{"id":"3b5d9005-a311-40d7-bfe1-14cf28c40b38","arxiv_id":"2412.05615","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A parameterized quantum circuit trained with classical optimizers achieved the lowest test MSE on a synthetic sales forecast task, while a variational quantum linear solver failed to produce usable forecasts.","lead":"This paper tests two quantum computing approaches for time-series forecasting on a synthetic sales dataset, comparing a parameterized quantum circuit (PQC) and a variational quantum linear solver (VQLS) against classical baselines. It finds that the PQC model achieved the lowest test error among the methods tried, while VQLS produced essentially random forecasts on the main dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed PQC advantage rests on a single unreplicated train/test split; without repeated runs or interval estimates, the Table 1 ordering may be initialization luck.","rationale":"The reader's weakest assumption correctly identifies the single-split, no-repeated-runs design as the load-bearing weakness: the entire positive conclusion about quantum-model advantage depends on one unreplicated comparison. I agree that this is the most important issue. I considered other potential concerns, such as the unspecified measurement observable for the PQC, the questionable O(n^2) Pauli-decomposition remark in Section 2.3.1, and the absence of code/data release, but none of these is as directly fatal to the paper's central forward-looking claim as the lack of any measure of variability. A single MSE table entry cannot distinguish a genuine advantage from initialization luck or a favorable test period. The paper is exploratory, its limitations are largely disclosed, and the VQLS negative result is reported honestly, so the appropriate verdict remains CONDITIONAL rather than REJECT. The suggested concrete test--repeated runs with different seeds and test windows--would settle whether the reported ordering is reproducible and therefore whether the 'potential advantage' framing is justified.","tokens_in":6826,"tokens_out":4352,"duration_ms":44824,"concrete_test":"Obtain or rerun the experiments with fixed seeds; then run each of the four models at least 30 times with different random parameter initializations on the same train/test split and report mean and 95% bootstrap CI of test MSE. Additionally, vary the test window (e.g., rolling-origin evaluation using test years 2022 and 2023) and check whether PQC-COBYLA remains below the linear model's interval. If the intervals overlap or the ordering flips in a majority of runs, the Table 1 ordering is not evidence of advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim is the test-set ordering in Table 1: PQC-COBYLA MSE 0.02106, PQC-L-BFGS-B 0.04418, linear 0.05177, NN 0.05767. Section 3 presents this ordering as evidence that the hybrid quantum-classical approach 'shows promise' and indicates 'a potential advantage.' The load-bearing assumption is that this ordering reflects the methods themselves, not one particular random initialization and one particular test window. The paper reports no seeds, no repeated runs, and no error bars; the dataset is a single synthetic monthly sales series of about 60 points, split once into one training and one test period (Figure 1a). Both PQC optimizers start from random circuit parameters, so the exact MSE values are seed-dependent. A gap of 0.021 vs 0.052 on a short, autocorrelated series could easily flip under a different initialization or a different test period. The conclusion 'promise' would require showing that the ordering is robust, not just that it occurs once. The paper's own caveat that the classical baselines are unregularized weakens the baselines, but the unreplicated comparison is the more fundamental issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript explores two hybrid quantum-classical methods for time-series forecasting on a synthetic monthly sales dataset. The first method trains a parameterized quantum circuit (PQC) to minimize MSE, using COBYLA and L-BFGS-B optimizers, and compares its test MSE with a classical linear regression and a small neural network. The second method applies the Variational Quantum Linear Solver (VQLS) to the normal equations from a sliding-window regression. On a single train/test split, the PQC with COBYLA reports the lowest test MSE (0.02106), followed by PQC with L-BFGS-B (0.04418), linear regression (0.05177), and the neural network (0.05767). VQLS is reported to give unsatisfactory results on the general synthetic series and is only qualitatively described as tracking monotonic trends. The authors conclude that the hybrid quantum-classical approach shows promise while acknowledging that the classical baselines lack regularization.","tokens_in":7061,"tokens_out":4437,"duration_ms":62429,"significance":"If the reported ordering of test MSE were robust, the result would be a modest but useful data point for NISQ-based forecasting, particularly because the PQC is trained in simulation on a realistic (though synthetic) sales series. The manuscript is honest in reporting the failure of VQLS on the general dataset and in stating that regularized classical models would likely perform better. However, the central positive claim currently rests on one unreplicated train/test split, on weak unregularized baselines, and on an incomplete experimental description. The paper does not supply code, data, seeds, error bars, or repeated trials, and the VQLS derivation contains a normalization inconsistency. These issues limit the strength of the evidence but are addressable in a revision.","major_comments":[{"comment":"The test-set MSE ordering in Table 1 is the sole evidence for the claim of a 'potential advantage', yet the experiments are performed on a single train/test split with no repeated runs, no seeds, and no error bars. Both PQC optimizers start from random parameters, so the reported values, including the gap between COBYLA (0.02106) and the linear model (0.05177), could be due to initialization luck or to the particular test window shown in Figure 1. The authors should run many repetitions with different seeds and initializations, report the distribution or at least the mean and confidence interval of the test MSE, and state the exact split and seeds used.","section":"Section 3, Table 1"},{"comment":"The comparison uses an unregularized classical linear model and a small neural network with no early stopping, dropout, or other regularization, and the text itself concedes that these models 'with proper regularization would likely lead to good test set MSE scores'. An advantage over deliberately weak baselines is not evidence of quantum utility. The revision should include properly regularized classical baselines, such as ridge regression, ARIMA/ETS, or a neural network with early stopping, with hyperparameters selected on the training set, and report their test MSE under the same protocol.","section":"Section 3, Table 1 and text after it"},{"comment":"The PQC experiment is not described with enough detail to be reproduced or evaluated: the number of qubits, ansatz depth, parameter initialization strategy, optimizer stopping criteria, learning rate for L-BFGS-B, and COBYLA iteration limits are not stated. More importantly, the text says the model minimizes the squared difference between an observable's expectation value and the labels, but it does not specify which observable is measured or how expectation values are mapped to real-valued forecasts on the original scale. These details are load-bearing because the entire positive result depends on the PQC implementation.","section":"Section 2.1 and Figure 2"},{"comment":"The cost function derivation is inconsistent: the text defines |ψ⟩ = M|x(ϕ)⟩, then writes C = ⟨ψ|ψ⟩ − |⟨b|ψ⟩|² and asserts that the first term simplifies to 1 because |ψ⟩ is a unit vector. However, M|x(ϕ)⟩ is not unit-norm in general, so C does not reduce to 1 − |⟨b|ψ⟩|². The authors must either normalize |ψ⟩ or use the correct unnormalized cost, and they should check whether this affects the reported VQLS results and the rescaling formula w ≈ (‖b‖/‖A|w⟩‖)|w⟩.","section":"Section 2.3.4, Cost Function"},{"comment":"The failure of VQLS is attributed to an 'extremely high condition number' and to the cost of the Pauli decomposition, but no condition number, matrix size, or numerical result for VQLS is reported. The reader cannot assess whether the failure stems from the ill-conditioning of A, from the normalization error in the cost function, or from the optimizer. The authors should report the condition number of A for the tested window sizes, the VQLS test MSE or a comparison against random predictions, and, ideally, the same quantities for the monotonic and geometric time series that are said to work.","section":"Section 3, VQLS paragraph"}],"minor_comments":[{"comment":"The caption states that the shaded area is used for both training and scaling adjustments; please clarify explicitly that the scaling constants are computed using only the training portion and are then applied to the test portion, since the final sentence of the caption indicates test data may fall outside the scaled range.","section":"Figure 1 caption"},{"comment":"The sentence 'We did VQLS experiments with 2, 3, and 4 qubits, that represents time series of length n of 4, 9, and 16' is unclear about the relationship between the qubit count, the series length, and the window size m used in the normal equations; please define the matrix dimension explicitly.","section":"Section 2.3.3"},{"comment":"The text contains inconsistencies in names and formatting, such as 'Forse AI' versus 'Force AI', the author name 'M¯ arti¸nˇs K¯ alis', and the omitted circuit measurement details in Figure 4; the manuscript would benefit from a careful proofreading pass.","section":"Throughout"},{"comment":"The statement that 'in reality, we need at most n^2 non-zero terms of M' is attributed to a reference about teleportation and dense coding, which is not obviously the source of this linear-algebra fact; please provide a more direct citation or a derivation.","section":"Section 2.3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is exploratory and would be suitable for a quantum-machine-learning venue only if the central comparison is made statistically solid and the methods are fully described. The absence of repeated runs and of regularized classical baselines is the main barrier; both are addressable. I would also encourage the editor to ask for the code and data or, failing that, a detailed experimental appendix, because the current manuscript does not allow reproduction of the reported MSE values."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a small, honest empirical note. The authors train a PQC on a synthetic monthly sales series and report a lower test MSE than a linear model and a small neural network. They also try VQLS and report that it fails. The PQC result is the kind of thing people in this area want to see, but it rests on a single unreplicated split with no repeated runs and no error bars. The paper itself admits the classical baselines are unregularized and would likely improve.\n\nWhat it does well: it is clearly written, the preprocessing (differencing, scaling) is sensible, and it reports the negative VQLS result rather than burying it. The candor about the baselines and the overfitting of the neural network is genuine. That honesty is worth something.\n\nThe bigger problem is in the VQLS section. The cost function is defined as C = ⟨ψ|ψ⟩ − |⟨b|ψ⟩|² with |ψ⟩ = M|x(φ)⟩, and the authors then say |ψ⟩ is a unit vector so the first term simplifies to 1. That is not true. M|x⟩ is not normalized. The correct VQLS cost includes a denominator ⟨x|M†M|x⟩. Omitting it lets the cost be trivially minimized by shrinking the norm of x, which likely explains the 'indistinguishable from random' results. The paper blames the high condition number, but the cost bug is the more probable culprit. This should be corrected before any conclusion about VQLS is drawn.\n\nThe main PQC comparison also needs repetition. The difference between 0.021 and 0.052 on a short autocorrelated series could flip with another random initialization or a different test window. No seeds are reported. That is a real limitation, and it cuts against the 'potential advantage' phrasing.\n\nWho is this for? Someone curious about whether PQC training behaves as advertised on small time series. It is not a breakthrough, and it does not claim to be. I would send it to a serious referee because the question is legitimate and the authors are transparent, but the VQLS section needs a fix and the PQC claim needs repeated runs. If the authors can show the gap survives multiple runs and correct the cost, this could become a useful short report.","headline":"An honest but underpowered empirical note: one PQC run beats two unregularized baselines on a synthetic sales series, and the VQLS negative result is likely due to a cost-function bug, not conditioning alone.","tokens_in":7598,"tokens_out":4748,"would_cite":false,"duration_ms":42287,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Q12","62M10","81P68"],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"This paper reports that a parameterized quantum circuit trained with COBYLA achieves the lowest test mean squared error on a synthetic sales-forecasting task, beating classical linear regression and a small neural network, while a…","keywords":["time-series forecasting","hybrid quantum-classical","parameterized quantum circuit","variational quantum linear solver","COBYLA","L-BFGS-B","mean squared error","NISQ"],"falsifier":"Run the same four models on the same dataset with, say, 50 different random initializations and several different train/test split points; compute the distribution of test MSE for each model. If the PQC with COBYLA does not have the lowest median test MSE, or if the differences are within the spread of the distributions, the paper's central claim that the PQC outperformed the classical baselines on this task is refuted.","tokens_in":6681,"feed_emoji":"📈","tokens_out":4068,"duration_ms":31438,"temperature":0.7,"pith_summary":"This paper asks whether hybrid quantum-classical methods can be practical for time-series forecasting, and tests two approaches on a synthetic monthly sales dataset. A parameterized quantum circuit (PQC) trained to predict next-month sales from a 12-month window is compared against classical linear regression and a small neural network. On the single test split used, the PQC with the COBYLA optimizer posted the lowest test mean squared error, with the PQC trained by L-BFGS-B second and both classical models higher. A second quantum approach, the Variational Quantum Linear Solver (VQLS), is also implemented but fails to produce reasonable forecasts on general series because the coefficient matrix is ill-conditioned; it only behaves on strictly growing or declining series. The paper's contribution is an early empirical benchmark suggesting PQC-based forecasting can be competitive, while cautioning that the advantage rests on one unreplicated split and on optimizer choice.","feed_headline":"Quantum circuit edges out classical models in sales forecast test","feed_subtitle":"A parameterized quantum circuit with COBYLA posted lowest test error; VQLS solver fails on ill-conditioned series.","key_machinery":"The central objects are the Parameterized Quantum Circuit (PQC) used as a regression model, and the Variational Quantum Linear Solver (VQLS) used to solve the normal-equation system Aw = b for sliding-window forecasting. The PQC encodes 12 past monthly values as rotation angles on 12 qubits, applies layered entangling CNOT gates and trainable RX/RY rotations, and minimizes the squared error between the expectation value of an observable and the true next value; training is done classically with L-BFGS-B or COBYLA. The VQLS approach decomposes the matrix A = X^T X into a weighted sum of tensor products of Pauli operators, prepares a normalized |b>, and uses a Hadamard test with a parameterized ansatz to minimize 1 - |<b|M|x(φ)>|^2. The paper's key finding is that the PQC machinery works well enough on this small problem to outperform the classical baselines on test MSE, while the VQLS machinery breaks down because A is ill-conditioned and the Pauli decomposition has O($n^{2}$) terms.","core_discovery":"On the paper's own terms, the central claim is that a parameterized quantum circuit with a COBYLA optimizer achieves a test-set MSE of 0.02106 on this synthetic sales-forecasting task, beating the classical linear model (0.05177) and a two-layer neural network (0.05767); a PQC with L-BFGS-B achieves 0.04418. The authors state that the PQC with COBYLA performed best on the test set, followed by PQC with L-BFGS-B. They also report that the VQLS linear-solver approach is not viable for general time-series matrices because the high condition number and the O($n^{2}$) number of Pauli terms make it ineffective; it only produces sensible extrapolations for strictly monotonic series. The paper frames these results as evidence that hybrid quantum-classical forecasting 'shows promise' but with heavy caveats about regularization, hyperparameter tuning, and the unreplicated experimental design.","pith_inferences":["The reported advantage is fragile: because the comparison rests on a single train/test split and no repeated runs, the probability that the ordering is due to initialization luck or a particular test period is not quantified; a reasonable extension would be to run 10–50 random seeds and report median and quartiles.","The neural network's overfitting suggests that a properly regularized classical model (early stopping, dropout, or weight decay) could erase the quantum advantage entirely, so the headline claim 'PQC beats classical' is better read as 'an unregularized small NN can overfit.'","The VQLS observation that monotonic series work while realistic series fail suggests a testable threshold: measure the condition number of the sliding-window matrix for various series and see if VQLS error correlates with condition number.","A practical next step is to combine the PQC with classical post-processing (e.g., residual correction) to see if the hybrid model's errors are complementary to classical models' errors."],"forward_implications":["If PQC-based forecasting is genuinely competitive, quantum models could be integrated into ensembles of classical forecasters to diversify errors and provide regularization.","The success of COBYLA over L-BFGS-B suggests optimizer choice is a dominant factor in PQC training outcomes; future work should systematically compare optimizers.","The VQLS failure on ill-conditioned matrices implies that quantum linear-solver approaches need preconditioning or orthogonalization (e.g., Gram-Schmidt) before they can be applied to time-series data.","The neural network's severe overfitting (training MSE 0.00003 vs test MSE 0.05767) indicates that simple classical baselines with proper regularization would likely close or invert the apparent quantum advantage.","The work motivates larger-scale benchmarks with repeated cross-validation to establish whether the observed test-MSE ordering is stable."],"supporting_citations":[{"why":"Supplies the parameterized quantum circuit architecture used as the first forecasting model.","marker":"[4]"},{"why":"Defines the Variational Quantum Linear Solver and its cost function, which the paper adapts for time-series regression.","marker":"[3]"},{"why":"Provides the Pauli-matrix decomposition used to represent the normal-equation matrix in the VQLS approach.","marker":"[5]"},{"why":"Supports the claim that an arbitrary n-by-n matrix requires O(n^2) Pauli terms, contributing to the VQLS failure.","marker":"[6]"}],"fun_headline_variants":["COBYLA-tuned quantum circuit posts lowest forecast error","Quantum circuit with COBYLA edges classical models in forecast test","VQLS solver fails on complex series; PQC with COBYLA succeeds","Quantum forecasting: COBYLA circuit best, VQLS not viable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the single train/test split of the synthetic sales series is representative, so the observed test-MSE ordering is not a statistical fluke of initialization or the particular test period.","fun_headline_variants_meta":{"raw":{"variants":["COBYLA-tuned quantum circuit posts lowest forecast error","Quantum circuit with COBYLA edges classical models in forecast test","VQLS solver fails on complex series; PQC with COBYLA succeeds","Quantum forecasting: COBYLA circuit best, VQLS not viable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001063,"raw_usage":{"total_tokens":4391,"prompt_tokens":813,"completion_tokens":3578,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":3502}},"tokens_in":429,"tokens_out":3578,"duration_ms":24792,"temperature":1.0,"reasoning_tokens":3502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:31:31.415279+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same four models on the same dataset with, say, 50 different random initializations and several different train/test split points; compute the distribution of test MSE for each model. If the PQC with COBYLA does not have the lowest median test MSE, or if the differences are within the spread of the distributions, the paper's central claim that the PQC outperformed the classical baselines on this task is refuted.","supporting_citations":[{"cited_title":"Parameterized quantum circuits as ma chine learning models","cited_arxiv_id":null,"evidence_quote":"Supplies the parameterized quantum circuit architecture used as the first forecasting model."},{"cited_title":"A probabilistic and information theoretic inter- pretation of quantum evolutions","cited_arxiv_id":null,"evidence_quote":"Provides the Pauli-matrix decomposition used to represent the normal-equation matrix in the VQLS approach."}],"review_version":1}