acceptodds
Under review as a conference paper at ICLR 2027

PatchTransferNet: A Mechanism-Driven Demand Forecasting Model

Abstract

Classical time series forecasting models are mostly designed for slowly varying, smooth, continuous signals (M1). By contrast, retail demand is discrete and nonnegative, often intermittent or sparse. We model retail demand as an observation-driven counting process (M2), and then develop PatchTransferNet (PtNet), which translates the M2 recursion directly through mean-preserving convex decoding and aggregate-consistent supervision. Sparse series put the two targets in conflict. Per-day medians can be all zero under a positive window total, so daily supervision misses the total and aggregate supervision cannot distribute it. An idealized mixed-Poisson window turns this tension into a two-fingerprint rule. Firstly, the zero-mass share gates aggregate supervision on sparse populations, and then the benefit grows with the count autocorrelation , the idealization places the crossing at for sales cycle . The rule places the optimum correctly on all three demand datasets. Two of them sit at opposite ends of the range, and the third splits by the reporting metric, where the daily caliber selects and the official aggregate metric selects . On FreshRetailNet-50K, PtNet reaches 27.94% WAPE, beats every single-stage forecaster in all five seeds by at least 1.4 pt (worst seed vs. best baseline ), and shrinks its under-forecast to , matching the best two-stage pipeline. On the public M5 benchmark it ties the best published store-item-level score (). When transplanted into a linear backbone on ten classical benchmarks, the anchor's gain tracks , so each data type rewards its own inductive bias.

Primary area:
learning on time series and dynamical systems
Keywords:
Demand forcasting, Time Series Forecasting, Observation-Driven Counting Process, Aggregate Supervision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_diKMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PatchTransferNet (PTNet), a retail demand forecasting model combining multi-scale patch attention, convex level anchoring, and aggregate-level supervision. It introduces a two-fingerprint decision rule based on zero-mass share and lag-one autocorrelation to guide loss selection. Experiments on three retail datasets and ten classical time-series benchmarks demonstrate competitive forecasting performance and investigate the effectiveness of the proposed inductive bias across different data regimes.

Critical strengths
  1. The paper explicitly considers the discrete, sparse, and intermittent characteristics of retail demand, providing a meaningful alternative to conventional continuous-signal forecasting approaches.

  2. PTNet achieves competitive performance on retail forecasting benchmarks, particularly on FreshRetailNet-50K, where it outperforms the compared single-stage baselines. The ablation studies and cross-domain experiments further support the proposed design.

Critical weaknesses
  1. In the abstract, "Firstly" appears, but "Secondly, Thirdly, Finally" does not appear afterwards, which may appear a bit weak in logic.

  2. The organization, layout, and content of the entire paper are a bit rough. The paper is not fully written, for example, there is still half a page left on page nine, but it has not been effectively utilized, and the introduction and contribution sections are written too vaguely.

  3. There are multiple 1-2 words occupying a line in the article (such as line 43, 87, 122, 256, 296, 409), which need to be adjusted and modified.

  4. The innovation of the paper's methodology is insufficient. The core components of PTNet are all reasonable, but the article needs to clarify what advantages these components can bring.

  5. Most baseline models use fixed random seed results, but for PTNet, 5 seed experiments were conducted.

  6. The learning rate of PatchTST on M5 is 2×10−32\times10^{-3}, and the WAPE is 74.90. However, the appendix indicates that the WAPE can reach 70.51 at 10−310^{-3}, which suggests that the baseline results in the main table of the article may not be the optimal results. The author needs to explain why the main table adopts a poor configuration.

  7. Section 5.4 claims that four datasets Rlast2R_{\text{last}}^2 exceeded the set threshold and benefited from anchoring, but in Table 5 of Appendix, only the Traffic, Weather, and Electricity datasets Rlast2>0.04R_{\text{last}}^2 > 0.04 exceeded 0.04. Although Solar is below the threshold, there has been some improvement, and the author needs to explain this.

Questions for authors

Please see Weaknesses.

Reviewer_uHh1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PatchTransferNet, a neural demand forecasting model motivated by observation-driven count processes. The architecture constructs historical patches at multiple temporal scales, uses future-covariate queries to combine their raw sales means through attention, and applies multiplicative, additive, and quantile-spread corrections to produce nonnegative, ordered quantile forecasts. The training objective blends daily pinball loss with a loss on aggregated forecasts. The paper’s main conceptual claim is that the usefulness of aggregate supervision can be predicted from two population fingerprints: lag-one count autocorrelation and the proportion of series exhibiting a zero daily median but a positive aggregate median. An idealized mixed-Poisson model motivates the threshold (1/(W+1)). Experiments cover three demand datasets and an anchor-transplant diagnostic on ten classical forecasting benchmarks. The strongest empirical result is on FreshRetailNet-50K, where all five reported PatchTransferNet runs outperform the reported single-stage baselines.

Critical strengths
  1. The use of historical sales means as raw-scale anchors, combined with calendar-dependent attention and correction heads, gives the model an interpretable connection to demand levels. Nonnegative outputs and noncrossing quantiles are useful structural properties.
  2. This paper recognizes that the sum of daily optimal quantiles need not equal the optimal quantile of the aggregate. The allocation-unidentifiability observation and conditional zero-median example provide useful motivation for studying the balance between daily and aggregate supervision.
  3. Meaningful empirical gains on FreshRetailNet-50K. The five reported FRN runs achieve WAPE between 27.94% and 29.60%, all below the strongest reported single-stage baseline score of 31.06%. The median is 28.88%. The no-static-input ablation also remains competitive at 29.38%, providing evidence that the full improvement cannot be attributed solely to static identifiers.
Critical weaknesses
  1. The theoretical correlation and the empirical fingerprint do not generally estimate the same statistical quantity. Appendix A.2 draws a persistent level (\mu) per series and assumes conditionally independent Poisson observations. This correlation reflects heterogeneity between series. In contrast, Table 3 defines the empirical fingerprint as the average of lag-one ACFs computed separately for each series. For a series with a fixed (\mu) and conditionally iid observations, its demeaned sample ACF approaches zero as its history grows, even when the marginal correlation above is large. Averaging such ACFs does not recover the between-series level share. Consequently, identifying the two quantities requires additional temporal modeling assumptions and an appropriate estimator. This is a statistical-object mismatch, rather than merely finite-sample attenuation. It undermines the current justification for reading the (1/(W+1)) crossing directly on the reported fingerprint.
  2. Proposition 3 identifies when one variance component exceeds another. It does not derive a comparison of expected forecasting losses, an optimal loss weight, or the absence of allocation cost. Likewise, the absence of zero daily medians does not eliminate incompatibility between daily and aggregate medians.
  3. The no-static ablation is useful, but it does not establish how much of the gain comes from the anchor architecture versus applying aggregate supervision to an otherwise strong forecaster. The comparison set also lacks a directly relevant neural count-likelihood forecaster. Appendix B reports M5 PatchTST WAPE of 74.90 at the shared learning rate (2\times10^{-3}), while also reporting 70.51 at (10^{-3}). A shared learning rate does not necessarily provide a balanced comparison across architectures.
Questions for authors
  1. How does the between-series level share in Appendix A.2 correspond to the average within-series ACF used in Table 3? If the latent level is intended to vary over time within each series, please specify that process and derive the corresponding relationship.
  2. What additional assumptions connect the variance crossing to improvement in expected pinball loss? How does the claim about near-zero zero-mass share accommodate the positive-median Poisson example above?
  3. Section 4.3 uses a single whole-horizon aggregate for SPR, whose horizon is 21 days, while the theoretical crossing uses (W=7). Why is the relevant variance-comparison window seven rather than 21 days in this experiment?
  4. Please enumerate the future covariates supplied to each model. Appendix B explicitly gives PatchTransferNet future weather and discount information, while describing the TSLib baselines through historical encoder channels and future calendar marks. Also, what evidence is available for the preferred loss weights across seeds, and for a strong baseline trained with the same aggregate objective?
Reviewer_PidEOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how the structure of retail demand should inform forecasting model design and training objectives. It models demand as an observation-driven count process and proposes PatchTransferNet, which combines patch-level historical sales means through convex weights and refines the resulting anchor with correction heads.

The paper also examines when to combine daily pinball loss with a loss on windowed sales totals. It proposes two diagnostic statistics: lag-one count autocorrelation and the share of series whose daily median is zero despite a positive median window total. Under an idealized mixed-Poisson model, it derives a signal-to-noise crossing at ρ = 1/(W+1). Together, these statistics motivate a rule for choosing whether to emphasize aggregate supervision.

Experiments on FreshRetailNet-50K, an internal retail dataset, and M5 show different preferred loss weights, broadly consistent with the proposed rule. A separate experiment transfers the historical anchor into DLinear on ten conventional forecasting benchmarks to examine where this inductive bias helps or hurts.

Critical strengths
  1. A practical guideline for choosing aggregate supervision. The paper uses autocorrelation and zero-mass share to guide the loss weight λ. An idealized analysis gives a threshold of ρ₀=1/(W+1). Cross-domain experiments also examine where the proposed bias hurts performance, helping clarify its scope.

  2. Strong single-stage performance, supported by multi-seed results and ablations. PtNet performs competitively with two-stage pipelines on FreshRetailNet-50K, using 1.33M parameters. Across five seeds, it outperforms the best single-stage baseline by at least 1.4 WAPE points. Replacing the anchor-derived heads with a single mean head worsens WAPE by 3.41 points, suggesting that the decoding design contributes to the gains.

Critical weaknesses
  1. Limited independent validation of the decision rule. The threshold ρ₀=1/(W+1) is derived under an idealized model, while the R²_last boundaries are empirical. Results on three demand datasets are encouraging, but do not establish how reliably the rule selects λ on unseen datasets. Testing it on additional public datasets would strengthen the claim.

  2. Missing comparisons with count-distribution objectives. Despite motivating the method through a negative-binomial count process, the paper trains with pinball loss. Comparisons with negative-binomial likelihood, zero-inflated, or hurdle objectives would help determine whether the proposed supervision offers advantages over established approaches to sparse demand.

Questions for authors
  • Although ρ₀ is derived under an idealized model, its applicability to real data and the empirical R²_last thresholds are evaluated on the datasets used to support the rule. Have you tested whether these criteria predict the preferred λ or the direction of the anchor’s benefit on additional held-out datasets, without adjusting the thresholds? Such evidence would help distinguish a predictive guideline from a post-hoc explanation and would strengthen my assessment of soundness.

  • Baseline comparability on FreshRetailNet-50K. Could you clarify which results in Table 1 were obtained from your own runs and which were taken from Wang et al. (2025)? Were the input features, data splits, and tuning budgets comparable across PtNet, the strongest single-stage baselines, and the two-stage TFT+TimesNet pipeline? This would help assess how much of the reported improvement can be attributed to the proposed method.

Comments on the paper

1 comment

Dear Reviewers,

I'm writing to gently remind you that the rebuttal discussion phase deadline is approaching. Please take a moment to read the authors' rebuttal, and feel free to respond if any concerns remain that you'd like them to address. Thank you very much for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.