REVIEW 3 major objections 5 minor 19 references
Variable-Horizon Workforce Demand Forecasting with an Aggregate Demand Constraint for Construction Workforce Planning
T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash
Pith's one-line read This paper claims that construction workforce demand can be forecast over task-specific variable horizons while guaranteeing that each forecast sums to the pre-allocated workforce total, and that this constraint-preserving retrieval…
desk verdict A competent applied forecasting paper whose central constraint-preservation claim hinges on the unstated provenance of the planned total Y_i; worth refereeing but needs that clarification. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the pair formed by a discrete cosine transform (DCT) coefficient vector for the observed demand segment and the cumulative-residual-allocation alignment (CRAA) step. The DCT representation, fit to the prefix normalized by the planned cumulative input, lets tasks of different scale and length be compared by shape; K-means clustering plus cosine similarity selects the reference set. CRAA then maps each reference's normalized residual allocation vector to the target's remaining horizon by treating the reference's cumulative allocation as a piecewise-linear curve $A_q:[0,1]\to[0,1]$ and taking the increment over each target interval. This preserves nonnegativity and sum-to-one, so the weighted average $\mathbf{p}_i$ remains on the probability simplex, and the final scaling by $M_i$ guarantees the aggregate constraints in Equations (13) and (14).
What would settle it
Reconstruct how the planned cumulative input $Y_i$ is obtained from the raw records. If $Y_i$ is the total observed workforce over the full task or is back-fitted from actual demand, then the constraint equations hold tautologically and the reported error reductions are partly an artifact. A decisive test: run CP-RAF using only budgets fixed before task start as $Y_i$; if its accuracy advantage over the baselines collapses, the aggregate constraint was carrying the forecast.
Extended reading notes
Core claim
The central claim is that the variable-horizon, aggregate-constrained forecasting problem can be reformulated as shape retrieval plus residual allocation, and that this reformulation makes the operational constraint hold by construction rather than by post-hoc correction. For a target task with observed prefix $y_{i,1:n_i}$, known planned cumulative input $Y_i$, and residual horizon $L_i = T_i - n_i$, CP-RAF represents the normalized prefix in a DCT basis, assigns the task to a K-means cluster of completed reference tasks, and selects references by cosine similarity in coefficient space. Each reference's normalized residual allocation vector is aligned to the target horizon by cumulative-residual-allocation alignment (CRAA), which reads increments from the reference's cumulative allocation curve; the aligned vectors are averaged with similarity weights to form a proportion vector $\mathbf{p}_i$ on the simplex. Scaling by the remaining planned input $M_i = Y_i - \sum_{m=1}^{n_i} y_{i,m}$ yields $\hat{\mathbf{y}}_i = M_i \mathbf{p}_i$, so $\sum_j \hat{y}_{i,n_i+j} = M_i$ and the full forecast equals $Y_i$ by construction. The paper's evidence is that this construction also forecasts well: best results on all five metrics at 5, 7, 10, and 15 days, significant Diebold–Mariano differences in 30 of 40 baseline comparisons, and stable MAE under task-specific horizons.
Load-bearing premise
The method assumes the planned cumulative workforce input for each task is known at the forecast origin and is a genuine plan, not the eventual actual headcount; the paper never says how that planned total is obtained from field records.
Editorial extensions
If this is right
- At fixed horizons of 5, 7, 10, and 15 days, CP-RAF ranks first on MAE, RMSE, iRMSSE, $R^2$, and aRMSE, with the advantage over the best baseline growing from 1.416 to 5.188 in MAE as the horizon lengthens.
- The Diebold–Mariano tests favor CP-RAF over all eight baselines at the 7- and 10-day horizons at the 5% level, and at the 15-day horizon at the 0.5% level; 30 of the 40 model–horizon comparisons are significant.
- In the variable-horizon setting, CP-RAF produces one forecast per task through its planned completion, covering residual horizons up to 291 days, with MAE between 32.741 and 35.156 across progress ratios 0.1 to 0.9.
- The aggregate demand constraint is enforced by construction: Equations (13) and (14) guarantee the forecast sums to the remaining planned workforce input and to the full planned cumulative input for every task.
- The method's online inference latency stays below 0.05 s at the 99th percentile with 653 reference tasks and below 0.13 s with one million references, so the forecasting loop fits within a daily planning cycle.
Reading between the lines
- The paper leaves implicit that the constraint guarantee is only as good as the planned cumulative input $Y_i$; if $Y_i$ is revised or poorly estimated in practice, the forecast remains internally consistent but can be systematically off against actual demand. An obvious extension is to quantify sensitivity to error in $Y_i$.
- Because CRAA transfers allocation proportions rather than levels, the method should transfer to other units of workforce demand, such as person-hours or crew counts, without redesign as long as the planned total is expressed in the same units.
- The cluster structure itself could be used as an operational diagnostic: a manager can inspect which historical allocation shapes, such as front-loaded, late-concentrated, or stable mid-task, were retrieved for a target task and judge whether the retrieved pattern is plausible for the current job.
- A testable extension suggested by the paper's failure mode is to add a separate spike-timing correction: the results show temporal displacement of sharp demand events inflates RMSE and aRMSE even when MAE is stable, so a post-processing step that aligns predicted peaks to observed peak windows could combine the constraint guarantee with lower pointwise error.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes CP-RAF, a task-level workforce demand forecasting method that produces variable-horizon forecasts while ensuring that the sum of forecast daily demands equals a pre-specified remaining workforce input. The method represents the observed prefix of each task by DCT coefficients normalized by the planned cumulative workforce input, retrieves completed reference tasks in a learned cluster, aligns their residual allocation profiles to the target horizon via CRAA, and scales the similarity-weighted profile by the remaining planned amount. The authors evaluate CP-RAF on shipyard field data (770 tasks, 93,915 task-day observations) in fixed-horizon (3, 5, 7, 10, 15 days) and variable-horizon settings, reporting that CP-RAF ranks first on all five metrics for horizons 5-15, that Diebold-Mariano tests are significant in 30 of 40 baseline-horizon comparisons, and that the aggregate constraint is preserved by construction (Eqs. (13)-(14)).
Significance. If the planned input Y_i is genuinely a pre-specified quantity known at the forecast origin, the paper makes a useful operational contribution: it gives a simple, interpretable mechanism for variable-horizon forecasting under a sum constraint, with a clean construction in Eqs. (13)-(14) and a sensible evaluation framework that includes DM tests and sensitivity analyses. The empirical claims are strengthened by the DM results and by the honest discussion of failure modes (temporal displacement of spikes). The significance is conditional, however, on resolving the provenance of Y_i; the current manuscript does not establish that the method is not being given the realized future sum, which would make the constraint preservation and the reported baseline advantages partly artifacts.
major comments (3)
- [Section 4.1 and Eqs. (2), (4)] The source of the planned cumulative workforce input Y_i is never described. Section 4.1 defines only the observed daily workforce demand (number of workers assigned per day) and explains how missing days were zero-filled; no planning record, contract total, budget allocation, or pre-specified workforce input is introduced. If Y_i is instead the realized cumulative demand of the full task, then M_i in Eq. (2) is the realized residual demand, making the aggregate error zero by construction in Eqs. (13)-(14), and Eq. (4) makes theta_i, the cluster assignment, the cosine similarities, and the weights in Eq. (6) depend on observations after the forecast origin. The authors must identify the data source for Y_i, explain when and how Y_i is recorded, and verify that Y_i is not equal to the realized full-series sum—for example, by reporting the distribution of Y_i minus the observed total for the test tasks and by showing that results are robust to a prefix-only scaling.
- [Section 4.3 and Table 4] The fixed-horizon comparison is not fair with respect to information available to the models: CP-RAF uses Y_i and M_i to scale every forecast, whereas none of the eight baselines receive any aggregate demand information. If planned totals are available operationally, a fair benchmark should also provide the same aggregate information to the baselines, e.g., by post-hoc proportional scaling of each baseline forecast so that it sums to M_i. Without such a control, the reported MAE/RMSE reductions on days 5-15 in Table 4 may largely reflect the value of the constraint information rather than the superiority of the retrieval-allocation mechanism.
- [Table 6 and Section 1] Exp-case 2 does not include any baseline over the same variable horizons, so Table 6 cannot separate allocation skill from knowledge of the residual total; if Y_i is leaked, the stable MAE values in Table 6 are exactly an oracle-constrained allocation. Moreover, the paper's statement in Section 1 that the proposed model 'produced lower forecast errors than the baseline models evaluated in the fixed-length experiments' compares different experimental designs (variable length versus fixed length) and is therefore not supported by Table 6. The authors should either add a sum-scaled baseline evaluated over the same variable horizons or explicitly limit the claim to 'low absolute error levels' rather than a comparison.
minor comments (5)
- [Table BII in Appendix B] The test split row reports Max = 0.00 while Q75 = 2.34 and the minimum is also 0.00; this is internally inconsistent and likely a formatting error that should be corrected.
- [Appendix A, Eq. (A.1)] The displayed definition S_{q,h} = sum_{b=1}^h S_{q,h} is circular; it should be S_{q,h} = sum_{b=1}^h a_{q,b}.
- [Table 5 note] The significance legend uses '*** p < .005' and '** p < .01'; a threshold of .005 for three stars is unconventional, and using p < .001 would be more standard and easier to compare with other work.
- [Section 4.3] The sentence beginning '3 presents the standard parameterization...' appears immediately after Table 3 and is grammatically incomplete; it should be revised to refer to Table 3 explicitly.
- [Section 4.1 and Table BI] The paper does not report how many tasks were excluded as ineligible for reference retrieval at each stage, nor the criteria-based counts; reporting these exclusions would help readers assess the generality of the method.
Circularity Check
No significant circularity: the aggregate constraint is an input by definition, but the unverified provenance of Y_i leaves a conditional future-leakage risk.
-
other
[Section 3.1, Eq. (2); Section 3.3, Eqs. (13)-(14); Section 4.1]
"Here, T_r denotes the full planned series length and Y_r > 0 is the planned cumulative workforce input known at the forecast origin. ... The source data consisted of work-performance records manually completed by onsite personnel."
The forecast's aggregate is fixed as an input: Eq. (13) makes the sum of the forecast equal to M_i, and Eq. (2) defines M_i = Y_i minus the observed prefix. This is a construction, not a learned prediction, and the paper explicitly says the constraint is satisfied by construction. Circularity would arise only if Y_i is actually the realized full-series total rather than a plan known at the forecast origin. The paper never documents how Y_i is obtained from the field records, so the reader cannot verify that Eqs. (4) and (13)-(14) do not leak post-origin information into the retrieval and the final sum. The reported accuracy gains are therefore conditional on the untested provenance of Y_i; no other fitted quantity is renamed as a prediction.
full rationale
CP-RAF is largely self-contained against its stated inputs. The DCT representation, clustering, retrieval, CRAA alignment, and similarity-weighted averaging are all fitted or computed from observed prefixes and completed reference tasks. Hyperparameters are tuned on validation data, and the eight baselines are external, domain-aligned models. The aggregate-demand preservation is explicitly a by-construction property: the forecast is M_i times a unit-sum allocation profile, so Eqs. (13)-(14) follow by algebra rather than by fitting. That by itself is not circular, because the paper takes Y_i as a known planned total and never claims to predict it. The only concern is that the dataset section never identifies the source of Y_i in the field data; if Y_i were instead the realized task total, both the normalization in Eq. (4) and the final sum in Eq. (13) would encode the answer. Since the paper defines Y_i as planned and known in advance, this is a provenance/leakage risk rather than a demonstrated circular reduction. Consequently, the circularity score is 1 rather than 0.
Assumptions & free parameters
free parameters (4)
- Number of K-means clusters (k) =
6 to 11 depending on horizon and progress ratio (Appendix C)
- DCT coefficient dimension (d_dct) =
12, 24, or 32 depending on configuration (Appendix C)
- Ridge penalty (lambda_ridge) =
Not reported
- Softmax temperature (tau) =
Not reported
assumptions (5)
- domain assumption DCT coefficients of the normalized observed prefix capture the temporal shape relevant for forecasting.
- domain assumption Reference tasks with similar prefix shape have similar residual allocation patterns.
- domain assumption Planned total duration T_i and planned cumulative input Y_i are known and reliable at the forecast origin.
- domain assumption K-means clustering in DCT coefficient space yields meaningful reference sets.
- standard math Piecewise-linear interpolation of cumulative allocation curves preserves the allocation distribution.
Cite this review
Pith. "Pith review of Variable-Horizon Workforce Demand Forecasting with an Aggregate Demand Constraint for Construction Workforce Planning." pith.science (2026). https://pith.science/paper/3IOD2LUO
@misc{pith2026260805551,
author = {Pith},
title = {Pith review of: Variable-Horizon Workforce Demand Forecasting with an Aggregate Demand Constraint for Construction Workforce Planning},
year = {2026},
howpublished = {\url{https://pith.science/paper/3IOD2LUO}},
note = {Machine review of arXiv:2608.05551}
}
read the original abstract
Workforce planning is a recurring operational decision during construction projects that requires accurate forecasts of the future workforce demand for individual tasks. However, in practice, tasks have different completion dates, resulting in variable forecast horizons. In addition, the sum of the predicted daily workforce demands must equal the total workforce allocation specified in advance. Most existing machine-learning (ML)-based forecasting models assume fixed-length outputs and do not explicitly impose an aggregate demand constraint, making them unsuitable for these operational requirements. To address this problem, this study proposes constraint-preserving residual allocation forecasting (CP-RAF). The CP-RAF represents an observed workforce demand time series as a coefficient vector and retrieves completed tasks with similar temporal shapes. Then, it estimates the allocation profile over the remaining task duration using similarity-weight averaging. The predefined remaining workforce demand for each task was distributed according to the estimated profile, and the forecast horizon was adjusted while retaining profile characteristics. This procedure accommodates variable forecast horizons while preserving the aggregate demand constraints. CP-RAF was evaluated using workforce demand field data. The results showed that CP-RAF outperformed eight baseline models in medium- and long-horizon fixed-length forecasting and maintained low forecast errors under variable-length forecasting. By directly incorporating operational constraints into the forecasting procedure, the proposed method provides a framework suitable for workforce allocation in construction practices.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[2]
Hyperparameter combination Progress ratio 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 (k,ddct)(10,24) (10,24) (11,32) (10,24) (8,24) (11,24) (8,24) (11,32) (10,24) 24 A preprint - August 7, 2026 D Supplementary results for alternative distance measures and basis functions Table DI: Workforce demand forecasting performance across nine progress ratios using four di...
work page 2026
-
[8]
URLhttps: //arxiv.org/abs/2304.08424. Domingos S. de O. Santos Júnior, João F. L. de Oliveira, and Paulo S. G. de Mattos Neto. An intelligent hybridization of ARIMA with machine learning models for time series forecasting.Knowledge-Based Systems, 175:72–86,
-
[9]
doi: 10.1016/j.knosys.2019.03.011. Francis X. Diebold. Comparing predictive accuracy, twenty years later: A personal perspective on the use and abuse of Diebold–Mariano tests. Working Paper 18391, National Bureau of Economic Research,
-
[11]
doi: 10.1080/15623599.2022.2112898. Rob J. Hyndman, Roman A. Ahmed, George Athanasopoulos, and Han Lin Shang. Optimal combination forecasts for hierarchical time series.Computational Statistics & Data Analysis, 55(9):2579–2589,
arXiv 2022
-
[13]
doi: 10.3389/fieng.2023.1267244. Ayodeji Ogungbire and Sudip K. Mitra. Workforce forecasting for state transportation agencies: A machine learning approach.International Journal of Transportation Science and Technology, 17:345–360,
-
[14]
Horizon-Aware Forecasting of Passenger Assistance Demand for Rail Station Workforce Planning
doi: 10.1016/j.ijtst.2024.05.004. Michael Sheehan and Irina Timoshenko. Horizon-aware forecasting of passenger assistance demand for rail station workforce planning.arXiv preprint arXiv:2604.16464,
work page Pith review arXiv 2024
-
[17]
doi: 10.1108/09699981111098
-
[18]
doi: 10.20469/ijaps.2.50005-3. G. Peter Zhang. Time series forecasting using a hybrid ARIMA and neural network model.Neurocomputing, 50:159–175,
Show all 19 references
-
[2003]
doi: 10.1016/S0925-2312(01)00702-0. A Computational detail for CRAA Let Piecewise-linear cumulative evaluation, Sq,0 = 0, Sq,h = h∑ b=1 Sq,h, h= 1,...,L q (A.1) 22 A preprint - August 7, 2026 For0 ≤v < 1, setx =vLq, b =⌊x⌋, andγ =x−b . The cumulative allocation curves in Equat...
2026 doi
-
[2005]
doi: 10.1080/01446190500204911. James M. W. Wong, Albert P. C. Chan, and Y. H. Chiang. Modeling and forecasting construction labor demand: Multivariate analysis.Journal of Construction Engineering and Management, 134(9):664–672,
-
[2006]
21 A preprint - August 7, 2026 Sung Hee Chung, Dong Chan Jung, Sun Nyung Yoon, and Dae Lee
doi: 10.1108/14714170610710622. 21 A preprint - August 7, 2026 Sung Hee Chung, Dong Chan Jung, Sun Nyung Yoon, and Dae Lee. A dynamic forecasting model for nursing manpower requirements in the medical service industry.Service Business, 4(3–4):225–236,
2026 doi
-
[2008]
doi: 10.1061/(ASCE)0733-9364(2008)134:9(664). James M. W. Wong, Albert P. C. Chan, and Y. H. Chiang. Construction manpower demand forecasting: A comparative study of univariate time series, multiple regression and econometric modelling techniques. Engineering, Construction and...
2008 doi
-
[2011]
doi: 10.1016/j.csda.2011.03.006. Guido J. L. Micheli, Alessandro Martino, Federico Porta, Andrea Cravello, Marco Panaro, and Alberto Calabrese. Workforce planning in project-driven companies: A high-level guideline.Frontiers in Industrial Engineering, 1:1267244,
2011 doi
-
[2013]
Ahmad Alqatawna, Bilal Abu-Salih, Nour Obeid, and M
doi: 10.5592/otmcj.2013.2.1. Ahmad Alqatawna, Bilal Abu-Salih, Nour Obeid, and M. Almiani. Incorporating time-series forecasting techniques to predict logistics companies’ staffing needs and order volume.Computation, 11(7):141,
2013 doi
-
[2016]
doi: 10.5959/eimj.v8i3.421. Anil L. Agarwal, B. L. Rajput, and E. S. Satpute Mangesh. Model formulation to estimate manpower demand for the real-estate construction projects in india.Organization, Technology & Management in Construction: An International Journal, 5(2):828–833,
-
[2019]
Nan Cao and Michael C
doi: 10.1080/20479700 .2017.1390182. Nan Cao and Michael C. P. Sing. Workforce forecasting in the building maintenance and repair work: Evaluating machine learning and LSTM models.Journal of Building Engineering, 95:110125,
2017
-
[2023]
Bashir Azimi Nayebi, Rafat Mohebbifar, Jalil Azimian, and Sima Rafiei
doi: 10.3390/computation11070141. Bashir Azimi Nayebi, Rafat Mohebbifar, Jalil Azimian, and Sima Rafiei. Estimating nursing staff requirement in an emergency department of a general training hospital: Application of workload indicators of staffing need (WISN).International Jou...
-
[2024]
Albert P
doi: 10.1016/j.jobe.2024.110125. Albert P. C. Chan, Yat Hung Chiang, Stephen W. K. Mak, Lennon H. T. Choy, and M. W. W. James. Forecasting the demand for construction skills in Hong Kong.Construction Innovation, 6(1):3–19,
2024
-
[2025]
Hady Elkholosy, Rana Ead, Ahmed Hammad, and Simaan AbouRizk
doi: 10.1016/j.sca.2024.100099. Hady Elkholosy, Rana Ead, Ahmed Hammad, and Simaan AbouRizk. Data mining for forecasting labor resource requirements: A case study of project management staffing requirements.International Journal of Construction Management, 24(5):561–572,
2024
Reviewed August 8, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.