Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

FreshRetailNet-50K: A Stockout-Annotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper introduces the first open stockout-annotated hourly fresh-retail dataset and shows that recovering censored demand before forecasting cuts systematic underestimation from 7.37% to near zero.

desk verdict A genuinely new retail dataset with hourly stockout labels that fills a real gap; the paper's main claims hold up, but the label-quality premise needs an audit and the synthetic evaluation is underspecified. read the letter →

arxiv 2505.16319 v5 pith:SSNSLVIE submitted 2025-05-22 cs.LG

classification cs.LG
keywords stockoutannotationcensoreddemandlatentrecoveryforecastingfreshretailhourlytimeseriesmissingnotatrandombenchmarkdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

FreshRetailNet-50K is a new public benchmark of 50,000 store-product hourly sales series from 898 fresh-retail stores, covering 863 perishable SKUs, with each hour labeled as stocked or stocked-out. The paper's central claim is that these stockout annotations make it possible to recover the demand that was censored during stockouts, and that feeding that recovered demand into a second-stage forecaster removes the systematic underestimation that ordinary sales-based models exhibit. The authors demonstrate the claim with a two-stage pipeline: first a latent demand recovery model fills the censored hours, then a forecaster predicts next-week totals from the recovered daily series. In their main experiment, the TFT forecaster's weighted absolute percentage error drops from 31.75% (raw sales) to 29.02% after TimesNet-based recovery, while the weighted percentage error moves from -7.37% (underestimation) to +2.58%, and to +0.57% with iTransformer recovery. If the dataset is sound, it gives the field a standard playground for censorship-aware demand modeling, inventory optimization, and causal retail analytics.

What carries the argument

The load-bearing object is the stockout annotation itself: an hourly binary stock-availability indicator $s_t$ derived from warehouse management system stock levels, recorded alongside hourly sales $y_t$. The paper formalizes the censoring mechanism as $s_t = 1$ when inventory is positive and $s_t = 0$ at stockout, and the recovery objective as $d = y \odot s + \hat{d} \odot (1 - s)$, where $\hat{d}$ is an imputed demand estimate. The two-stage pipeline then aggregates the reconstructed hourly demand into daily totals and trains a seven-day-ahead forecaster; the counterfactual is the same forecaster trained on raw sales aggregates.

What would settle it

Take a random sample of stockout hours and compare the warehouse stock-level records against an independent ground truth (physical inventory counts or order-level timestamps showing stock was actually present). If a meaningful fraction of labeled stockouts show positive available stock, or if labeled non-stockout hours show zero sales due to reasons other than stock absence, the dataset's central assumption fails. A simpler computational test: withhold stockout labels on a held-out set, train the recovery stage only on non-stockout hours, and check whether the forecast bias reappears.

Watch

Extended reading notes

Core claim

The central discovery on the paper's own terms is that a large-scale, hour-level stockout-annotated dataset makes latent demand recovery tractable and demonstrably useful: models trained on recovered daily demand are more accurate and far less downward-biased than models trained on raw censored sales. The stockout indicators act as a mask that separates true zero demand from supply-limited zero sales, breaking the missing-not-at-random mechanism that plagues standard imputation and forecasting. The paper reports that in the full-dataset TFT setting, recovery via TimesNet improves WAPE from 31.75% to 29.02% and WPE from -7.37% to +2.58%; recovery via iTransformer brings WPE to +0.57%. It also introduces a decoupling score measuring the correlation between stockout ratios and recovered demand, showing that raw sales carry a strong negative correlation (-0.57) that TimesNet recovery reduces to near zero (0.07).

Load-bearing premise

The entire pipeline and evaluation rest on the assumption that the warehouse system's hourly stock levels correctly identify every stockout, and that sales equal true demand whenever stock is available; if stockout labels are noisy, or zero sales have other causes, the recovered demand targets and the reported improvements inherit that error.

Editorial extensions

If this is right

  • Researchers can now benchmark latent demand recovery methods against a real missing-not-at-random dataset with ground-truth stockout masks, rather than synthetic censoring.
  • Forecasting models trained on recovered demand should exhibit near-zero bias in fresh retail, breaking the stockout-underestimation feedback loop.
  • The hourly resolution and covariates (promotions, weather, holidays) enable studies of how censoring itself depends on price, weather, and time of day.
  • The dataset supports a decoupling evaluation: a recovery method is good if the recovered demand is uncorrelated with the stockout ratio.
  • Daily-aggregated benchmarks like M5 and Favorita cannot detect or correct this censoring; the paper positions FreshRetailNet-50K as the first benchmark that can.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The stockout annotations may themselves be a training target: a model predicting the probability of stockout from covariates could be used as an early-warning system for replenishment.
  • The two-stage design suggests a natural test: if recovery quality is the driver, end-to-end models that jointly impute and forecast under a censoring mask should match or beat the pipeline, a comparison the paper leaves for future work.
  • Because the labels come from warehouse management system stock levels, the dataset's value hinges on label fidelity; an independent audit of a random sample of stockout labels against physical inventory or order timestamps would directly test that.
  • The decoupling score could be reused as a generic diagnostic for any retail dataset with stockout indicators, not just this one.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces FreshRetailNet-50K, a dataset of roughly 50,000 store-product hourly time series from 898 stores and 863 perishable SKUs over March–June 2024, with stockout annotations, promotion, weather, and calendar covariates. The authors formalize two tasks: latent demand recovery during stockouts and censoring-robust 7-day demand forecasting from recovered daily totals. They benchmark several imputation and forecasting methods, reporting that TimesNet-based recovery gives the best reconstruction accuracy and that feeding recovered demand into TFT reduces WAPE from 31.75% to 29.02% and WPE from -7.37% to +2.58% on the full dataset. The manuscript concludes that the dataset enables new research on censored demand estimation, and it releases data and code.

Significance. If the stockout labels are indeed reliable, FreshRetailNet-50K fills a genuine gap: no existing public retail benchmark provides hourly stockout annotations with rich covariates, and the two-stage recovery-forecasting pipeline directly addresses a well-known downward-bias problem in perishable demand forecasting. The open data and code release, the large multi-city scale, and the explicit formalization of recovery and forecasting tasks are valuable contributions. However, the dataset's central value rests on the correctness of the stockout labels and on the realism of the synthetic evaluation, and both points are currently insufficiently supported. The forecasting gains, while plausible, need statistical substantiation before the quantitative claims can be taken at face value.

major comments (4)
  1. [§3.1, §4.1 Eq. (1)] The stockout annotations are described as 'verified' in the abstract and contributions, but the manuscript provides no audit or validation of these labels. Section 3.1 states only that inventory dynamics are captured via warehouse management systems, and Eq. (1) defines the stockout indicator purely from the reported inventory level (s_t = 1 if i_t > 0, 0 otherwise). Since Eq. (2) uses this indicator as ground truth for recovered demand, and Section 5.3 evaluates forecasting only on periods believed to be stockout-free, any error in the inventory-derived labels — from shrink/waste write-offs, order cancellations, picking errors, or system lags — propagates directly into the recovered demand targets and all downstream metrics. Please add a validation study: for example, a random sample of stockout and non-stockout hours audited against order-level or store-operation records, with precision, recall, and error analysis, or a reconciliation of inventory changes against sales plus recorded waste. Without such evidence, the central claim of 'verified' annotations is not established.
  2. [§5.1, Table 2] The latent-demand recovery evaluation is performed on synthetic censored regions, not on the real stockout intervals. Section 5.1 says only that 'non-stockout periods' are used as ground truth and that synthetic censored regions are generated 'matching empirical patterns,' but it does not specify the mask-generation mechanism, the parameters, the random seed, or how well the synthetic masks reproduce the empirical stockout distribution. Because Table 2 and the decoupling scores in §5.2 are computed entirely on these synthetic data, the paper never directly measures recovery accuracy on the actual stockout periods that the dataset is meant to support. This is a load-bearing gap: the contribution is precisely the availability of real stockout annotations, yet those annotations are not used to evaluate recovery. Please provide a complete description of the simulation, a fidelity check against the empirical censoring pattern, and, if possible, a validation on real stockout periods using auxiliary information (e.g., orders that were placed but unfilled, or a holdout subset with manual demand reconstruction).
  3. [§5.3, Table 3] The forecasting comparison in Table 3 has no confidence intervals, significance tests, or multi-seed variability. Several headline differences are small in absolute terms: for TFT overall, WAPE values are 29.02% (TimesNet), 29.26% (iTransformer), and 29.54% (ImputeFormer), and the differences may be within run-to-run noise. Relatedly, the abstract's '2.73% improvement in prediction accuracy' is the absolute percentage-point reduction in WAPE (31.75% to 29.02%), not a relative improvement; the relative improvement is about 8.6%. Please report standard errors or confidence intervals over multiple seeds and state the improvement metric unambiguously. This is needed to support the paper's quantitative claims about the two-stage pipeline.
  4. [§4.2 Eq. (6)] The Decoupling Score ρ_DS is introduced as a core metric for causal validity, but its definition is unclear as printed. Eq. (6) contains 'P earson(SRi, di)' (likely a typo for Pearson correlation) and the weight formula 'wi = µiP µi' does not parse. It is also not specified whether the correlation is computed across time, across store-product pairs, or both, nor why a near-zero correlation is the desired property for recovered demand. Since ρ_DS is used to argue that TimesNet 'eliminates spurious demand and stockout linkages,' please define the metric precisely and justify its interpretation.
minor comments (6)
  1. [Abstract] The GitHub link in the abstract has a stray closing brace: 'https://github.com/Dingdong-Inc/frn-50k-baseline})' should be 'https://github.com/Dingdong-Inc/frn-50k-baseline'.
  2. [§3.2] The power-law claim reports α = 2.83 with a Kolmogorov-Smirnov p < 0.03; a p-value this low is usually interpreted as evidence against the power-law null rather than support for it. Please clarify the test setup and interpret the result accordingly.
  3. [§4.1 Eq. (1)] The notation for s_t is confusing: s_t = 1 denotes stock availability, but the text calls s a 'censoring indicator.' Since Eq. (2) uses (1-s) to select censored periods, please rename s to an availability indicator or define the censoring mask explicitly to avoid ambiguity.
  4. [§5.2, Table 2] Table 2 reports 'Raw Sale' only in the row label with no WAPE or WPE values; providing the raw-sales baseline would make the improvement attributable to recovery methods easier to assess.
  5. [Figure 3] The caption for Figure 3, 'Demand Concentration and Long-Tail Distribution,' does not explain what is plotted (e.g., rank-frequency plot, Lorenz curve, or histogram); please make the axes and quantity explicit.
  6. [Conclusion] There is a typo in the final paragraph: 'perishabble' should be 'perishable.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the recovery and forecasting claims rest on empirical evaluations, not on definitions that re-import their targets.

full rationale

No load-bearing step reduces to its own inputs. Equations (1)-(3) define the stockout indicator from inventory records and the recovery target from observed sales plus a model estimate; these are modeling definitions, not predictions smuggled from the data. The synthetic MNAR evaluation in Section 5.1 uses non-stockout periods as ground truth and censoring masks 'matching empirical patterns'; this is a limitation in external validity because real stockout recovery is never directly measured, but it does not make the reported WAPE/WPE results equivalent to the inputs by construction. The Decoupling Score (Eq. 6) is an evaluation metric relating recovered demand to stockout ratios; the raw-sales negative value is partly definitional because sales are zero during stockouts, but the paper does not use rho_DS as a derivation of the recovered demand itself. Forecasting is assessed on stockout-free periods, so the evaluation labels there are observed sales; the improvement from using recovered-demand training data is an empirical outcome, not a tautology. The paper contains essentially no self-citation chain: the cited baselines and missing-data references are external, and no uniqueness theorem is imported from the authors' prior work. The strongest weakness is the unverified quality of the stockout labels, but that is a data-quality and correctness concern, not a circularity of the kind this pass is asked to flag.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central experimental claims rest on domain assumptions about data quality and on an evaluation design (synthetic censoring) that is not fully specified. No invented entities are introduced.

free parameters (1)
  • Empirical stockout mask statistics = not fully reported
    Section 5.1 creates synthetic censored regions 'matching empirical patterns'; the diurnal U-shaped stockout curve (below 2% to 26%), promotion and weather coefficients are fitted from the same dataset and used to make evaluation masks, but the procedure is not specified.
assumptions (4)
  • domain assumption Hourly stock level records from the warehouse management system accurately reflect product availability and are correctly annotated as stockouts.
    Invoked in Sections 3.1 and 4.1 (Equation 1 defines censoring indicator s_t based on inventory level i_t > 0). No audit of label quality is provided.
  • domain assumption During non-stockout hours, observed sales equal true demand, with no unobserved lost sales.
    Section 5.1 states evaluation is conducted only during operational periods without stockouts, and raw sales are treated as ground truth for the synthetic MNAR simulation.
  • ad hoc to paper Synthetic censored regions generated from non-stockout periods are representative of real MNAR censoring.
    Section 5.1 says masks are generated 'matching empirical patterns' but the exact generative mechanism, parameters, and validation are not described, making the simulation a custom evaluation assumption.
  • domain assumption Future covariates (promotions, weather, calendar) are known over the 7-day forecast horizon.
    Section 4.2 assumes P, W, C are known for T+1 to T+7 because marketing plans, weather forecasts, and holiday schedules are available in advance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FreshRetailNet-50K: A Stockout-Annotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail." pith.science (2026). https://pith.science/paper/SSNSLVIE

@misc{pith2026250516319,
  author       = {Pith},
  title        = {Pith review of: FreshRetailNet-50K: A Stockout-Annotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SSNSLVIE}},
  note         = {Machine review of arXiv:2505.16319}
}
read the original abstract

Accurate demand estimation is critical for the retail business in guiding the inventory and pricing policies of perishable products. However, it faces fundamental challenges from censored sales data during stockouts, where unobserved demand creates systemic policy biases. Existing datasets lack the temporal resolution and annotations needed to address this censoring effect. To fill this gap, we present FreshRetailNet-50K, the first large-scale benchmark for censored demand estimation. It comprises 50,000 store-product time series of detailed hourly sales data from 898 stores in 18 major cities, encompassing 863 perishable SKUs meticulously annotated for stockout events. The hourly stock status records unique to this dataset, combined with rich contextual covariates, including promotional discounts, precipitation, and temporal features, enable innovative research beyond existing solutions. We demonstrate one such use case of two-stage demand modeling: first, we reconstruct the latent demand during stockouts using precise hourly annotations. We then leverage the recovered demand to train robust demand forecasting models in the second stage. Experimental results show that this approach achieves a 2.73% improvement in prediction accuracy while reducing the systematic demand underestimation from 7.37% to near-zero bias. With unprecedented temporal granularity and comprehensive real-world information, FreshRetailNet-50K opens new research directions in demand imputation, perishable inventory optimization, and causal retail analytics. The unique annotation quality and scale of the dataset address long-standing limitations in retail AI, providing immediate solutions and a platform for future methodological innovation. The data (https://huggingface.co/datasets/Dingdong-Inc/FreshRetailNet-50K) and code (https://github.com/Dingdong-Inc/frn-50k-baseline}) are openly released.

Figures

Figures reproduced from arXiv: 2505.16319 by the authors.

Figure 1
Figure 1. Various assortments of highly substitutable perishable goods. Each row presents multiple [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Temporal Demand Patterns in FreshRetailNet-50K [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. WPE Over Time by Different Demand Recovery Method On Low Sale 6 Limitation FreshRetailNet-50K provides valuable insights into perishable retail forecasting but should be contextualized with considerations. The temporal scope of the dataset may not fully capture seasonal trends or long-term shifts in consumer behavior and supply chain dynamics relevant to fresh produce retail. Furthermore, the sparse and volatile dat… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SCOPE: Supply-Chain Operations through Coupled Policies for End-to-End Coordination

    cs.AI 2026-07 conditional novelty 6.0 of 10

    A tokenized multi-interface policy that coordinates assortment, sourcing, replenishment frequency, and routing outperforms decomposed industrial baselines on Dingdong and JD.com fresh-retail data under a shared proxy utility.

Reference graph

Works this paper leans on

44 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    Walmart recruit- ing - store sales forecasting

    Walmart Competition Admin and Will Cukierski. Walmart recruit- ing - store sales forecasting. https://kaggle.com/competitions/ walmart-recruiting-store-sales-forecasting , 2014. Kaggle

  2. [2]

    Censored demand estimation in retail

    Muhammad J Amjad and Devavrat Shah. Censored demand estimation in retail. Proceedings of the ACM on Measurement and Analysis of Computing Systems, 1(2):1–28, 2017

  3. [3]

    Fore- casting with temporal hierarchies

    George Athanasopoulos, Rob J Hyndman, Nikolaos Kourentzes, and Fotios Petropoulos. Fore- casting with temporal hierarchies. European Journal of Operational Research, 262(1):60–74, 2017

  4. [4]

    Inventory management with partially observed nonstationary demand

    Erhan Bayraktar and Michael Ludkovski. Inventory management with partially observed nonstationary demand. Annals of Operations Research, 176(1):7–39, 2010

  5. [5]

    Kolmogorov–smirnov test: Overview

    Vance W Berger and YanYan Zhou. Kolmogorov–smirnov test: Overview. Wiley statsref: Statistics reference online, 2014

  6. [6]

    Recurrent neural networks for multivariate time series with missing values

    Zhengping Che, Sanjay Purushotham, Kyunghyun Cho, David Sontag, and Yan Liu. Recurrent neural networks for multivariate time series with missing values. Scientific reports, 8(1):6085, 2018

  7. [7]

    Power-law distributions in empirical data

    Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ Newman. Power-law distributions in empirical data. SIAM review, 51(4):661–703, 2009

  8. [8]

    SAITS: Self-Attention-based Imputation for Time Series

    Wenjie Du, David Cote, and Yan Liu. SAITS: Self-Attention-based Imputation for Time Series. Expert Systems with Applications, 219:119619, 2023

Show all 44 references
  1. [9]

    Dua and C

    D. Dua and C. Graff. Uci individual household electric power consumption dataset. https://archive.ics.uci.edu/ml/datasets/individual+household+electric+ power+consumption, publisher = UCI Machine Learning Repository, 2015

  2. [10]

    Corpo- ración favorita grocery sales forecasting

    Corporación Favorita, inversion, Julia Elliott, and Mark McDonald. Corpo- ración favorita grocery sales forecasting. https://kaggle.com/competitions/ favorita-grocery-sales-forecasting , 2017. Kaggle

  3. [11]

    Gp-vae: Deep probabilistic time series imputation

    Vincent Fortuin, Dmitry Baranchuk, Gunnar Rätsch, and Stephan Mandt. Gp-vae: Deep probabilistic time series imputation. In International conference on artificial intelligence and statistics, pages 1651–1661. PMLR, 2020

  4. [12]

    Offline dynamic inventory and pricing strategy: Addressing censored and dependent demand

    Korel Gundem and Zhengling Qi. Offline dynamic inventory and pricing strategy: Addressing censored and dependent demand. arXiv preprint arXiv:2504.09831, 2025

  5. [13]

    M5 forecasting - accuracy

    Addison Howard, inversion, Spyros Makridakis, and vangelis. M5 forecasting - accuracy. https://kaggle.com/competitions/m5-forecasting-accuracy , 2020. Kaggle

  6. [14]

    Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting

    Shiyang Li, Xiaoyong Jin, Yao Xuan, Xiyou Zhou, Wenhu Chen, Yu-Xiang Wang, and Xifeng Yan. Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting. Advances in neural information processing systems, 32, 2019

  7. [15]

    Temporal fusion transformers for interpretable multi-horizon time series forecasting

    Bryan Lim, Sercan Ö Arık, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4):1748–1764, 2021

  8. [16]

    Missing data assumptions

    Roderick J Little. Missing data assumptions. Annual Review of Statistics and Its Application, 8(1):89–107, 2021

  9. [17]

    Statistical analysis with missing data

    Roderick JA Little and Donald B Rubin. Statistical analysis with missing data. John Wiley & Sons, 2019

  10. [18]

    Pristi: A conditional diffusion framework for spatiotemporal imputation

    Mingzhe Liu, Han Huang, Hao Feng, Leilei Sun, Bowen Du, and Yanjie Fu. Pristi: A conditional diffusion framework for spatiotemporal imputation. In2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 1927–1939. IEEE, 2023. 11

  11. [19]

    Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting

    Shizhan Liu, Hang Yu, Cong Liao, Jianguo Li, Weiyao Lin, Alex X Liu, and Schahram Dustdar. Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. In International conference on learning representations, 2021

  12. [20]

    itransformer: Inverted transformers are effective for time series forecasting

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625, 2023

  13. [21]

    Multivariate time series imputation with generative adversarial networks

    Yonghong Luo, Xiangrui Cai, Ying Zhang, Jun Xu, et al. Multivariate time series imputation with generative adversarial networks. Advances in neural information processing systems, 31, 2018

  14. [22]

    M5 accuracy competi- tion: Results, findings, and conclusions

    Spyros Makridakis, Evangelos Spiliotis, and Vassilios Assimakopoulos. M5 accuracy competi- tion: Results, findings, and conclusions. International Journal of Forecasting, 38(4):1346–1364, 2022

  15. [23]

    Demand estimation from censored observations with inventory record inaccuracy

    Adam J Mersereau. Demand estimation from censored observations with inventory record inaccuracy. Manufacturing & Service Operations Management, 17(3):335–349, 2015

  16. [24]

    Reducing fresh fish waste while ensuring availability: Demand forecast using censored data and machine learning

    Vera Lucia Miguéis, André Pereira, Joao Pereira, and Gonçalo Figueira. Reducing fresh fish waste while ensuring availability: Demand forecast using censored data and machine learning. Journal of cleaner Production, 359:131852, 2022

  17. [25]

    Continuous inventory control with stochastic and non- stationary markovian demand

    Walid W Nasr and Ibrahim J Elshar. Continuous inventory control with stochastic and non- stationary markovian demand. European Journal of Operational Research, 270(1):198–217, 2018

  18. [26]

    Imputeformer: Low rankness- induced transformers for generalizable spatiotemporal imputation

    Tong Nie, Guoyang Qin, Wei Ma, Yuewen Mei, and Jian Sun. Imputeformer: Low rankness- induced transformers for generalizable spatiotemporal imputation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2260–2271, 2024

  19. [27]

    Brazilian e-commerce public dataset by olist

    Olist and André Sionek. Brazilian e-commerce public dataset by olist. https://www.kaggle. com/dsv/195341, 2018

  20. [28]

    Pedregal and Juan R

    Diego J. Pedregal and Juan R. Trapero. Censored data forecasting: Applying tobit exponential smoothing with time aggregation, 2024

  21. [29]

    A cross-temporal hierarchical frame- work and deep learning for supply chain forecasting

    Sushil Punia, Surya P Singh, and Jitendra K Madaan. A cross-temporal hierarchical frame- work and deep learning for supply chain forecasting. Computers & Industrial Engineering, 149:106796, 2020

  22. [30]

    Coherent probabilistic forecasting of temporal hierarchies

    Syama Sundar Rangapuram, Shubham Kapoor, Rajbir Singh Nirwan, Pedro Mercado, Tim Januschowski, Yuyang Wang, and Michael Bohlke-Schneider. Coherent probabilistic forecasting of temporal hierarchies. In International Conference on Artificial Intelligence and Statistics, pages 93...

  23. [31]

    Inference and missing data

    Donald B Rubin. Inference and missing data. Biometrika, 63(3):581–592, 1976

  24. [32]

    Arima models

    Robert H Shumway, David S Stoffer, Robert H Shumway, and David S Stoffer. Arima models. Time series analysis and its applications: with R examples, pages 75–163, 2017

  25. [33]

    Predicting demand for new products in fashion retailing using censored data

    MS Sousa, ALD Loureiro, and VL Miguéis. Predicting demand for new products in fashion retailing using censored data. Expert Systems with Applications, 259:125313, 2025

  26. [34]

    Csdi: Conditional score-based diffusion models for probabilistic time series imputation

    Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. 2021

  27. [35]

    A behavioral remedy for the censorship bias

    Jordan Tong, Daniel Feiler, and Richard Larrick. A behavioral remedy for the censorship bias. Production and Operations Management, 27(4):624–643, 2018

  28. [36]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 12

  29. [37]

    Gaussian processes for regression

    Christopher Williams and Carl Rasmussen. Gaussian processes for regression. Advances in neural information processing systems, 8, 1995

  30. [38]

    Times- net: Temporal 2d-variation modeling for general time series analysis

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Times- net: Temporal 2d-variation modeling for general time series analysis. arXiv preprint arXiv:2210.02186, 2022

  31. [39]

    Autoformer: Decomposition trans- formers with auto-correlation for long-term series forecasting

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition trans- formers with auto-correlation for long-term series forecasting. Advances in neural information processing systems, 34:22419–22430, 2021

  32. [40]

    Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023

  33. [41]

    Score- cdm: Score-weighted convolutional diffusion model for multivariate time series imputation

    Shunyang Zhang, Senzhang Wang, Hao Miao, Hao Chen, Changjun Fan, and Jian Zhang. Score- cdm: Score-weighted convolutional diffusion model for multivariate time series imputation. arXiv preprint arXiv:2405.13075, 2024

  34. [42]

    sasdim: self-adaptive noise scaling diffusion model for spatial time series imputation

    Shunyang Zhang, Senzhang Wang, Xianzhen Tan, Ruochen Liu, Jian Zhang, and Jianxin Wang. sasdim: self-adaptive noise scaling diffusion model for spatial time series imputation. arXiv preprint arXiv:2309.01988, 2023

  35. [43]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11106–11115, 2021

  36. [44]

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International conference on machine learning, pages 27268–27286. PMLR, 2022. 13

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.