pith. sign in

arxiv: 1907.01960 · v1 · pith:VPT3OVEFnew · submitted 2019-06-27 · 💻 cs.OH

Fashion Retail: Forecasting Demand for New Items

Pawan Kumar Singh , Yadunath Gupta , Nilpa Jha , Aruna Rajan This is my paper

Pith reviewed 2026-05-25 14:14 UTC · model grok-4.3

classification 💻 cs.OH
keywords fashion demand forecastingnew item predictionattribute-based modelsneural networksretail sales datageneralized forecastinginventory planning
0
0 comments X

The pith

Demand for new fashion items can be forecasted from their attributes using models trained on historical sales.

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

The paper studies large-scale fashion sales records to determine which attributes of clothing and footwear most strongly influence demand. It then trains models that take only those attributes as input to predict demand for entirely new designs that have no prior sales history. This matters because fashion trends shift quickly, bulk production requires advance commitment, and unsold inventory creates large losses. The authors test the approach across multiple neural network designs, machine learning algorithms, and loss functions to show consistent results. A sympathetic reader would see this as a way to move beyond item-specific historical forecasting to attribute-driven prediction for unseen styles.

Core claim

By analyzing historical sales data the authors extract the clothing and footwear attributes and merchandising factors that drove past demand, then construct generalized models that forecast demand for new items solely from those attributes; the models maintain robust performance when different neural architectures, machine learning methods, and loss functions are substituted.

What carries the argument

Generalized forecasting models that map new-item attributes to predicted demand, trained on historical sales records of existing items.

If this is right

  • Retailers could commit to production quantities for new designs before any sales data exist for those specific items.
  • Forecasting can be performed at the level of abstracted attributes rather than individual stock-keeping units.
  • The same modeling pipeline works across varied neural architectures, standard machine learning algorithms, and different loss functions.
  • Inventory risk from overproduction or underproduction of transient fashion items can be reduced by attribute-based planning.

Where Pith is reading between the lines

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

  • The method might be tested by holding out entire seasonal collections or color palettes to check whether attribute signals remain stable across trend cycles.
  • Retailers could combine the attribute models with short-term social-media signals to adjust forecasts closer to launch.
  • If attribute importance rankings prove stable, they could guide design teams on which features to emphasize in new collections.

Load-bearing premise

The attributes and factors that drove demand for past items will also drive demand for completely new designs and styles that never appeared in the training data.

What would settle it

Train the models on one set of items and test accuracy on a separate collection of new styles with no shared attributes or designs; a sharp drop in predictive accuracy on the new styles would falsify the generalization claim.

Figures

Figures reproduced from arXiv: 1907.01960 by Aruna Rajan, Nilpa Jha, Pawan Kumar Singh, Yadunath Gupta.

Figure 1
Figure 1. Figure 1: DNN Model Architectures In the data, we see long tail behaviour that is typical character￾istic of retail, with fewer items contributing to a majority of the sales. Due to this, we see variation of sales over several orders of magnitude. To address this high variance problem, we train our models at different scales - log and linear, and try a different set of loss functions. See [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 2
Figure 2. Figure 2: Salient Features of Data: Sales have Poisson Distribution in linear scale (a) and Normal Distribution in log scale (b). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of derived features on forecast 4.4 Deployment in industrial setting We have tested and deployed our models for the following fashion retail use cases at Myntra-Jabong. We also talk about futuristic scenarios where we are working to deploy our models. • Seasonal assortment Planning: Fashion Retailers have to plan their assortment a year in advance due to manufacturing lead times. At the time planner… view at source ↗
Figure 4
Figure 4. Figure 4: Actual vs. Forecasted: (a) wMAPE=0.34, is an example of good forecast, (b) wMAPE = 0.37, is an example of good [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

Fashion merchandising is one of the most complicated problems in forecasting, given the transient nature of trends in colours, prints, cuts, patterns, and materials in fashion, the economies of scale achievable only in bulk production, as well as geographical variations in consumption. Retailers that serve a large customer base spend a lot of money and resources to stay prepared for meeting changing fashion demands, and incur huge losses in unsold inventory and liquidation costs [2]. This problem has been addressed by analysts and statisticians as well as ML researchers in a conventional fashion - of building models that forecast for future demand given a particular item of fashion with historical data on its sales. To our knowledge, none of these models have generalized well to predict future demand at an abstracted level for a new design/style of fashion article. To address this problem, we present a study of large scale fashion sales data and directly infer which clothing/footwear attributes and merchandising factors drove demand for those items. We then build generalised models to forecast demand given new item attributes, and demonstrate robust performance by experimenting with different neural architectures, ML methods, and loss functions.

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

2 major / 0 minor

Summary. The paper claims that by analyzing large-scale fashion sales data, one can infer which clothing/footwear attributes and merchandising factors drive demand, then build generalized models (via neural architectures, ML methods, and loss functions) that forecast demand for entirely new item designs/styles, achieving robust performance.

Significance. If the generalization result holds with proper validation, the work would be significant for fashion retail by enabling demand forecasts for novel designs without historical sales data, potentially reducing unsold inventory and liquidation costs in a high-variability domain.

major comments (2)
  1. [Abstract] Abstract: The claim of demonstrating 'robust performance' by experimenting with different neural architectures, ML methods, and loss functions is unsupported by any quantitative metrics, validation details, dataset descriptions, or error analysis, preventing evaluation of whether the models actually generalize to new items.
  2. [Abstract] Abstract: The central OOD generalization claim requires that demand drivers identified from existing items apply to new designs/styles never seen in training. No information is given on attribute vocabulary size, whether new styles introduce unseen attribute values, or whether the train/test division is temporal (future seasons) versus random, so the reported performance cannot be interpreted as evidence for the generalization step.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments on our abstract. We agree that additional details would strengthen the abstract and will revise it accordingly. Below we address the specific points.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim of demonstrating 'robust performance' by experimenting with different neural architectures, ML methods, and loss functions is unsupported by any quantitative metrics, validation details, dataset descriptions, or error analysis, preventing evaluation of whether the models actually generalize to new items.

    Authors: The abstract is intended as a concise overview. The full paper provides quantitative results, including performance metrics for various models, validation procedures, dataset descriptions, and error analyses in the dedicated Experiments and Results sections. To address this concern and allow readers to better evaluate the claims from the abstract alone, we will revise the abstract to include key quantitative metrics and a brief mention of the validation approach. revision: yes

  2. Referee: [Abstract] Abstract: The central OOD generalization claim requires that demand drivers identified from existing items apply to new designs/styles never seen in training. No information is given on attribute vocabulary size, whether new styles introduce unseen attribute values, or whether the train/test division is temporal (future seasons) versus random, so the reported performance cannot be interpreted as evidence for the generalization step.

    Authors: We agree that the abstract lacks these specifics. The manuscript details the attribute vocabulary, confirms that the model handles new combinations of attributes, and uses a temporal train/test split to ensure forecasting for future unseen items. We will update the abstract to briefly describe the temporal validation and attribute-based generalization to clarify the OOD aspect. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML modeling with no derivations or self-referential steps

full rationale

The paper presents an empirical ML study that infers demand drivers from historical sales data on existing items and trains models to predict demand for new items given their attributes. No equations, derivations, fitted parameters renamed as predictions, or self-citations appear in the provided text. The central claim rests on experimental results across neural architectures, ML methods, and loss functions rather than any deductive chain. This is a standard data-driven generalization task whose validity is assessed externally via held-out performance, making the work self-contained with no load-bearing reductions to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities are described.

pith-pipeline@v0.9.0 · 5728 in / 1029 out tokens · 57051 ms · 2026-05-25T14:14:05.988682+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

19 extracted references · 19 canonical work pages · 4 internal anchors

  1. [1]

    [n. d.]. Autoregressive integrated moving average (ARIMA). https://en.wikipedia. org/wiki/Autoregressive_integrated_moving_average. Accessed: 2019-05-02

    work page 2019

  2. [2]

    [n. d.]. H&M, a Fashion Giant, Has a Problem: $4.3 Billion in Unsold Clothes. https://www.nytimes.com/2018/03/27/business/hm-clothes-stock-sales.html. Ac- cessed: 2019-05-02

    work page 2018

  3. [3]

    [n. d.]. One Hot Encoding. https://scikit-learn.org/stable/modules/generated/ sklearn.preprocessing.OneHotEncoder.html. Accessed: 2019-05-02

    work page 2019

  4. [4]

    James Bergstra, Daniel Yamins, and David Daniel Cox. 2013. Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures. (2013)

    work page 2013

  5. [5]

    Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’16) . ACM, New York, NY, USA, 785–794. https://doi.org/10.1145/2939672.2939785

    work page doi:10.1145/2939672.2939785 2016

  6. [6]

    Anna Veronika Dorogush, Vasily Ershov, and Andrey Gulin. 2018. CatBoost: gra- dient boosting with categorical features support. arXiv preprint arXiv:1810.11363 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  7. [7]

    Valentin Flunkert, David Salinas, and Jan Gasthaus. 2017. DeepAR: Proba- bilistic forecasting with autoregressive recurrent networks. arXiv preprint arXiv:1704.04110 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  8. [8]

    Cheng Guo and Felix Berkhahn. 2016. Entity embeddings of categorical variables. arXiv preprint arXiv:1604.06737 (2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  9. [9]

    Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. 2012. Improving neural networks by preventing co- adaptation of feature detectors. arXiv preprint arXiv:1207.0580 (2012)

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  10. [10]

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. LightGBM: A Highly Efficient Gradient Boosting Decision Tree. In NIPS

    work page 2017

  11. [11]

    Ellen C. Mik. 2019. New Product Demand Forecasting, A Literature Study . Master’s thesis. Vrije Universitat, Amsterdam. (In preparation)

    work page 2019

  12. [12]

    Maria Elena Nenni, Luca Giustiniano, and Luca Pirolo. 2013. Demand forecasting in the fashion industry: a review. International Journal of Engineering Business Management 5 (2013), 37

    work page 2013

  13. [13]

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer

    work page

  14. [14]

    In NIPS-W

    Automatic differentiation in PyTorch. In NIPS-W

    work page

  15. [15]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour- napeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine KDD 2019 Workshop, August 2019, Anchorage, Alaska - USA Pawan Kumar Singh, Yadunath Gupta, Nilpa Jha, and Aruna Rajan L...

    work page 2011

  16. [16]

    Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. 2018. How does batch normalization help optimization?. In Advances in Neural Infor- mation Processing Systems. 2483–2493

    work page 2018

  17. [17]

    Leslie N Smith. 2017. Cyclical learning rates for training neural networks. In 2017 IEEE Winter Conference on Applications of Computer Vision (W ACV). IEEE, 464–472

    work page 2017

  18. [18]

    Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to Sequence Learning with Neural Networks. In Advances in Neural Information Processing Systems 27, Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger (Eds.). Curran Associates, Inc., 3104–3112. http://papers.nips.cc/ paper/5346-sequence-to-sequence-learning-with-neural-n...

    work page 2014

  19. [19]

    Sébastien Thomassey and Antonio Fiordaliso. 2006. A hybrid sales forecasting system based on clustering and decision trees. Decision Support Systems 42, 1 (2006), 408–421. A APPENDIX We list down results on some more article types for different types of models/loss functions used, and find that XGBoost with an MSE loss function consistently outperforms ot...

    work page 2006