A QUBO-Based Optimization Framework for ATM Cash Replenishment Scheduling

Davide Caputo; Davide Corbelletto; Davide Veronelli; Francesca Cibrario; Giulia Montani; Luca Toffanetti; Riccardo Aiolfi; Valeria Zaffaroni

REVIEW 4 major objections 5 minor 12 references

A QUBO-Based Optimization Framework for ATM Cash Replenishment Scheduling

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper shows that ATM cash replenishment can be scheduled with a QUBO optimization model, solved on a GPU, to cut refill costs by 15–18% while keeping average service levels above 99.8%.

desk verdict Real industrial QUBO case study, but the service-level penalty is sign-inverted and the empirical claims aren't reproducible as written. read the letter →

arxiv 2607.22675 v1 pith:QE5N6CAA submitted 2026-07-10 math.OC quant-ph

Riccardo Aiolfi , Giulia Montani , Valeria Zaffaroni , Luca Toffanetti , Davide Veronelli , Francesca Cibrario , Davide Caputo , Davide Corbelletto This is my paper · ORCID

classification math.OCquant-ph

keywords ATMcashreplenishmentQUBOGPU-acceleratedoptimizationlogisticsservicelevelthresholdpolicyco-locationdiscountscheduling

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

This paper sets out to show that ATM cash replenishment scheduling can be recast as a Quadratic Unconstrained Binary Optimization (QUBO) problem and solved with a GPU-accelerated solver, producing cheaper schedules than the threshold-based replenishment rules banks commonly use. On a real dataset of 276 ATMs across four months, the authors report cost reductions between 15% and 18% while keeping average service levels at 99.8–99.9%. The QUBO formulation captures fixed and variable refill costs, discounts for servicing co-located machines on the same day, and penalties for violating operational and service-level constraints. The practical promise is a decision-support tool that runs daily, converts the optimized schedule into trigger thresholds, and does not sacrifice cash availability for savings.

What carries the argument

The load-bearing object is the QUBO objective H(x) = λ1·H_cost + λ2·H_discount + λ3·H_feasibility + λ4·H_service, where x_{i,t} = 1 means ATM i is refilled on day t. H_cost is linear in fixed/variable costs, H_discount is a quadratic pair term for ATMs at the same address refilled the same day, H_feasibility is a quadratic at-most-one penalty per ATM, and H_service is a linear penalty built from precomputed shortages Γ_{i,t} = min(0, projected balance − contingency). The model also maps the binary plan into daily thresholds that trigger actual service tickets, and the resulting QUBO matrices are solved by a GPU-accelerated QUBO solver.

What would settle it

Take a single ATM with known starting cash, a fixed contingency level, and deterministic daily withdrawals such that the projected balance drops below the contingency level on day 3. Build the QUBO from the paper's equations with p_{i,t}=1 for all t, solve it, and check whether the chosen refill day is 3 or a later day. If it is later, the service-penalty sign in the paper does not do what the text claims.

Watch

Extended reading notes

Core claim

The central claim is that the ATM cash replenishment problem — choosing which machines to refill on which days, with co-location discounts and shortage avoidance — can be expressed as a QUBO instance that a GPU solver can handle at network scale. The objective combines linear costs for each refill, a quadratic term for same-day co-location discounts, an at-most-one penalty so each ATM is refilled at most once in the horizon, and a linear service penalty computed from projected cash balances. Solving this QUBO and translating the binary decisions into daily activation thresholds yields, in the authors' retrospective simulation, 15–18% lower replenishment costs than a threshold-based baseline

Load-bearing premise

The load-bearing premise is that the service-level penalty, through its sign and the tunable coefficients p_{i,t}, makes the optimizer choose the first day a cash-out is projected; if that sign is implemented as written in Eq. (12), the penalty actually rewards later refills, so the claimed service-level maintenance is only as good as that tuning.

Editorial extensions

If this is right

If the paper is right, a daily operating pipeline that solves this QUBO and converts the result to thresholds can cut refill costs by double digits without lowering average availability.
The reported average service levels around 99.8–99.9% suggest the cost savings do not come from tolerating frequent cash-outs.
The minimum service level drops more than the baseline in some months (e.g., 96.99% in October), meaning the approach trades off worst-case availability for average cost; penalty weights control that trade.
Because the model runs on a rolling horizon and only needs forecasts, it can be embedded into an existing forecast-plus-threshold operational system.

Reading between the lines

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

The sign of the service-level penalty as printed (Γ ≤ 0 times positive λ4) gives a reward for refilling on days after the projected shortage, so the reported service levels likely depend on the implementation's coefficient tuning rather than the equations alone; this should be checked before real deployment.
The at-most-one-refill-per-ATM constraint means the planning horizon cannot exceed the typical inter-replenishment interval; for longer horizons the model would need to be re-cast, for instance by adding a binary state variable per day.
The quadratic co-location discount overestimates the actual discount for sites with more than two co-located ATMs; the authors note this can be recalibrated, but the magnitude of the approximation error is not quantified.
A natural next test is to run the same pipeline with different penalty weights and report the Pareto frontier between cost and minimum service level, which would help operators pick a risk posture.

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.

Desk Editor's Note

read the letter

Dear colleague,

The headline: this is a genuine industrial case study of QUBO-based ATM replenishment, but the main reported result — 15–18% savings at 99.8% service — is not supported as written because the service-level penalty in Eq. (12) is inverted.

What is new and good: the co-location discount is encoded as a quadratic interaction between machines at the same address, which is a natural fit for QUBO. The “dual-meaning” time indices for Saturday-service and non-Saturday ATMs, with the corresponding forecast aggregation and discount eligibility rules, are a thoughtful piece of operational modeling. The rolling-threshold translation from binary schedule to daily activation thresholds is also reasonable. The evaluation uses a real dataset of 276 Italian ATMs over four months, which is more than most QUBO papers.

The soft spots are serious. First, Γ_i,t is defined as min{0, S̃_i,t − Fix_i} ≤ 0, and H_service = λ4 Σ p_i,t Γ_i,t x_i,t. With p>0, minimizing this term rewards replenishment on days with negative Γ — i.e., late days. The text says p can be tuned so the first critical day dominates, but for any positive p, later days give more negative contributions. So the mechanism does the opposite of what the paper claims. The reported ASL cannot come from this term. It may come from the threshold translation, but the paper needs to say so.

Second, nothing is reproducible. MegaQUBO is proprietary, no code or data are released, and the penalty weights, contingency levels, threshold baseline rules, and forecast parameters are unreported. Third, the baseline is an uncharacterized threshold heuristic, and the cost KPI is computed under the same approximate cost model the optimizer minimizes — including the co-location discount term the paper itself admits overestimates the contractual discount for groups larger than two. So the savings number is partly circular.

The paper is not a takedown; the authors clearly know the operations. But as written, the central empirical claim is not testable. I'd treat it as a promising technical report, not a publishable research result. If the sign is fixed, the parameters released, and a MILP or stronger heuristic baseline added, this could become a useful case study. Right now I'd send it to referees only with an explicit request for major revision and re-review; a desk reject would be defensible, but there is enough real modeling here to deserve referee time.

Referee Report

4 major / 5 minor

Summary. The paper proposes a QUBO model for ATM cash replenishment scheduling, encoding fixed and variable costs, co-location discounts, an at-most-one replenishment constraint per ATM, and a service-level term as a binary quadratic objective. The QUBO instances are solved with MegaQUBO, a proprietary GPU-accelerated solver, and evaluated on 276 Italian ATMs over four months of 2022 against a threshold-based baseline. The paper claims consistent cost reductions of 15%-18% while maintaining an average service level around 99.8%-99.9%.

Significance. If the claims are valid, the work demonstrates a practically relevant industrial application of QUBO/GPU optimization to cash logistics, a domain where exact methods scale poorly and nonlinear cost interactions are common. The empirical setup uses real operational data and a realistic cost structure, and the QUBO formulation is a natural fit for the problem. However, the manuscript does not release code or data, does not specify the penalty weights or service-level coefficients, and defines its central KPI imprecisely. The empirical claims are therefore not reproducible in their current form, and the practical significance is conditional on the authors clarifying those details.

major comments (4)

[Sec. 3.5.2, Eqs. (11)-(12)] The service-level term is not a penalty but a reward for selecting a projected-shortage day, since Gamma is non-positive. The text says p_i,t can be tuned so the first critical day dominates; this is possible in principle because p_i,t is indexed by (i,t), so the sign inversion does not by itself invalidate the model. However, the paper gives no rule or values for p_i,t, nor for lambda1-lambda4. As written, the ASL claim is an unverifiable tuning assertion rather than a property of the model. Please specify the tuning procedure or report the coefficient values, and consider redefining Gamma as max(0, Fix_i - S_tilde_i,t) so the term behaves as a true penalty for late replenishment.
[Sec. 4.3, Eq. (14)] The ASL formula as printed, ASL = 1 - Sum lost_i,t / W*_i,t, sums per-observation lost-demand ratios; with thousands of ATM-days this would produce values far from 0.99 and is not a fraction of total demand. The reported Table 1 numbers are therefore not defined by Eq. (14). Please correct to either 1 - (Sum lost)/(Sum W*) or 1 - (1/M)Sum lost/W*, and recompute Table 1 accordingly. Also define 'minimum service level' used in the same table.
[Sec. 3.4 and Sec. 4.3] The relationship between the optimized objective and the reported cost KPI is unclear. Section 4.3 defines total cost as the sum of per-replenishment fixed and variable costs, while Eq. (5) defines the true contractual cost including the co-location discount. The QUBO discount term is admitted in Sec. 3.4 to overestimate the contractual discount when more than two ATMs are co-serviced. The authors should state which cost measure Table 2 reports, and also give the true contractual cost from Eq. (5) to quantify the impact of the overestimated discount on the 15-18% savings.
[Sec. 4.2 and 4.4] The threshold-based baseline is not fully specified: no formula or parameter values are given for the threshold, nor is it stated whether it includes a safety margin or co-location discounts. Without this, the 15-18% improvement has no well-defined comparator. Please provide the exact baseline policy used to generate Tables 1-2, including any tuning of the threshold and the contingency margin.

minor comments (5)

[Sec. 3.5.2] 'Strictly negative thereafter' is imprecise: Gamma_i,t equals -Fix_i once S_tilde_i,t reaches zero, so the sequence is non-increasing, not strictly negative throughout.
[Eq. (14)] Add explicit parentheses or summation indices to the ASL expression; the current notation is ambiguous and should match the intended definition.
[Table 2] The heading contains the typo 'Cost ineSolution'; it should be 'Cost in Euro' or 'Cost (EUR)'.
[Sec. 4.1] The rolling-horizon simulation over the four months is described only through a 9-day example; please specify the horizon length, the overlap between consecutive windows, and whether forecasts are updated daily.
[Sec. 5] The claim that the pipeline has been 'translated and used into a production-grade decision-support pipeline' is not accompanied by deployment details or evidence; either remove or substantiate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the QUBO cost reduction is an empirical benchmark, not an identity; the service-penalty sign issue is a correctness flaw outside this pass.

full rationale

The paper's central claim is an empirical benchmark: a QUBO schedule generated by MegaQUBO is compared, under a retrospective simulation, with a threshold-based policy. The total-cost KPI is described independently ('the total cost for the entire benchmark period is the sum of the individual costs for every replenishment scheduled by the model'), and the service-level KPI is an out-of-sample counterfactual based on realized withdrawals (Eqs. 13–14). No equation reduces the reported savings to the model's own fitted parameters or to a self-citation. The co-location discount term in the objective is admitted to overestimate the contractual discount for >2 co-located ATMs (Sec. 3.4), but the text does not state that the KPI cost evaluation uses this approximate term; without that identity, the overestimation is an optimizer-bias/correctness issue, not a demonstrable circularity. The self-citations to Data Reply's MegaQUBO page and case studies are promotional background and are not used to derive the 15–18% saving; they do not make the derivation circular. The sign of H_service (Eqs. 11–12) is a substantive modeling error — Gamma <= 0 makes the penalty reward shortage-day replenishment — but it is an error in the claimed mechanism, not an instance of a result being equivalent to its input by construction. No circular step is therefore established.

Assumptions & free parameters 10 free parameters · 8 assumptions · 0 invented entities

The central claim rests on a half-dozen hand-set or unreported coefficients plus domain assumptions about at-most-once replenishment and forecast accuracy. The most serious is the sign/role of p_i,t in the service penalty, which is asserted but not derived.

free parameters (10)

F (fixed replenishment cost) = 52
Set in Section 4.2 "to ensure consistency with realistic operational settings"; savings scale directly with it.
c (variable loading cost coefficient) = 0.000069
Chosen cost parameter in Section 4.2; affects the trade-off between early and late replenishment.
s (residual-handling cost coefficient) = 0.000333
Chosen cost parameter in Section 4.2; makes later replenishment cheaper by reducing counted residual cash.
D (co-location discount) = 26.69
Chosen in Section 4.2; drives the quadratic discount term and part of the reported savings.
λ1–λ4 (objective weights) = not reported
Introduced in Eq. (6) and said to "need to be tuned appropriately"; never given values, yet they determine feasibility and service behavior.
p_i,t (service-level coefficients) = not reported
Introduced in Eq. (12); the paper says they can be tuned so the first critical day dominates, but values and construction are absent.
Fix_i (contingency cash level) = not reported
Used in Eq. (11) to define shortage; not specified for the 276 ATMs.
Safety contingency margin in threshold translation = not reported
Section 3.1 mentions a margin "that absorbs demand variability" without quantifying it.
Threshold-based baseline parameters = not reported
Section 4.2 describes the baseline qualitatively; the trigger threshold and safety level are not specified.
Forecast model and hyperparameters = not reported
Section 4.1 says a time-series forecasting model produces withdrawal estimates, but the model, training window, and accuracy are not described.

assumptions (8)

standard math Penalty reformulation with sufficiently large λ reproduces the constrained optimum.
Standard QUBO penalty theory invoked in Section 3.2, Eqs. (2)–(4).
domain assumption Each ATM is replenished at most once per planning horizon.
Section 3.5.1 says historical data indicate this is rare, but the model enforces it; valid multi-replenishment schedules are excluded.
domain assumption All replenishments happen at end of day, and this is a conservative approximation.
First simplifying assumption in Section 3.4; shifts the timing of cash availability.
domain assumption The no-replenishment trajectory S̃_i,t can be used to evaluate residual cash at a replenishment day.
Eqs. (8)–(9) in Section 3.4; valid only under at-most-once replenishment and ignores demand feedback after refill.
domain assumption Forecasted withdrawals W_i,j are accurate enough to set thresholds and penalties.
Used throughout Section 3.3 and 4.1; the paper provides no forecast validation.
ad hoc to paper The quadratic co-location discount approximates the contractual discount.
Section 3.4 explicitly states the quadratic term overestimates the real discount when more than two co-located ATMs are serviced together.
ad hoc to paper p_i,t > 0 can make the first critical shortage day dominate all later days.
Section 3.5.2 asserts this without derivation; with Γ≤0 the printed equations favor later days instead.
domain assumption Dual-meaning time indices preserve service when withdrawal forecasts are aggregated.
Section 4.1 aggregates weekend withdrawals into Friday for ATMs without Saturday service; this assumes the aggregation exactly captures multi-day demand.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A QUBO-Based Optimization Framework for ATM Cash Replenishment Scheduling." pith.science (2026). https://pith.science/paper/QE5N6CAA

@misc{pith2026260722675,
  author       = {Pith},
  title        = {Pith review of: A QUBO-Based Optimization Framework for ATM Cash Replenishment Scheduling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QE5N6CAA}},
  note         = {Machine review of arXiv:2607.22675}
}

read the original abstract

The management of cash replenishment in Automated Teller Machine (ATM) networks requires scheduling recharges in order to minimize operational costs while maintaining high service levels and avoiding cash-outs, under uncertain and time-varying withdrawal demand. This work formulates the ATM cash replenishment problem through a Quadratic Unconstrained Binary Optimization (QUBO) model, which naturally captures nonlinear cost interactions, while incorporating operational constraints through penalty terms. The objective function combines fixed and variable replenishment costs with co-location discounts, as well as penalties for a late replenishment that could cause a service interruption. The resulting QUBO instances are solved using MegaQUBO, a GPU-accelerated QUBO solver. An empirical evaluation on a real dataset of 276 ATMs located in Italy, covering four representative months of 2022 (April, May, October, and November), benchmarks the proposed approach against a threshold-based operational policy. Results show consistent cost reductions of approximately 15%-18% while maintaining an excellent average service level (around 99.8%-99.9). Overall, the study demonstrates that QUBO-based optimization, coupled with GPU-based solving, can provide a practically deployable decision-support tool for large-scale ATM cash logistics.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 1 canonical work pages

[1]

A Comparative Study of Integer Linear Programming and Dynamic Programming Approaches for ATM Cash Replenishment Optimization

Ozer, Ahmet Burak and Toroslu, Ismail Hakki and Karagoz, Pinar. A Comparative Study of Integer Linear Programming and Dynamic Programming Approaches for ATM Cash Replenishment Optimization. Journal of Advanced Management Science. 2020

2020
[2]

A Quantum Annealing Approach for Dynamic Multi-Depot Capacitated Vehicle Routing Problem

Hussain, Hira and Javaid, Muhammad Bilal and Khan, Faraz Shabbir and Dalal, Ahmad and Khalique, Abida. A Quantum Annealing Approach for Dynamic Multi-Depot Capacitated Vehicle Routing Problem. arXiv preprint arXiv:2005.12478. 2020

arXiv 2005
[3]

Quadratic and Higher-Order Unconstrained Binary Optimization of Railway Rescheduling for Quantum Computing

Zaman, Mujahid and others. Quadratic and Higher-Order Unconstrained Binary Optimization of Railway Rescheduling for Quantum Computing. Quantum Information Processing. 2022. doi:10.1007/s11128-022-03670-y

doi:10.1007/s11128-022-03670-y 2022
[4]

Multi-Objective Variational Quantum Optimization for Constrained Problems: An Application to Cash Management

Díez-Valle, Pablo and others. Multi-Objective Variational Quantum Optimization for Constrained Problems: An Application to Cash Management. arXiv preprint arXiv:2302.04196. 2023

arXiv 2023
[5]

Quadratization in discrete optimization and quantum mechanics

Dattani, Nike. Quadratization in discrete optimization and quantum mechanics. arXiv preprint arXiv:1901.04405. 2019. arXiv:1901.04405

arXiv 1901
[6]

and G \'o mez-Santill \'a n, C

Cruz-Reyes, L. and G \'o mez-Santill \'a n, C. and Rangel-Valdez, N. and Hern \'a ndez-Lerma, O. Optimization of ATM cash replenishment with group-demand forecasts. Expert Systems with Applications. 2015. doi:10.1016/j.eswa.2014.12.053

work page doi:10.1016/j.eswa.2014.12.053 2015
[7]

Reformulating a Problem

D-Wave Systems Inc. Reformulating a Problem
[8]

Penalty Models

D-Wave Systems Inc. Penalty Models

Show all 12 references

[9]

European Journal of Operational Research , author=

A new approach for modeling and solving set packing problems , year=. European Journal of Operational Research , author=
[10]

MegaQUBO

Data Reply. MegaQUBO
[11]

Algoritmi Quantum inspired per ENEL

Quantum-inspired algorithms for ENEL. Algoritmi Quantum inspired per ENEL
[12]

Unlocking Predictive Power in Corporate Credit Markets

Unlocking Predictive Power in Corporate Credit Markets. Unlocking Predictive Power in Corporate Credit Markets

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

A Comparative Study of Integer Linear Programming and Dynamic Programming Approaches for ATM Cash Replenishment Optimization

Ozer, Ahmet Burak and Toroslu, Ismail Hakki and Karagoz, Pinar. A Comparative Study of Integer Linear Programming and Dynamic Programming Approaches for ATM Cash Replenishment Optimization. Journal of Advanced Management Science. 2020

2020

[2] [2]

A Quantum Annealing Approach for Dynamic Multi-Depot Capacitated Vehicle Routing Problem

Hussain, Hira and Javaid, Muhammad Bilal and Khan, Faraz Shabbir and Dalal, Ahmad and Khalique, Abida. A Quantum Annealing Approach for Dynamic Multi-Depot Capacitated Vehicle Routing Problem. arXiv preprint arXiv:2005.12478. 2020

arXiv 2005

[3] [3]

Quadratic and Higher-Order Unconstrained Binary Optimization of Railway Rescheduling for Quantum Computing

Zaman, Mujahid and others. Quadratic and Higher-Order Unconstrained Binary Optimization of Railway Rescheduling for Quantum Computing. Quantum Information Processing. 2022. doi:10.1007/s11128-022-03670-y

doi:10.1007/s11128-022-03670-y 2022

[4] [4]

Multi-Objective Variational Quantum Optimization for Constrained Problems: An Application to Cash Management

Díez-Valle, Pablo and others. Multi-Objective Variational Quantum Optimization for Constrained Problems: An Application to Cash Management. arXiv preprint arXiv:2302.04196. 2023

arXiv 2023

[5] [5]

Quadratization in discrete optimization and quantum mechanics

Dattani, Nike. Quadratization in discrete optimization and quantum mechanics. arXiv preprint arXiv:1901.04405. 2019. arXiv:1901.04405

arXiv 1901

[6] [6]

and G \'o mez-Santill \'a n, C

Cruz-Reyes, L. and G \'o mez-Santill \'a n, C. and Rangel-Valdez, N. and Hern \'a ndez-Lerma, O. Optimization of ATM cash replenishment with group-demand forecasts. Expert Systems with Applications. 2015. doi:10.1016/j.eswa.2014.12.053

work page doi:10.1016/j.eswa.2014.12.053 2015

[7] [7]

Reformulating a Problem

D-Wave Systems Inc. Reformulating a Problem

[8] [8]

Penalty Models

D-Wave Systems Inc. Penalty Models