In traditional time series forecasting, we typically ask: "What is the most likely value for tomorrow?" This leads to point estimates—single numbers that represent the conditional mean of a distribution. However, in high-stakes environments like power grid management, pandemic planning, or financial trading, the "average" outcome is often the least interesting one. Decision-makers need to know the risks at the edges: the probability of a massive electricity price spike, the worst-case scenario for hospital admissions, or the 90th percentile of consumer demand.
To address this, researchers have shifted toward probabilistic forecasting, with Quantile Regression (QR) emerging as the primary mechanism for quantifying uncertainty without making restrictive assumptions about the shape of the data.
The Intuition: Slicing the Distribution
The core idea behind quantile regression is to estimate specific percentiles (quantiles) of the target variable's distribution. While a standard regression model minimizes Mean Squared Error (MSE) to find the mean, QR uses the pinball loss function.
The pinball loss works by applying asymmetric penalties to errors. If you want to estimate the 90th percentile, the loss function penalizes under-estimation much more heavily than over-estimation. This forces the model to "hedge" its predictions upward until only 10% of the actual data points fall above the predicted line. As Pierre Alquier and Xiaoyin Li demonstrate in Prediction of Quantiles by Statistical Learning and Application to GDP Forecasting, this allows for the construction of non-parametric confidence intervals that remain valid even when the underlying data is "weakly dependent" or non-Gaussian.
From Linear Models to Deep Architectures
Early applications of QR relied on linear models or ensembles. Tarik S. Salem and colleagues utilized Quantile Regression Forests to manage intra-hour electricity imbalances, showing that tracking the full distribution of samples in tree leaves provides a robust uncertainty measure.
However, the field has rapidly evolved toward deep learning. Modern frameworks like GluonTS, introduced by A. Alexandrov and Konstantinos Benidis, treat time series as "global" problems, training neural networks across thousands of related sequences to learn shared patterns of volatility.
The integration of QR into Transformers has been particularly transformative. Sifan Wu and Peilin Zhao developed the Adversarial Sparse Transformer (AST), which uses sparse attention to handle long sequences and a discriminator to ensure the generated quantile paths look like real-world "spiky" data. Similarly, Zhiyuan Zhao and Boris N. Oreshkin introduced the Temporal-Aligned Transformer (TAT), which uses multi-quantile loss to predict peak demand during volatile e-commerce promotions.
Why It Works: Handling Non-Linear Uncertainty
The primary advantage of quantile-based methods is their ability to capture heteroscedasticity—the phenomenon where uncertainty changes over time or across different conditions.
In solar power forecasting, for instance, Benjamin F. Hobbs and Qingyu Xu observed a "U-shaped" uncertainty profile in the Pro2R project: variability is low when it is very sunny or very dark, but extremely high at intermediate irradiance levels. Quantile regression naturally maps these non-linear shifts in risk, whereas a standard mean-based model would produce a constant (and likely incorrect) error margin.
This flexibility extends to biological signals. In scTEL, Yuanyuan Chen and Xiaodan Fan use quantile loss to distinguish between reliable protein markers and technical noise in single-cell data, providing a range of certainty for biological discoveries.
Bridging Prediction and Optimization
The ultimate goal of predicting quantiles is often to make better decisions. This is the "Predict and Optimize" philosophy. Zahra Namazian and Peter J. Stuckey demonstrate this in beauty retail inventory management, where they sample from predicted quantiles to simulate thousands of future demand scenarios. By optimizing against these scenarios rather than a single point forecast, they significantly reduce stockouts and operational costs.
Similarly, in humanitarian aid, the Foresight system developed by Rahul Nair uses quantile regressors to help the Danish Refugee Council assess the risk levels of forced displacement, moving from reactive crisis response to proactive planning.
Calibration and Reliability
A common challenge with quantile regression is ensuring that the "90th percentile" predicted by the model actually corresponds to the 90th percentile in reality. Charles Marx and Willie Neiswanger addressed this with Modular Conformal Calibration (MCC). By using quantiles as a base modeling strategy and passing them through a calibration framework, they produce uncertainty intervals that are both "sharp" (informative) and theoretically guaranteed to be accurate.
As Qi Dong notes in a comprehensive survey of electricity load forecasting, the field is decisively moving away from "black-box" point predictions. By embracing quantile regression, researchers are building systems that don't just predict the future, but quantify our ignorance of it.