Pith. sign in

REVIEW 3 major objections 4 minor 34 references

OTPTO: Joint Product Selection and Inventory Optimization in Fresh E-commerce Front-End Warehouses

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Jointly learning which products to stock and in what quantity, rather than forecasting sales first, raises fully fulfilled orders from 61.57% to 65.91% on JD.com's 7Fresh platform.

desk verdict A cleanly presented optimize-then-predict inventory pipeline whose headline 4.34pp gain is undermined by an internal baseline inconsistency in Algorithm 2. read the letter →

arxiv 2505.23421 v1 pith:NKOO3MWT submitted 2025-05-29 cs.LG

Zheming Zhang , Yan Jiang , Qingshan Li , Ai Han This is my paper
classification cs.LG
keywords fullorderfulfillmentratefront-endwarehouseinventoryoptimizationproductselectionoptimize-then-predict-then-optimizepredict-then-optimizemixedintegerprogrammingLightGBM
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

The paper claims that fresh e-commerce front-end warehouses should be stocked by learning the decisions that would have been optimal in hindsight, rather than by forecasting sales and then converting the forecast into inventory. Its OTPTO method first solves a mixed-integer program (OM1) over historical orders to obtain the product set and quantities that maximize the full order fulfillment rate, then trains two predictors on those optimal decisions, and finally applies a constraint-preserving post-processor to the predictions. On a seven-day test at JD.com's 7Fresh, the method raises the share of orders delivered in a single shipment from 61.57% to 65.91%, shrinking the gap to the hindsight-optimal rate by 5.27 percentage points. The reported gains persist across five other warehouses, and removing the sales-prediction features drops performance by 8.40 percentage points, showing that OTPTO builds on a strong existing forecaster.

What carries the argument

OM1 is the engine of the pipeline: a 0-1 mixed integer linear program that, given the arrival-ordered orders of one day, selects stocking variables $y_i$ and $x_i$ and order-fulfillment variables $z_{oi}$ and $p_o$, maximizing the average fraction of orders fully fulfilled subject to capacity limits on SKU types, total quantity, and minimum stock per SKU. Its optimal solutions become the labels and decision features that PM1 and PM2 learn, which is what aligns prediction with the operational objective. PM1 and PM2 are LightGBM models, one binary classifier for product selection and one regressor for inventory quantity, trained on OM1 outputs and a set of decision, sales-prediction, clustering, and SKU-order cross-features. OM2 is a greedy post-processing algorithm that re-imposes the three capacity constraints on the predicted plan before it is executed.

What would settle it

Simulate one week of OTPTO's posted stocking plans against the actual order stream while allowing a realistic deviation that OM1 forbids, such as a mid-day restock of a top-selling SKU or substitution of a missing SKU with an equivalent product, and compare the realized full order fulfillment rate with PTO's; if the OTPTO margin disappears or reverses, the OM1 assumption set is the failing link.

Watch

Extended reading notes

Core claim

The central claim is that the best way to stock a small front-end warehouse is to treat the inventory decision itself as the prediction target. For each historical day, OM1, a 0-1 mixed integer program, chooses which SKUs to stock ($y_i$) and in what quantity ($x_i$) so as to maximize the number of orders that can be fully fulfilled by that warehouse alone, given limits on SKU types, total units, and minimum per-SKU stock. Two LightGBM models, PM1 and PM2, then learn to reproduce those optimal product-selection and stocking-quantity decisions from features that include sales forecasts, historical decision statistics, SKU clusters, and order-basket cross-features. A greedy post-processor OM2 maps the predictions back onto the feasibility constraints. In the test week, the pipeline reaches 65.91% full order fulfillment on average versus 61.57% for predict-then-optimize, cutting the gap to the hindsight-optimal 82.41% by 5.27 percentage points, and the same qualitative advantage shows up on five other warehouses.

Load-bearing premise

The load-bearing premise is that the OM1 formulation (fixed daily stock, arrival-order handling, no mid-day restocking, no substitution, and no customer-acceptable split delivery) matches how the warehouse really fulfills orders; if the real system deviates, both the optimal labels and the OPT benchmark are biased, and the measured 4.34-point gain may not show up in actual operations.

Editorial extensions

If this is right

  • Warehouse managers can move from sales-forecast-then-stock to decision learning, because the prediction targets are the inventory decisions themselves, so the training loss matches the operational goal.
  • A test week gain of 4.34 percentage points in full order fulfillment means fewer split shipments for customers and lower last-mile delivery cost per order.
  • The GMV tie-breaking and label-smoothing steps handle the fact that OM1 has many equally good optimal solutions, making the training labels stable enough for supervised learning.
  • OTPTO's plans include more low-selling SKUs alongside best-sellers, indicating that order-basket complementarity, not sales volume alone, drives full order fulfillment.
  • Because the sales-prediction features contribute the largest ablation gain (8.40 percentage points), the method is not a replacement for demand forecasting but a better consumer of it.

Reading between the lines

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

  • The measured gain likely comes from modeling which SKUs co-occur in orders rather than from better demand forecasts, since the OM1 objective rewards stocking the combination that completes more baskets.
  • A natural next test is an A/B deployment that compares actual split-shipment rates between an OTPTO-managed warehouse and a PTO-managed one, which would tell whether the 5.27-point gap reduction to the hindsight optimum transfers to field operations.
  • The same optimize-then-predict-then-optimize pattern should apply to other capacity-constrained retail decisions, such as limited shelf space in convenience stores or assortment planning under display constraints, whenever historically optimal decisions can be computed offline.
  • The 82.41% hindsight-optimal ceiling implies roughly 16.5 percentage points of fulfillment are lost even with perfect lookahead under OM1's assumptions, so relaxing those assumptions (allowing substitution or mid-day restocking) may be where the next large gain lies.
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

3 major / 4 minor

Summary. This paper proposes OTPTO, a three-phase approach for joint product selection and inventory quantity decisions in fresh e-commerce front-end warehouses. The first phase solves a 0-1 mixed-integer program (OM1) on historical order data to produce optimal stocking labels and an OPT benchmark. The prediction phase trains two LightGBM models, PM1 for product selection and PM2 for stocking quantities, using specialized sample, label, and feature strategies. The final phase applies a post-processing algorithm (OM2) to enforce capacity constraints. The method is evaluated on one test week from JD 7Fresh, reporting a 4.34 percentage-point improvement in full-order fulfillment over a greedy PTO baseline, together with an ablation study and a robustness plot across five other warehouses.

Significance. If the reported comparison is correct, the paper makes a useful applied contribution by offering a concrete way to align prediction with an order-fulfillment objective and by demonstrating the approach on real e-commerce data. The strengths are the clearly specified MIP, the out-of-sample evaluation relative to the training period, real data from six warehouses, and a detailed ablation of sample, label, and feature strategies. The main unresolved issue is that the PTO baseline's definition in Section 5 conflicts with the implementation shown in Algorithm 2, so the headline 4.34 percentage-point gain may not support the claimed joint selection-and-inventory improvement over a genuine sales-driven PTO. The single-week evaluation without error bars also tempers the strength of the empirical conclusion.

major comments (3)
  1. [Section 5 and Algorithm 2 (Appendix A.1)] The PTO baseline is described in Section 5 as first sorting SKUs in descending order of predicted sales volume, but the PTO branch of Algorithm 2 (lines 15-20) sorts SKUs in descending order of \hat{y}_{ti}, the PM1 product-selection probability. If the implementation follows Algorithm 2, then OTPTO and PTO use the same PM1-based selection, and the 4.34 percentage-point difference shown in Table 2 isolates only the quantity model (PM2 with clipping) rather than a joint selection-and-inventory comparison against a sales-driven PTO. Please implement the baseline exactly as defined in the text, rerun the experiment, and report the resulting difference, or clearly state that the PTO baseline uses PM1 for selection.
  2. [Section 4.2, Eq. (19)] Equation (19) is an equal-weight linear combination of the full-order fulfillment rate and the GMV share of fulfilled orders, not a lexicographic maximization of the fulfillment rate. Since the GMV term is normalized by total GMV, a solution with one fewer fully fulfilled order can tie with (or in edge cases outrank) a fulfillment-maximizing solution when its fulfilled GMV share is much higher. Because Equation (19) generates both the OM1 training labels for PM1/PM2 and the OPT column in Table 2, the labels and the reported 'optimal' fulfillment rate may not correspond to the stated primary objective. Please solve the problem in two stages (maximize fulfillment rate first, then GMV among optimal solutions) or demonstrate that the weighted objective cannot change the optimal fulfillment rate on the actual data.
  3. [Section 5.1, Table 2; Section 5.3, Figure 9] The headline empirical claim is based on a single seven-day test window (2023-09-01 to 2023-09-07) with no confidence intervals, standard errors, or significance tests. The daily improvements are consistent in sign and magnitude (3.56 to 5.50 percentage points), which is encouraging, but one week is a narrow basis for the statement that OTPTO 'significantly' outperforms PTO. The robustness analysis for the other five warehouses is presented only as a plot; please report the numerical gap values (and ideally per-warehouse tables) and, if feasible, add a multi-week or rolling-window evaluation.
minor comments (4)
  1. [Section 4.1, Eq. (6)] Constraint (6) caps inventory by the realized sales volume d_i, which makes OM1 a hindsight optimization; this is appropriate for label generation, but the manuscript should state explicitly that both the training labels and the OPT benchmark are conditional on the OM1 model and realized demand, not on a forward-looking feasible policy.
  2. [Algorithm 2] The rounding and capacity loop in lines 23-32 can leave the total inventory below N and may drop SKUs that were selected by the top-K rule; please specify the exact tie-breaking, rounding, and capacity-reallocation rules so the post-processing is fully reproducible.
  3. [Table 3] Some ablation removals improve performance on individual days (e.g., A5 on 2023-09-01: 73.09% vs. 72.11%; A6 on 2023-09-05 and 09-07), so the claim that each strategy contributes should be qualified with per-day variability or a significance test.
  4. [Overall manuscript] The manuscript contains several presentation issues, including garbled text in figures (e.g., the label sequences in Figures 3-7 as rendered in the preprint) and minor typos (e.g., 'optimizin g' in the Abstract); these should be cleaned before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the out-of-sample OTPTO evaluation is not reduced to its training labels, though the PTO baseline pseudocode is internally inconsistent.

full rationale

The OTPTO derivation chain is not circular. OM1 is a 0-1 mixed-integer program solved independently on historical order data; its solutions are used only to construct labels and features for PM1 and PM2. The reported full order fulfillment rates are then computed on the held-out week 2023-09-01 to 2023-09-07, so the central 4.34-percentage-point improvement over PTO is an out-of-sample result rather than a fitted value renamed as a prediction. The OPT benchmark is obtained by re-solving the same OM1 on the test week's actual orders, which is a consistent simulation upper bound and not a piece of training data; using the same order-fulfillment model to generate labels and to evaluate both methods does not make the comparison true by construction. There is no load-bearing self-citation: the methodological inspiration is the external reference Qi et al. [7], and LightGBM [33,34] is a standard external tool. The most notable issue is an internal inconsistency, not circularity: Section 5 states that PTO 'first sort[s] the SKUs in descending order according to the predicted sales volume,' but Algorithm 2 in Appendix A.1 implements the PTO branch by sorting SKUs by \hat{y}_{ti}, the PM1 product-selection probability, and Table 5 reports different SKU counts for OTPTO (about 347) and PTO (about 299), which suggests the pseudocode does not match the executed baseline. If the code follows Algorithm 2, the reported comparison is against a mis-specified baseline rather than a genuine predict-then-optimize method; this is a reproducibility and external-validity concern, not a circular reduction of the measured improvement to the method's own inputs. Similarly, OM1's objective (19) is an equal-weighted sum of the fulfillment rate and GMV share rather than a lexicographic maximization of the stated primary objective, which may bias the generated labels, but again this is a modeling-fidelity risk rather than circular reasoning. Overall, no step in the claimed derivation reduces by definition or by self-citation to its own inputs, so the circularity score is 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the OM1 model's fidelity and on the learnability of its solutions from the available features; the hyperparameters (lambda, mu, gamma, rho) are fitted to the training data and would require sensitivity analysis to establish robustness.

free parameters (7)
  • label clustering lambda = 80
    Number of K-Means clusters used to group SKUs for cross-sectional label smoothing (Algorithm 1, Part 1, line 22); set by search, affects which SKUs get y* set to 1.
  • cross-section smoothing threshold mu = 0.8
    If more than mu fraction of a cluster is stocked, all SKUs in the cluster are labeled stocked (Algorithm 1, Part 1, line 33); tuned by search and directly alters training labels.
  • time-series smoothing threshold gamma = 0.8
    If a SKU is stocked on more than gamma of its active sales days in the training period, its y* is set to 1 for every day (Algorithm 1, Part 2, line 52); tuned by search.
  • feature clustering rho = 4
    Number of K-Means clusters for SKU stocking-frequency features in PM1/PM2 (Section 4.2.3); tuned by search.
  • big-M constant M = 1e5
    Used to linearize logical constraints in OM1 (Section 4.1); chosen large enough to dominate any feasible xi - c_oi.
  • small constant delta = 1e-3
    Used in OM1 to break ties at xi = c_oi (Section 4.1); chosen smaller than the unit gap of integer variables.
  • LightGBM hyperparameters (learning_rate, n_estimators, etc.) = see Table 4
    PM1 and PM2 rely on these settings, which were presumably selected by validation search; the paper provides no sensitivity analysis.
assumptions (5)
  • domain assumption The front-end warehouse inventory problem is a single-period newsvendor problem because fresh goods require daily restocking and leftover spoilage is not modeled (Section 2).
    The entire OM1/OTPTO pipeline is built on daily independent stocking decisions; if multi-day dynamics or spoilage costs matter, the objective and labels change.
  • domain assumption The mixed-integer program OM1 with constraints (2)-(16) exactly represents order fulfillment in a front-end warehouse.
    Used to generate all training labels and the OPT benchmark; if the model misses real-world fulfillment dynamics, the whole pipeline is affected.
  • domain assumption The existing sales forecasting model PM0 provides accurate sales predictions and is available at decision time.
    PM0 outputs are used both as features (Section 4.2.3) and as a cap in the post-processing quantity rule qtty=max(B,min(x_hat,q_hat)) (Algorithm 2, line 13); the ablation shows removing these features causes the largest drop.
  • domain assumption The feature set available before each day is sufficient to predict the OM1 optimal decisions.
    PM1/PM2 map features to y* and x*; if the OM1 decisions are driven by unobserved information (e.g., specific order sequences), the learned model cannot replicate them.
  • domain assumption The three-month training period is representative of the test week.
    The paper uses 2023-06-01 to 2023-08-31 for training and 2023-09-01 to 2023-09-07 for testing, with no discussion of seasonality or product lifecycle effects.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OTPTO: Joint Product Selection and Inventory Optimization in Fresh E-commerce Front-End Warehouses." pith.science (2026). https://pith.science/paper/NKOO3MWT

@misc{pith2026250523421,
  author       = {Pith},
  title        = {Pith review of: OTPTO: Joint Product Selection and Inventory Optimization in Fresh E-commerce Front-End Warehouses},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NKOO3MWT}},
  note         = {Machine review of arXiv:2505.23421}
}
read the original abstract

In China's competitive fresh e-commerce market, optimizing operational strategies, especially inventory management in front-end warehouses, is key to enhance customer satisfaction and to gain a competitive edge. Front-end warehouses are placed in residential areas to ensure the timely delivery of fresh goods and are usually in small size. This brings the challenge of deciding which goods to stock and in what quantities, taking into account capacity constraints. To address this issue, traditional predict-then-optimize (PTO) methods that predict sales and then decide on inventory often don't align prediction with inventory goals, as well as fail to prioritize consumer satisfaction. This paper proposes a multi-task Optimize-then-Predict-then-Optimize (OTPTO) approach that jointly optimizes product selection and inventory management, aiming to increase consumer satisfaction by maximizing the full order fulfillment rate. Our method employs a 0-1 mixed integer programming model OM1 to determine historically optimal inventory levels, and then uses a product selection model PM1 and the stocking model PM2 for prediction. The combined results are further refined through a post-processing algorithm OM2. Experimental results from JD.com's 7Fresh platform demonstrate the robustness and significant advantages of our OTPTO method. Compared to the PTO approach, our OTPTO method substantially enhances the full order fulfillment rate by 4.34% (a relative increase of 7.05%) and narrows the gap to the optimal full order fulfillment rate by 5.27%. These findings substantiate the efficacy of the OTPTO method in managing inventory at front-end warehouses of fresh e-commerce platforms and provide valuable insights for future research in this domain.

Figures

Figures reproduced from arXiv: 2505.23421 by the authors.

Figure 1
Figure 1. Overview of The Front-end Warehouse Model [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. OTPTO Pipeline 4.1 The First Optimization Phase In general, the stocking strategy actually implemented is often not optimal. Inspired by Qi et al [7], we construct a joint optimization model for product selection and inventory quantity based on historical transaction data to obtain the training labels and inventory-related features of the prediction phase, that is, the optimal stocking results. For the independent p… view at source ↗
Figure 3
Figure 3. OM1’s inventory decision of the same SKU across day [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Product selection results from OM1 on 2023-06-21 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Correlation between sale quantity and optimal inventory quantity across second categories Sales Prediction Features The results of the OM1 model show a strong correlation between the optimal inventory lev￾els x ⋆ i and the actual sales on the corresponding day, with th…
Figure 6
Figure 6. Figure 6: Distribution of Stocking Frequency SKU-Order Cross-Features Through the comparative analy￾sis of OM1 outcomes across various days, frequently stocked SKUs are typically: a) appear in a larger number of orders; b) related to orders that contains only a small number of S…
Figure 7
Figure 7. Figure 7: Sales and types in orders. The points in the graph re [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Product selection results of OTPTO and PTO on 2023- [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Gap between the OTPTO/PTO method and the optimal so [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

  1. [1]

    Comparative analys is of online fresh food shopping behavior during normal and covid-19 crisis periods

    Miaojia Lu, Ran Wang, and Peiyang Li. Comparative analys is of online fresh food shopping behavior during normal and covid-19 crisis periods. British F ood Journal, 124(3):968–986, 2022

    work page 2022

  2. [2]

    Has covid-19 accelerated the e-commerce of agricultural products? evidence from sal es data of e-stores in china

    Jianxin Guo, Songqing Jin, Jichun Zhao, Hongbiao Wang, a nd Fang Zhao. Has covid-19 accelerated the e-commerce of agricultural products? evidence from sal es data of e-stores in china. F ood Policy, 112:102377, 2022

    work page 2022

  3. [3]

    2023-2024 china fresh food e-commerce operation big data and development prospects research report

    iiMedia New Retail Industry Research Center. 2023-2024 china fresh food e-commerce operation big data and development prospects research report. Technical report, iiMedia Research, 2023

    work page 2023

  4. [4]

    Location sele ction of fresh e-commerce’s front warehouse under new retail model

    Fengjiao Wan, Jiahui Qin, and Xingyu Wang. Location sele ction of fresh e-commerce’s front warehouse under new retail model. In 2022 International Conference on Artificial Intelligence, Internet and Digital Economy (ICAID 2022), pages 831–839. Atlantis Press, 2022

    work page 2022

  5. [5]

    Introduction to inventory manage- ment

    Dinesh Shenoy, Roberto Rosas, Dinesh Shenoy, and Robert o Rosas. Introduction to inventory manage- ment. Problems & Solutions in Inventory Management , pages 3–11, 2018

    work page 2018

  6. [6]

    Elmachtoub, Paul Grigas, and Ambuj Tewari

    Othman El Balghiti, Adam N. Elmachtoub, Paul Grigas, and Ambuj Tewari. Generalization bounds in the predict-then-optimize framework. Mathematics of Operations Research, 48(4):2043–2065, NOV 2023

    work page 2023

  7. [7]

    A practical end-to-end inventory management model with deep learning

    Meng Qi, Y uanyuan Shi, Y ongzhi Qi, Chenxin Ma, Rong Y uan, Di Wu, and Zuo-Jun Shen. A practical end-to-end inventory management model with deep learning. Management Science, 69(2):759–773, 2023

    work page 2023

  8. [8]

    The multi -product newsvendor problem: Review and extensions

    Mengting Mu, Junlin Chen, Y u Yang, and Jian Guo. The multi -product newsvendor problem: Review and extensions. In 2019 6th International Conference on Behavioral, Economic and Socio-Cultural Comput- ing (BESC), pages 1–4. IEEE, 2019

    work page 2019

Show all 34 references
  1. [9]

    Optimal and heuristic solutions fo r the multi-item newsvendor problem with a single capacity constraint

    Steven J Erlebacher. Optimal and heuristic solutions fo r the multi-item newsvendor problem with a single capacity constraint. Production and Operations Management, 9(3):303–318, 2000

    2000

  2. [10]

    A multi-product newsvendor problem with budget and loss constraints

    Yanju Zhou, Xiaohong Chen, Xuanhua Xu, and Changjun Y u. A multi-product newsvendor problem with budget and loss constraints. International Journal of Information Technology & Decisio n Making, 14(05):1093–1110, 2015

    2015

  3. [11]

    Analysis of two substitute products newsvendor problem with a budget constraint

    Linlin Zhang, Guoqing Zhang, and Zhong Yao. Analysis of two substitute products newsvendor problem with a budget constraint. Computers & Industrial Engineering , 140:106235, 2020

    2020

  4. [12]

    The risk-avers e newsvendor problem with random capacity

    Meng Wu, Stuart X Zhu, and Ruud H Teunter. The risk-avers e newsvendor problem with random capacity. European Journal of Operational Research, 231(2):328–336, 2013

    2013

  5. [13]

    Optimal robust policy for feature-based newsvendor

    Luhao Zhang, Jincheng Yang, and Rui Gao. Optimal robust policy for feature-based newsvendor. Man- agement Science, 70(4):2315–2329, 2024

    2024

  6. [14]

    Constructing decision rules fo r multiproduct newsvendors: An integrated estimation-and-optimization framework

    Alba V Olivares-Nadal. Constructing decision rules fo r multiproduct newsvendors: An integrated estimation-and-optimization framework. European Journal of Operational Research, 315(3):1021–1037, 2024

    2024

  7. [15]

    Bilevel optimization for feature selection in the data-driven newsvendor problem

    Breno Serrano, Stefan Minner, Maximilian Schiffer, an d Thibaut Vidal. Bilevel optimization for feature selection in the data-driven newsvendor problem. European Journal of Operational Research, 315(2):703– 714, 2024

    2024

  8. [16]

    Chan, Vivek F

    Carri W. Chan, Vivek F. Farias, Nicholas Bambos, and Gab riel J. Escobar. Optimizing intensive care unit discharge decisions with patient readmissions. Operations Research, 60(6):1323–1341, Nov-Dec 2012

    2012

  9. [17]

    Timothy C. Y . Chan, Tim Craig, Taewoo Lee, and Michael B. Sharpe. Generalized inverse multiobjective optimization with application to cancer therapy. Operations Research, 62(3):680–695, May-Jun 2014

    2014

  10. [18]

    Business analytics for flexible r esource allocation under random emergencies

    Mallik Angalakudati, Siddharth Balwani, Jorge Calzad a, Bikram Chatterjee, Georgia Perakis, Nicolas Raad, and Joline Uichanco. Business analytics for flexible r esource allocation under random emergencies. Management Science, 60(6, SI):1552–1573, JUN 2014

    2014

  11. [19]

    In f ree float: Developing business analytics support for carsharing providers

    Sebastian Wagner, Tobias Brandt, and Dirk Neumann. In f ree float: Developing business analytics support for carsharing providers. Omega-International Journal of Management Science , 59(A, SI):4–14, MAR 2016

    2016

  12. [20]

    Mersereau, Andres Garro, Albe rte Dapena Mora, and Martin Novoa Vidal

    Jeremie Gallien, Adam J. Mersereau, Andres Garro, Albe rte Dapena Mora, and Martin Novoa Vidal. Initial shipment decisions for new products at zara. Operations Research, 63(2):269–286, Mar-Apr 2015. 13

    2015

  13. [21]

    A practical inventory control policy using operational statistics

    LH Liyanage and JG Shanthikumar. A practical inventory control policy using operational statistics. Operations Research Letters, 33(4):341–348, JUL 2005

    2005

  14. [22]

    Analytics for an online retailer: Demand forecasting and price optimization

    Kris Johnson Ferreira, Bin Hong Alex Lee, and David Simc hi-Levi. Analytics for an online retailer: Demand forecasting and price optimization. Manufacturing & Service Operations Management , 18(1, SI):69–88, WIN 2016

    2016

  15. [23]

    Zhang, Hengchen Dai, Lingxiu Dong, Fangfang Q i, Nannan Zhang, Xiaofei Liu, Zhongyi Liu, and Jiang Yang

    Dennis J. Zhang, Hengchen Dai, Lingxiu Dong, Fangfang Q i, Nannan Zhang, Xiaofei Liu, Zhongyi Liu, and Jiang Yang. How do price promotions affect customer behavior on retailing platforms? evidence from a large randomized experiment on alibaba. Production and Operations Managem...

    2018

  16. [24]

    Production planning with patterns: A problem from processed food manuf acturing

    Mili Mehrotra, Milind Dawande, Srinagesh Gavirneni, M ehmet Demirci, and Sridhar Tayur. Production planning with patterns: A problem from processed food manuf acturing. Operations Research, 59(2):267– 282, Mar-Apr 2011

    2011

  17. [25]

    Capacity m anagement in agricultural commodity processing and application in the palm industry

    Onur Boyabatli, Jason Nguyen, and Tong Wang. Capacity m anagement in agricultural commodity processing and application in the palm industry. Manufacturing & Service Operations Management , 19(4):551–567, FAL 2017

    2017

  18. [26]

    Schrotenboer, and Feng Chen

    Menglei Jia, Albert H. Schrotenboer, and Feng Chen. Sce nario predict-then-optimize for data-driven online inventory routing, 2024

    2024

  19. [27]

    Bo Tang and Elias B. Khalil. Pyepo: A pytorch-based end- to-end predict-then-optimize library for linear and integer programming, 2023

    2023

  20. [28]

    The impa ct of employee satisfaction on quality and profitability in high-contact service industries

    Rachel WY Yee, Andy CL Yeung, and TC Edwin Cheng. The impa ct of employee satisfaction on quality and profitability in high-contact service industries. Journal of operations management , 26(5):651–668, 2008

    2008

  21. [29]

    Customer rev isit intention to restaurants: Evidence from online reviews

    Xiangbin Yan, Jing Wang, and Michael Chau. Customer rev isit intention to restaurants: Evidence from online reviews. Information Systems Frontiers, 17:645–657, 2015

    2015

  22. [30]

    Financial performance measurement of supply chains: a review

    Masoud Rahiminezhad Galankashi and Farimah Mokhatab R afiei. Financial performance measurement of supply chains: a review. International journal of productivity and performance management, 71(5):1674– 1707, 2022

    2022

  23. [31]

    Anal yzing the relationship between con- sumer satisfaction and fresh e-commerce logistics service using text mining techniques

    Wei Hong, Changyuan Zheng, Linhai Wu, and Xujin Pu. Anal yzing the relationship between con- sumer satisfaction and fresh e-commerce logistics service using text mining techniques. Sustainability, 11(13):3570, 2019

    2019

  24. [32]

    Sustain ing customer loyalty of fresh food e-tailers: an empirical study in china

    Li Cui, Shanshan He, Hepu Deng, and Xiaoyi Wang. Sustain ing customer loyalty of fresh food e-tailers: an empirical study in china. Asia Pacific Journal of Marketing and Logistics , 35(3):669–686, 2023

    2023

  25. [33]

    W hy do tree-based models still outperform deep learning on typical tabular data? Advances in neural information processing systems , 35:507–520, 2022

    Léo Grinsztajn, Edouard Oyallon, and Gaël V aroquaux. W hy do tree-based models still outperform deep learning on typical tabular data? Advances in neural information processing systems , 35:507–520, 2022

    2022

  26. [34]

    Lightgbm: A highly efficient gradient boosting decisio n tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Che n, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decisio n tree. In I Guyon, UV Luxburg, S Bengio, H Wallach, R Fergus, S Vishwanathan, and R Garnett, editors,Advances in Neural Inf...

    2017

Pith tools

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