pith. sign in

arxiv: 2604.26668 · v1 · submitted 2026-04-29 · 📊 stat.ME

Nonlinear Probabilistic Forecast Reconciliation

Anubhab Biswas , Lorenzo Zambon , Lorenzo Nespoli , Giorgio Corani This is my paper

Pith reviewed 2026-05-07 10:42 UTC · model grok-4.3

classification 📊 stat.ME
keywords forecast reconciliationprobabilistic forecastingnonlinear constraintsUnscented Kalman Filterprojectionconditioningcoherent forecaststime series
0
0 comments X

The pith

Probabilistic forecasts with nonlinear constraints are reconciled by projecting samples onto the coherent manifold or conditioning them with UKF-inspired sampling.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes the first methods for probabilistic forecast reconciliation under nonlinear constraints by extending both projection and conditioning techniques from the linear case. Projection maps individual forecast samples onto the nonlinear manifold that satisfies the constraints, while conditioning uses an Unscented Kalman Filter-inspired sampler to adjust the joint distribution. A sympathetic reader cares because many real forecasting problems involve nonlinear relationships such as products, ratios, or physical laws, where standard linear reconciliation would produce incoherent or biased probabilistic predictions.

Core claim

We address probabilistic forecast reconciliation with nonlinear constraints for the first time. We extend both reconciliation via projection and conditioning to the case of nonlinear constraints. The projection approach reconciles forecast samples by mapping them onto the nonlinear coherent manifold. The conditioning approach adopts a sampling algorithm inspired to the Unscented Kalman Filter. Empirical evaluation on synthetic and real datasets shows both methods generally improve forecast accuracy, with the UKF-based approach achieving the best overall performance while being substantially faster than projection.

What carries the argument

Projection of forecast samples onto the nonlinear coherent manifold and UKF-inspired conditioning sampling that adjusts distributions to satisfy the nonlinear constraints.

If this is right

  • Both reconciliation approaches generally improve forecast accuracy on synthetic and real data.
  • The UKF-based conditioning method achieves the best overall performance among the tested approaches.
  • The UKF-based method runs substantially faster than the projection method.
  • Reconciliation becomes feasible for forecasting problems whose variables obey nonlinear rather than linear relationships.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The projection approach could become practical in high-dimensional settings if paired with efficient manifold optimization routines.
  • UKF conditioning may generalize to other sampling schemes that preserve more moments of the original forecast distribution.
  • Domains with physical or economic nonlinearities, such as energy production or supply-chain ratios, could now use coherent probabilistic forecasts where only point forecasts were feasible before.

Load-bearing premise

The nonlinear constraints are known exactly and can be evaluated or sampled from without introducing large approximation errors that invalidate the reconciled distributions.

What would settle it

Applying the projection or UKF methods to a dataset with known nonlinear constraints and observing no improvement or degradation in probabilistic scores such as CRPS relative to the original forecasts would falsify the accuracy claim.

Figures

Figures reproduced from arXiv: 2604.26668 by Anubhab Biswas, Giorgio Corani, Lorenzo Nespoli, Lorenzo Zambon.

Figure 1
Figure 1. Figure 1: Example of time series with nonlinear constraints; the free time series are view at source ↗
Figure 2
Figure 2. Figure 2: Example of nonlinear probabilistic reconciliation of the ratio constraint described in Fig. view at source ↗
Figure 3
Figure 3. Figure 3: Representation of nonlinear probabilistic reconciliation via conditioning view at source ↗
Figure 4
Figure 4. Figure 4: Simulation study: 3-D surfaces generated by the synthetic data - ripples (top left), saddle (top right), and view at source ↗
Figure 5
Figure 5. Figure 5: Number of acquired citizenships per year, existing foreigners and the citizenship ratio for Canton Ticino. view at source ↗
Figure 6
Figure 6. Figure 6: The structure of immigration rates disaggregated by Swiss Cantons contains both linear (solid) and view at source ↗
Figure 7
Figure 7. Figure 7: Improvement for UKF against the Base forecast and PBU for each Canton on the two datasets. view at source ↗
Figure 8
Figure 8. Figure 8: Structure of the Australian tourism case study. The bottom level contains quarterly tourism counts for view at source ↗
read the original abstract

Forecast reconciliation adjusts independently generated forecasts so that they satisfy some known constraints. While probabilistic forecast reconciliation is well established for linear constraints, some practical forecasting problems involve nonlinear relationships among variables. In this paper, we address probabilistic forecast reconciliation with nonlinear constraints for the first time. We extend both reconciliation via projection and conditioning to the case of nonlinear constraints. The projection approach reconciles forecast samples by mapping them onto the nonlinear coherent manifold. The conditioning approach adopts a sampling algorithm inspired to the Unscented Kalman Filter (UKF). We evaluate both methods on synthetic and real datasets. Empirically, both reconciliation approaches generally improve forecast accuracy. The UKF-based approach achieves the best overall performance while being substantially faster than the projection one.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper claims to be the first to address probabilistic forecast reconciliation with nonlinear constraints. It extends both the projection method (by mapping forecast samples onto the nonlinear coherent manifold) and the conditioning method (via a UKF-inspired sampling algorithm that propagates sigma points through the nonlinear constraint) to the nonlinear setting. Empirical evaluation on synthetic and real datasets shows that both approaches generally improve forecast accuracy, with the UKF-based method achieving the best performance while being substantially faster than projection.

Significance. If the approximation quality holds, the work fills a clear gap in probabilistic forecasting by handling nonlinear coherence constraints that arise in applications such as energy systems or hierarchical time series with multiplicative or trigonometric relations. Providing two complementary algorithmic approaches (projection and conditioning) with reported speed-accuracy trade-offs is practically useful. The empirical gains on both synthetic and real data strengthen the case for adoption, though the absence of theoretical error bounds or exact-sampling benchmarks limits the strength of the conclusions.

major comments (3)
  1. [§3.2] §3.2 (UKF conditioning): The sigma-point propagation and subsequent Gaussian reconstruction implicitly assume that the base forecast remains approximately Gaussian after the nonlinear mapping and that higher-order curvature effects are negligible. No error bounds, bias diagnostics, or comparisons to exact conditional sampling (MCMC or rejection sampling) are provided for strongly nonlinear or non-Gaussian cases. This is load-bearing because the central claim of reliable accuracy improvement rests on the reconciled samples being distributed according to the true conditional law.
  2. [§3.1] §3.1 (projection method): Independent per-sample optimization maps points onto the manifold but does not specify whether the mapping is measure-preserving or corresponds to any particular conditional distribution. Consequently, the empirical distribution of the projected samples may not equal the reconciled probabilistic forecast; no diagnostic (e.g., density estimation or moment matching against a known conditional) is reported.
  3. [§4] §4 (experiments): Accuracy improvements are reported without quantifying approximation error on controlled synthetic examples where the true nonlinear conditional can be computed exactly. This leaves open whether observed gains arise from correct reconciliation or from uncontrolled bias introduced by the UKF linearization or the projection heuristic.
minor comments (3)
  1. [§4] The precise functional form of the nonlinear constraints used in the real-data experiments should be stated explicitly (e.g., in a table or appendix) so that readers can reproduce the manifold.
  2. [§4] Add statistical significance tests or confidence intervals to the reported accuracy improvements to distinguish genuine gains from sampling variability.
  3. [§3.1] Clarify the computational complexity of the projection optimization step and whether warm-starting or analytic Jacobians are used.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed review. The comments correctly identify key limitations in the validation of our proposed methods. We agree that stronger empirical diagnostics and explicit discussion of approximations are needed, and we outline targeted revisions below. We address each major comment point by point.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (UKF conditioning): The sigma-point propagation and subsequent Gaussian reconstruction implicitly assume that the base forecast remains approximately Gaussian after the nonlinear mapping and that higher-order curvature effects are negligible. No error bounds, bias diagnostics, or comparisons to exact conditional sampling (MCMC or rejection sampling) are provided for strongly nonlinear or non-Gaussian cases. This is load-bearing because the central claim of reliable accuracy improvement rests on the reconciled samples being distributed according to the true conditional law.

    Authors: We agree that the UKF-based conditioning is an approximation relying on moment matching via sigma points, which does not guarantee exact conditional sampling for strongly nonlinear or non-Gaussian base forecasts. This mirrors standard limitations of the unscented transform. In the revised manuscript we will expand the discussion in §3.2 to explicitly state these assumptions and their potential impact. We will also add a new set of low-dimensional synthetic experiments comparing the UKF samples to reference distributions obtained via MCMC or rejection sampling, reporting bias and moment-matching diagnostics to quantify the approximation quality. revision: yes

  2. Referee: [§3.1] §3.1 (projection method): Independent per-sample optimization maps points onto the manifold but does not specify whether the mapping is measure-preserving or corresponds to any particular conditional distribution. Consequently, the empirical distribution of the projected samples may not equal the reconciled probabilistic forecast; no diagnostic (e.g., density estimation or moment matching against a known conditional) is reported.

    Authors: The projection approach is presented as a heuristic that enforces coherence by mapping each sample to the nonlinear manifold, extending the linear projection method. It does not claim to produce samples from the exact conditional distribution. We acknowledge the absence of distributional diagnostics. In revision we will add moment-matching checks (means, covariances) and, on selected synthetic examples, density or quantile comparisons against reference reconciled distributions to better characterize the empirical output of the projection procedure. revision: yes

  3. Referee: [§4] §4 (experiments): Accuracy improvements are reported without quantifying approximation error on controlled synthetic examples where the true nonlinear conditional can be computed exactly. This leaves open whether observed gains arise from correct reconciliation or from uncontrolled bias introduced by the UKF linearization or the projection heuristic.

    Authors: We concur that the current experiments would be strengthened by direct quantification of approximation error against exact conditionals. While our synthetic setups already include nonlinear constraints, we did not compute reference distributions. In the revised §4 we will introduce controlled low-dimensional synthetic cases (e.g., quadratic or multiplicative constraints) for which the true conditional can be obtained numerically or by dense sampling, and we will report metrics such as KL divergence, Wasserstein distance, or moment errors between the reconciled samples and these exact references. revision: yes

standing simulated objections not resolved
  • Derivation of theoretical error bounds or convergence guarantees for the UKF approximation and the projection heuristic under arbitrary nonlinear constraints.

Circularity Check

0 steps flagged

No significant circularity; direct algorithmic extensions of linear methods

full rationale

The paper's core contribution is the extension of projection (mapping samples onto the nonlinear coherent manifold) and conditioning (UKF-inspired sampling) from linear to nonlinear constraints. These steps are described as novel applications without any equations that reduce by construction to fitted parameters, self-defined quantities, or load-bearing self-citations. The abstract and description present the methods as independent algorithmic innovations, with empirical evaluation on synthetic and real datasets providing external validation. No self-definitional loops, renamed known results, or uniqueness theorems imported from prior author work are present in the provided material. This is a standard non-circular extension case.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract provides insufficient detail to enumerate specific free parameters or invented entities; the work rests on the domain assumption that nonlinear constraints are known and evaluable.

axioms (1)
  • domain assumption Nonlinear constraints among forecast variables are known and can be evaluated or sampled from
    The projection and conditioning methods presuppose that the constraints are given and tractable.

pith-pipeline@v0.9.0 · 5414 in / 1123 out tokens · 48699 ms · 2026-05-07T10:42:37.852869+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

  1. [1]

    Forecasting: Principles and Practice

    Athanasopoulos, G., Ahmed, R. A., & Hyndman, R. J. (2009). Hierarchical forecasts for aus- tralian domestic tourism.International Journal of Forecasting,25, 146–166. URL:https:// 24 Method Cantonal IR Cantonal CR Baseline Base 1 1 PBU 0.839 0.973 Projection OLS 1.088 1.267 WLS 0.837 0.96 Block 0.837 0.96 FULL 0.834 0.949 Conditioning UKF0.808 0.94 Table D...

    work page doi:10.1016/j.ijforecast.2008.07.004 2009

  2. [2]

    doi:10.1080/01621459.2020

    URL:https://doi.org/10.1080/01621459.2020.1736081. doi:10.1080/01621459.2020. 1736081.arXiv:https://doi.org/10.1080/01621459.2020.1736081. Wickramasuriya, S. L. (2024). Probabilistic forecast reconciliation under the gaus- sian framework.Journal of Business & Economic Statistics,42, 272–285. URL: https://doi.org/10.1080/07350015.2023.2181176. doi:10.1080/...

    work page doi:10.1080/01621459.2020.1736081 2020

  3. [3]

    Zambon, L., Azzimonti, D., Rubattu, N., & Corani, G. (2024b). Probabilistic reconciliation of mixed-type hierarchical time series. In N. Kiyavash, & J. M. Mooij (Eds.),Proceedings of the Fortieth Conference on Uncertainty in Artificial Intelligence(pp. 4078–4095). PMLR volume 244 ofProceedings of Machine Learning Research. URL:https://proceedings.mlr.pres...

    work page doi:10.1016/j.ejor.2024.05.024 2024