acceptodds
Under review as a conference paper at ICLR 2027

Learning to Staff from Operational Proxies: Towards Win-Win On-Demand Staffing

Abstract

On-demand staffing plans labor in advance to meet changing business needs. Ideally, it achieves a win-win outcome for businesses and workers: reducing unnecessary labor-hours while meeting business needs and keeping workloads manageable. However, historical schedules record assigned staffing rather than minimum requirements, so learning from them can reproduce overstaffing. In this paper, we propose ProxyStaff, a general supervision framework that learns staffing requirements from paired business and scheduling records. An operational proxy relates each role's labor to supported revenue and transactions, treating required roles as complementary stages. Its inverse yields minimum staffing requirements under the proxy. These labels enable different backbone models to learn future hourly staffing needs directly from planning-time information, through training from scratch or fine-tuning pretrained models. To evaluate ProxyStaff, we collect two real-world datasets from a restaurant chain with 13 stores and a milk-tea chain with 8 stores. We evaluate 6 backbone models for planning 1 or 7 days in advance, reflecting managers' routine lead times for publishing staffing needs. Across multiple backbones, ProxyStaff more often matches human-manager business support with fewer labor-hours than learning historical assignments or forecasting business before converting it to staffing. The results show that ProxyStaff can consistently reduce labor-hours by about 20–30% relative to human-manager schedules across both chains and planning horizons while matching or exceeding their business support under the proxy.

Primary area:
learning on time series and dynamical systems
Keywords:
On-demand staffing, Supervision design, Workforce scheduling, Time-series forecasting, Learning from operational data

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_doGAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

The work proposes ProxyStaff, a way to learn hourly staffing requirements from historical business and schedule records. Since schedules record assigned rather than required staffing, the authors fit an operational proxy that estimates how much revenue and how many transactions each staffing level per role has supported and invert it to obtain the minimum staffing for each past hour. Forecasting models are then trained to predict these labels one or seven days ahead. On two datasets collected from a restaurant chain and a milk-tea chain, with six backbones, ProxyStaff is compared with learning the managers’ schedules and with forecasting business and converting it through the same proxy, using support and labor metrics defined by the proxy. The datasets and code are not available for review; the authors plan to release them upon acceptance, subject to licensing.

Critical strengths
  1. The paper addresses a practical obstacle to learning staffing from data. Historical schedules record what managers assigned, not what was needed, so a model trained to imitate them also learns their overstaffing. ProxyStaff instead derives requirement labels from records that businesses already keep (hourly revenue, transactions, and schedules), so it needs no new data collection or optimal labels, and the labels can be used with any forecasting model.

  2. The comparison is controlled: three supervision schemes are compared within each of six backbones, with the same inputs, training budget, and calibration, on two real datasets at two planning horizons, with seed-level results. Because Business+ proxy uses the same proxy, it isolates the effect of changing the prediction target.

Critical weaknesses
  1. The results cannot be checked or reproduced during review. The code link is a placeholder, and the datasets, listed as a contribution, will be released only upon acceptance and subject to licensing. Because this is a method-and-experiment paper rather than a theoretical one, checking that the code is consistent with the paper is an important part of my assessment. The code involves no commercial data and could be shared anonymously now. If confidentiality is the concern, the data could be shared in a protected form, for example with stores anonymized and revenue and transactions rescaled by a secret constant, which leaves the labels and metrics unchanged.

  2. The claimed win-win is not supported. The paper frames the work as a win-win for businesses and workers, but neither side is checked against an observed outcome. The workers’ side, such as workload or income from fewer paid hours, is not measured. For the business side, the proxy takes the business recorded in past hours with a given staffing level (its 95th percentile) as the business that this staffing can handle. The records show that this business occurred, not that it was served well; customers may have waited or left, and staff may have been overloaded. Both the training labels and the evaluation rest on this assumption, so if it overstates what a smaller staff can handle, the reported savings would hide real understaffing. Checking the proxy against an outcome it does not use, such as service times or lost sales, would address this.

Questions for authors
  1. Address the weaknesses above, respectively.
  2. How general is ProxyStaff? The study covers two food-service chains whose assumptions, such as complementary roles and comparable productivity across stores, may not hold elsewhere. Please state the conditions the framework requires and which other settings, in staffing or in other applications, meet them, ideally with one example beyond food service. This would strengthen the case for the paper’s significance.
1 comment
AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer,

We would appreciate your feedback on our responses. Do these address your main concerns? We would be happy to clarify any remaining questions during the discussion period.

Thank you again for your time and consideration.

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

This paper is about using AI to plan staffing for places like restaurants: how many people each role needs each hour. The difficulty is to predict staffing needs without copying potentially inefficient schedules from the past. The idea is to work out the real requirements from the data. It uses historical sales, order counts, and staffing records to estimate how many people each role needs, then trains models to predict those needs a day or a week ahead. On restaurant and milk-tea data, several models match the managers’ business support scores with fewer staff hours.

Critical strengths
  1. It is great to see that the paper addresses a practical problem: historical schedules tell us how many people were assigned, but not how many were actually needed. Constructing staffing requirement labels from business records is a useful way to rethink the learning target, and the approach can work with different forecasting models.

  2. It is tested on two real-world chains and both day-ahead and week-ahead planning, so the results are more practical.

  3. The authors keep the inputs, training budgets, and calibration procedures matched within each model, making it easier to see what changing the supervision contributes.

  4. The writing is easy to follow.

Critical weaknesses
  1. My main conern is that the same proxy is used to make the labels and score the results. ProxyStaff learns this target directly, so the setup naturally plays to its strengths. The historical-schedule baseline uses the proxy for calibration and evaluation, but is not directly trained to predict the target. How much of the gain comes from matching the scoring rule better? Would the results still hold if the evaluation proxy were fitted on separate data or used different settings?

  2. I am also not sure how well the proxy captures real staffing capacity. The managers’ schedules score about 80%+ on hours and only around 60% on revenue and transactions. Are managers really short-staffed? Managers also tend to add staff when they expect more business, so the records mix together demand and capacity. The proxy pools stores within each chain, but staff productivity could vary quite a bit across locations.

  3. Predicting how many people are needed each hour is still encouraged, moving to making a schedule a store can actually use.

Questions for authors

How do you check that the proxy captures what staff can reasonably handle, rather than just how busy a store happened to be? Do you have any independent evidence from service quality or employee workload measures?

1 comment
AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer,

We would appreciate your feedback on our responses and the additional experiments, including the percentile sensitivity results. Do these address your main concerns? We would be happy to clarify any remaining questions during the discussion period.

Thank you again for your time and consideration.

Reviewer_FLjyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PROXYSTAFF constructs hourly staffing requirement labels from historical business and scheduling records. A monotone operational proxy maps each role's labor to supported revenue and transactions; its inverse gives the minimum staffing needed under that proxy. Forecasting models learn these requirements directly. Experiments on two service chains compare this approach with learning historical staffing and forecasting business before applying the same proxy, using six backbones, two planning horizons, and three seeds. With N-HiTS, PROXYSTAFF reduces labor by 23.7-24.2% on Restaurant and 30.1-31.9% on MilkTea while meeting the manager reference's proxy support thresholds. The study evaluates hourly requirements retrospectively; service quality, worker workload, and feasible shifts remain outside the evaluation.

Critical strengths
  1. The supervision target addresses a useful problem: historical schedules show assigned labor, but do not establish the minimum labor needed. Equations (5)-(6) explain why the proposed labels preserve the proxy's support decision. The labels can be used with different forecasting architectures, and the Business + proxy baseline provides a relevant comparison using the same operational assumptions.
  2. The experimental reporting is careful. Within each backbone, the methods share inputs, fitting budgets, and calibration procedures. Appendix B explains the seven-day exclusion needed to respect the week-ahead planning origin. Appendix C reports individual support measures and seed-level attainment: MLP, ResMLP, and N-HiTS each meet the support-and-labor criterion in all 12 runs. The paper also acknowledges that gains vary across backbones and that proxy support does not establish actual service capacity.
Critical weaknesses
  1. The method's dependence on the proxy is insufficiently tested. Equation (1) pools stores and estimates support from historically assigned labor, so managerial allocation decisions and store productivity can affect the labels and evaluation. Appendix A varies the percentile only for the manager reference, not for the learned methods. Restaurant revenue support changes from 64.5% at P95 to 98.0% at P100. This does not invalidate P95, but it leaves open whether the reported method rankings and savings persist under other reasonable proxy choices. I accept the paper's restriction to proxy-based evaluation; my concern is the robustness of the conclusions within that scope.
  2. The reported operating points do not establish when direct requirement labels improve the support-labor tradeoff. Equation (4) minimizes labor once reference support is met. On MilkTea, MLP with Business + proxy meets the constraints in every seed and saves 28.7%/29.4%, compared with 6.7%/11.1% for PROXYSTAFF, which provides higher support (Tables 3 and 6). N-HiTS benefits from PROXYSTAFF, so this is not evidence of uniform inferiority. However, the MLP result raises a question about whether the differences reflect a better attainable tradeoff or different levels of conservatism. The paper already calibrates on validation; reporting the selected settings and a small set of validation-selected operating points would help resolve this. I lean toward weak rejection because the useful supervision idea and positive N-HiTS results are not yet accompanied by enough evidence about sensitivity to the proxy and the conditions under which direct labels outperform business forecasting. These concerns appear addressable within the existing experimental framework. Robust rankings under nearby proxy choices and a clearer explanation of the support-labor tradeoff could change my recommendation.
Questions for authors
  1. Do the main conclusions hold at nearby proxy percentiles, such as P90 and P97.5? A focused comparison of PROXYSTAFF and Business + proxy with N-HiTS and MLP would be sufficient. Please apply each proxy consistently to label construction, calibration, and evaluation.
  2. What explains the MilkTea MLP savings gap when both methods meet all constraints? Please report the selected quantiles and scales and the scale-search procedure. Could a small set of operating points selected on validation clarify the tradeoff under the same test protocol, without selecting settings using test outcomes?
  3. Section 5.3 reconstructs full-time labor over the operating window while retaining flexible workers' recorded intervals. How is full-time labor reconstructed, and does the same convention enter Tables 3, 5, and 6? Clarifying whether the reference represents scheduled attendance or another accounting convention would resolve an ambiguity in the savings denominator.
1 comment
AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer,

We would appreciate your feedback on our responses and the additional experiments, including the percentile sensitivity results. Do these address your main concerns? We would be happy to clarify any remaining questions during the discussion period.

Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-up to Reviewers FLjy and Fcpy

Dear Reviewers FLjy and Fcpy,

We believe our rebuttal and revised manuscript have addressed all concerns raised in your reviews through clarifications and additional experiments. These updates are now incorporated into the manuscript and marked in blue. As the discussion deadline approaches, we respectfully ask you to reconsider your ratings in light of this evidence. If any concerns remain, we would appreciate the opportunity to address them. Thank you for your time and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.