acceptodds
Under review as a conference paper at ICLR 2027

CMQNet: Continuous and Monotone Quantile Networks

Abstract

Deep quantile regression has emerged as a powerful paradigm for characterizing conditional distributions and quantifying uncertainty. However, existing neural network approaches are often limited to estimating quantiles at discrete grids, resulting in discontinuity with respect to the quantile level and failing to structurally guarantee the non-crossing property. To address these challenges, we propose the Continuous and Monotone Quantile Network (CMQNet). Leveraging the fundamental theorem of calculus for absolutely continuous functions, CMQNet models the conditional quantile function as the summation of a baseline term and the integral of a non-negative derivative with respect to the quantile index. By modeling this derivative using a neural network with strictly positive activation functions, our framework naturally ensures both monotonicity and continuity by construction, thereby eliminating quantile crossing and admitting a well-defined conditional density. Furthermore, we introduce an efficient training strategy based on stochastic quantile sampling, allowing the model to learn the global quantile process beyond fixed grids. Extensive experiments on simulated datasets and real-world benchmarks demonstrate that CMQNet significantly outperforms state-of-the-art methods in terms of estimation accuracy and stability. The code can be found at https://anonymous.4open.science/r/CMQNet-C420/.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Deep Quantile Networks, Non-crossing, Fundamental Theorem of Calculus

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_g8PbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a quantile regression model that prevents quantile crossing in conditional quantile estimation. Its core strategy models nonnegative increments between quantile values, an approach previously used to ensure non-crossing quantile estimates. Although the proposed approach is reasonable, it offers little methodological novelty.

Critical strengths
  • The direction and strategy described in this paper are consistent with those of existing studies.

  • Comparative experiments were conducted across a variety of datasets.

Critical weaknesses
  • The use of nonnegative increments and integral formulations, as advocated in this paper, is not new.

  • Although the paper provides theoretical proofs, its poorly defined notation makes them nearly impossible to follow. Page 6 even uses symbols before they are properly defined, making the intended arguments difficult to discern. For example, ss and pp, introduced on line 296, appear to characterize the smoothness of a function class, but they are explained only in the appendix. The paper also fails to explain how this smoothness relates to Theorem 1. Similarly, FCMQ,K\mboxgrid\mathcal{F}_{CMQ,K}^{\mbox{grid}}, used around line 298 on page 6, appears to be introduced to calculate the complexity of the function class, yet it is used without a definition. These issues make the theoretical discussion very difficult to follow.

Questions for authors
  • Do the results in Theorem 1 establish a non-asymptotic bound for the proposed method? If not, how do they help demonstrate the method’s effectiveness?

  • Beyond the evaluation metrics reported in the simulation study, are there any comparative results using additional calibration metrics?

Reviewer_zEhaOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CMQNet models conditional quantiles using a median network and an integrated positive derivative network. It uses randomized quantile sampling and a fixed-grid implementation to produce continuous, non-crossing predictions.

Critical strengths

The positive-increment construction preserves monotonicity under the stated numerical implementation.

Critical weaknesses
  1. Both positive-derivative integration and randomized quantile training already appear in Brando et al. (2022), Eqs. (3)–(4). The substantive advance beyond changes in parameterization and numerical approximation is not convincingly established. This closest related method is also absent from the experiments.
  2. Eq. (4) defines a piecewise-linear quantile predictor, whose induced density is piecewise constant. Section 7.6 instead uses the reciprocal of the continuous derivative network. These generally describe different distributions, undermining the claimed smooth induced density and its evaluation.

Axel Brando, Joan Gimeno, José A. Rodríguez-Serrano, and Jordi Vitrià. Deep Non-Crossing Quantiles through the Partial Derivative. In Proceedings of the 25th International Conference on Artificial Intelligence and Statistics (AISTATS), Proceedings of Machine Learning Research, 151:7902–7914, 2022.

Questions for authors
  1. What substantive capability or theoretical guarantee does CMQNet add beyond Brando et al. (2022)? Please identify the precise differences and provide a direct comparison with that method.

  2. For τ∈(qj−1,qj)\tau\in(q_{j-1},q_j), Eq. (4) implies

∂τh^(x,τ)=MLPD(x,qˉj).\partial_\tau\widehat h(x,\tau)=\mathrm{MLP}_D(x,\bar q_j).

Why does Section 7.6 use 1/MLPD(x,τ)1/\mathrm{MLP}_D(x,\tau), rather than the reciprocal of this interval slope? Does the reported density integrate to one?

  1. Is Figure 4 evaluated using the ideal continuous model or the implemented grid predictor? Please report density results consistent with the predictor actually trained, and clarify whether the claimed smoothness survives that implementation.
Reviewer_sCFyOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Continuous and Monotone Quantile Networks (CMQNet) for conditional quantile estimation. The method constructs a continuous, non-crossing quantile function by integrating a non-negative derivative network around a learned conditional median. The integral is approximated numerically using a fixed grid and linear interpolation. The authors also introduce stochastic quantile sampling for pinball-loss training and establish theoretical approximation and estimation guarantees. Experiments on synthetic and real-world datasets show improved accuracy over existing deep quantile regression methods, including NQ-Net and DQRP.

Critical strengths
  1. The fomulation and framework is simple.
  2. The paper provides theoretical results on the approximation adn estimation errors fo the proposed estimator, providing theoretical justification fo the implemented framework.
Critical weaknesses
  1. Limited practical significance and scope of the paper. The proposed method provides additional theoretical properties, such as continuity and monotonicity of the quantile function. But those doesn't translate to practical advantage of the method in real applications also the experiments suggest little gain.

  2. Limited methodological novelty compared with prior work. The core idea of CMQNet appears closely related to Deep Non-Crossing Quantiles through the Partial Derivative. That work already represents a quantile function as the integral of a positive neural-network-parameterized derivative. It also employed stochastic sampling of quantile levels during training. Although the authors cite this work, the methodological differences are not sufficiently discussed.

  3. The experiments are limited to simulation and very easy tabular regression dataset.

Questions for authors

NA

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.