[ENH] Online Quantile Regression with Pinball-Loss Optimizers

[sisird864]

Motivation: River lacks native quantile regression for streaming tasks; this enables prediction intervals and robust regression. (Related: “Quantile Regression Approaches”.)

Proposal:

• New estimator(s) linear.QuantileSGDRegressor supporting multiple τ in one pass.
• Optimizer: per-sample pinball loss with optional forgetting factor.
• API parity with linear_model regressors (fit_one/learn_one/predict_one).
• Benchmarks: synthetic piecewise-stationary streams; compare vs. batch baselines.
• Docs: example + tutorial notebook.

Questions: preference for single-τ estimator vs multi-τ; default regularization; placement under linear?

[MaxHalford]

Hey there @sisird864. I agree that River could provide more utilities to provide prediction internals. That's the aim of the conf module, which implements conformal prediction algorithms.

Questions: preference for single-τ estimator vs multi-τ; default regularization; placement under linear?

I think single-τ estimator is more than fine. Not sure about the implications of your two other questions really. But I would start simple and build from there.