DeepAR

[timoschowski]

Summary

• Add support for QuantileOutput in DeepAR, enabling direct quantile regression (pinball loss) as an alternative to distribution-based outputs trained with NLL loss
• Refactor forward() into _forward_distribution() (existing path, unchanged) and _forward_quantile() (new path using median for autoregressive feedback)
• Wire QuantileForecastGenerator in the estimator's create_predictor() so end-to-end training and inference works out of the box
• Add unit tests and two comparison example scripts (synthetic data + electricity dataset)

Motivation

DeepAR currently only supports distribution-based outputs (e.g. StudentTOutput, NormalOutput). This PR gives users a simpler, non-parametric alternative — QuantileOutput — that directly predicts quantile values via pinball loss, similar to what MQ-CNN already supports. This is useful when users want specific quantile forecasts without assuming a parametric distribution family.

Usage

from gluonts.torch.model.deepar import DeepAREstimator
from gluonts.torch.distributions.quantile_output import QuantileOutput

estimator = DeepAREstimator(
    freq="H",
    prediction_length=24,
    distr_output=QuantileOutput(quantiles=[0.1, 0.5, 0.9]),
)
predictor = estimator.train(training_data)

Changes

File	Change
src/gluonts/torch/model/deepar/module.py	Widen distr_output type to Output, dispatch forward() to distribution/quantile paths, add _forward_quantile(), guard output_distribution() and log_prob()
src/gluonts/torch/model/deepar/estimator.py	Widen distr_output type, select QuantileForecastGenerator vs SampleForecastGenerator in create_predictor()
test/torch/model/test_deepar_modules.py	Add test_deepar_quantile_output() — shapes, loss, log_prob error
examples/deepar_quantile_comparison.py	Comparison on synthetic sine data: NormalOutput vs QuantileOutput
examples/deepar_electricity_studentt_vs_quantile.py	Comparison on electricity dataset: StudentTOutput vs QuantileOutput with GluonTS Evaluator

Design decisions

1. Median for autoregressive feedback: The P50 quantile (or closest to 0.5) is fed back into the RNN at each prediction step
2. Return normalized predictions: _forward_quantile() returns quantile values in scale-normalized space; QuantileForecastGenerator handles scale multiplication — consistent with how QuantileOutput.loss() already works
3. num_parallel_samples ignored for QuantileOutput: Quantile predictions are deterministic, no sampling needed
4. Branching via isinstance: Simple, localized, fully backward-compatible — no changes to the Output class hierarchy

Test plan

• python -m pytest test/torch/model/test_deepar_modules.py -v — all 11 tests pass (4 existing distribution tests + 6 RNN input tests + 1 new quantile test)
• examples/deepar_quantile_comparison.py — trains both models, produces forecasts, prints metrics, saves plot
• examples/deepar_electricity_studentt_vs_quantile.py — end-to-end on electricity dataset with GluonTS Evaluator metrics

🤖 Generated with Claude Code

[timoschowski]

[carlocav]

Hello. Is it possible to use quantile loss for computing the validation loss and keep likelihood for the training loss? Deepar usually is unstable, and minimizing a validation quantile loss while still training the model to predict the full distribution rather than quantiles could be useful. This is implemented in nixtla neuralforecast which however lacks essential functions such as generating dependent sample paths