{"id":"594813bc-3524-4dd7-8f70-bf40d80b7864","arxiv_id":"2412.13769","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"QuLTSF, a linear model with a 10-qubit variational circuit inserted between two linear layers, reports improved MSE and MAE on the Weather dataset across horizons 96 to 720.","lead":"A team proposes QuLTSF, a hybrid quantum-classical model that adds a small variational quantum circuit to a linear forecasting model. On one weather dataset, it reports lower forecast errors than the classical linear baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported advantage hinges entirely on comparing QuLTSF's single-run results to baselines copied from Zeng et al. without evidence that preprocessing, normalization, and split are identical; without error bars the small margins are not established.","rationale":"The reader identified the load-bearing weakness as the comparability of QuLTSF's self-reported results with baselines copied from Zeng et al., compounded by the absence of error bars. My reading of the paper agrees: the empirical claim rests entirely on Table 1, and the paper provides no evidence that the evaluation protocols match beyond the sequence length and prediction horizon. The small margins at T=720 make the lack of repeated-seed statistics especially consequential. I also note that the QuLTSF model has far fewer trainable parameters than the classical Linear model due to the 10-qubit bottleneck, so even if the numbers were generated under an identical pipeline, the improvement might come from implicit regularization rather than the quantum layer; however, the paper only claims superiority over the listed baselines, so the comparability issue is the primary blocker. The reader's CONDITIONAL verdict is appropriate: the requested re-runs and error bars are feasible and would settle the concern. Hence no verdict adjustment is needed.","tokens_in":7810,"tokens_out":4882,"duration_ms":48583,"concrete_test":"Run the released QuLTSF code and the Linear, NLinear, and DLinear baselines from Zeng et al. under a single script with identical data loading, normalization, split boundaries, and evaluation protocol on the Weather dataset; repeat each with at least five random seeds and report mean plus standard deviation. If the mean MSE advantage of QuLTSF over the best classical baseline falls within the seed noise, or if the ordering changes under either pipeline, the central claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim, stated in Section 5.5 as 'Our proposed QuLTSF outperform all the baseline models in all 4 cases', depends on Table 1, whose footnote says 'Other results are from (Zeng et al., 2023)'. The QuLTSF numbers were produced by the authors, but the classical baselines were not re-run under the same pipeline. The paper does not specify the normalization, data scaling, or exact split boundaries used for QuLTSF, and Section 5.1 only gives a 70/20/10 split. If the QuLTSF preprocessing differs from the pipeline used to generate the quoted baseline numbers, the MSE/MAE values are not comparable even though L and T match. This is not a hypothetical concern: the reported gains are small in several settings, e.g., at T=720 the MSE advantage over DLinear is 0.315 vs 0.323 and the MAE advantage over NLinear is 0.346 vs 0.348, so seed-to-seed or preprocessing variation could change the ranking. The absence of any error bars or repeated-seed statistics makes it impossible to tell whether the difference is systematic. Public code is a positive step, but it does not establish comparability with numbers taken from another paper. Therefore the headline claim of universal superiority is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QuLTSF, a hybrid quantum-classical model for long-term time series forecasting (LTSF), in which a classical linear layer maps the input to a 2N-dimensional vector, a 10-qubit hardware-efficient ansatz with K=3 layers processes an amplitude-encoded state, and a final classical linear layer outputs the T-step forecast. The authors evaluate QuLTSF on the Weather dataset with a fixed look-back length L=336 and prediction lengths T in {96,192,336,720}, and report that it outperforms Linear, NLinear, DLinear, FEDformer, Autoformer, and Informer on both MSE and MAE in all four settings (Section 5.5, Table 1). They also report experiments with varying sequence length L (Figures 2-3). The central claim is that inserting a variational quantum layer into a simple linear forecasting model yields consistent gains over classical linear and transformer baselines.","tokens_in":8135,"tokens_out":3936,"duration_ms":37172,"significance":"If the empirical claim were established, the paper would provide evidence that a variational quantum circuit can improve a classical linear forecasting pipeline, which is a useful contribution to the emerging QML-for-time-series literature. The model itself is transparent and easy to reproduce, the equations in Section 4 are coherent, and the authors provide public code, which are strengths. However, the significance is currently limited by the evidence base: the evaluation uses a single dataset, a single run per configuration, and baseline numbers quoted from prior work rather than recomputed under the same pipeline. Because several of the reported margins are small (e.g., T=720 MSE 0.315 vs 0.323 for DLinear; MAE 0.346 vs 0.348 for NLinear), the paper's central claim is not yet supported at the standard expected for a 'state-of-the-art' comparison. The architecture and training procedure are sound, but the empirical validation needs substantial strengthening before the superiority claim can be accepted.","major_comments":[{"comment":"The headline claim that \"QuLTSF outperform all the baseline models in all 4 cases\" rests on a comparison between QuLTSF results produced by the authors and baseline numbers quoted from (Zeng et al., 2023), as stated in the Table 1 footnote. The paper does not specify the normalization, scaling, or exact train/validation/test split boundaries used for QuLTSF; Section 5.1 only gives a 70/20/10 split. If the QuLTSF preprocessing pipeline differs from that used to generate the quoted baseline numbers, the MSE and MAE values are not comparable even when L and T match. The authors should re-run all baselines in their own pipeline, or at minimum demonstrate that their preprocessing and split are identical to those in the source of the quoted numbers.","section":"Table 1, Section 5.5"},{"comment":"No error bars, confidence intervals, or repeated-seed statistics are reported. Section 5.4 lists hyperparameters (N=10, K=3, batch size 16, learning rate 0.0001) but does not state how many random seeds or initializations were used. This matters because several of the reported gains are small: at T=720, QuLTSF's MSE is 0.315 versus DLinear's 0.323, and its MAE is 0.346 versus NLinear's 0.348. Without uncertainty quantification, the observed differences can easily be within seed-to-seed or optimization-run variation, so the claim of consistent superiority is not established. The authors should provide multiple-seed results with mean and standard deviation, and ideally a significance test.","section":"Section 5.4, Table 1"},{"comment":"The discussion introduces comparisons to PatchTST and MTST that qualify the earlier universal-superiority claim. For L=336, T=336, the paper reports QuLTSF MSE 0.248 versus PatchTST 0.249 and MTST 0.246; QuLTSF is not the best among these. The text says QuLTSF is \"outperforms or comparable\" to these models, but this contradicts the unqualified wording in Section 5.5 that QuLTSF outperforms all baselines in all cases. The authors should state the precise scope of the claim (e.g., limited to the baselines in Table 1) and ensure that comparisons to models outside Table 1 are accompanied by the same comparability evidence as the main table.","section":"Section 5.6"},{"comment":"The conclusion states that \"incorporating a hidden quantum layer yielded additional improvements,\" but no ablation is provided that isolates the effect of the quantum layer. QuLTSF contains trainable parameters in the input linear layer, the VQC, and the output linear layer, whereas the Linear baseline is a single linear map. It is therefore possible that any measured improvement (if real) comes from additional model capacity or the nonlinearity introduced by the input projection, rather than from the quantum circuit per se. A controlled comparison against a purely classical model with a comparable nonlinear hidden layer and a similar parameter count is needed to support the attribution in the conclusion.","section":"Section 4, Equations (2)-(4); Section 6"}],"minor_comments":[{"comment":"There are several typographical errors: \"import problem\" should be \"important problem,\" and \"strain's the models predictive capacity\" should be \"strains the model's predictive capacity.\" These should be corrected in revision.","section":"Section 1"},{"comment":"The figures plot MSE for varying L but do not show error bars, markers, or legend entries that distinguish repeated runs. It is unclear whether each curve is a single run, and the y-axis ranges make the small differences between curves difficult to assess. Adding standard-deviation bands or at least markers and a clear caption would help.","section":"Figures 2 and 3"},{"comment":"The abstract says \"extensive experiments on a widely used weather dataset,\" but only one dataset is used. Either the wording should be softened to reflect the single-dataset evaluation, or additional datasets should be added.","section":"Section 5.1"},{"comment":"Some reference names are typeset with spacing artifacts, e.g., \"V aswani\" for Vaswani and \"Y ou\" for You. These formatting issues should be fixed in the final manuscript.","section":"References"},{"comment":"When citing numbers from (Zhang et al., 2024) for PatchTST and MTST, the text says \"Table 2 in (Zhang et al., 2024)\" but the reader must cross-check Table 1 of this paper. It would be clearer to include these comparisons in a dedicated table with the same provenance and preprocessing details as Table 1.","section":"Section 5.6"}],"recommendation":"major_revision","confidential_remarks":"The paper proposes a straightforward hybrid linear-VQC architecture, and its main contribution is empirical. The core experiment is not yet convincing because of the reliance on baseline numbers from another paper without a shared pipeline, the absence of error bars, and the small margins. These issues are fixable by re-running baselines, adding multiple seeds, and including an ablation. If the authors perform this additional work, the paper could become a reasonable applied QML contribution. There is no circularity concern: the only self-cited reference (Binder et al., 2018) is background on recurrent quantum processes and does not enter the derivations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate first application of a hybrid linear-VQC model to long-term time series forecasting, with clear writing and working code. But the central superiority claim is not yet supported. Table 1 takes the linear and transformer baseline numbers straight from Zeng et al. 2023, and the QuLTSF runs are single-seed. The margins at T=720 (0.315 vs 0.323 MSE over DLinear) are too small to survive unknown preprocessing differences or seed noise. So I read the paper as a promising pilot rather than a demonstrated result.\n\nWhat's genuinely new is the application, not the components. Amplitude embedding, a hardware-efficient ansatz, and linear input/output layers are all standard QML pieces, but nobody has put them together for LTSF under the channel-independence setting. The experiments cover the standard four horizons and also vary L, which is more than many first-application papers do. The code on GitHub is a real asset.\n\nThe soft spots are the usual ones, but they matter here. No error bars, no multiple seeds, one dataset, and the baselines are not re-run under the same preprocessing. The paper doesn't specify normalization or split boundaries for QuLTSF beyond a 70/20/10 split, so the comparison is fragile. There's also no ablation: replacing the VQC with a random nonlinear layer or removing it entirely would tell you whether the quantum layer is doing anything. The hyperparameters (N=10, K=3, lr) are declared without a search, and given the small gains that's a real worry. To the authors' credit, the discussion section honestly admits that QuLTSF is only comparable to PatchTST and MTST at two settings; that undercuts the 'outperform all baselines' phrasing but also shows they're not cherry-picking the comparison.\n\nWho should read it: people working on QML for time series, and anyone who wants a compact example of how easy it is to overclaim from a single-table comparison. The referee should not desk-reject it; the topic is timely and the code makes it checkable. But the revision must include multi-seed statistics, re-run baselines under the same pipeline, at least one more dataset, and an ablation. Without those, the headline claim should not stand.","headline":"First QML shot at LTSF with public code, but the 'wins in all 4 cases' claim rests on borrowed baselines and single-run numbers.","tokens_in":8624,"tokens_out":2353,"would_cite":false,"duration_ms":21470,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid quantum-classical model called QuLTSF is claimed to beat classical linear and transformer baselines on the Weather benchmark for long-term time series forecasting.","keywords":["long-term time series forecasting","quantum machine learning","variational quantum circuit","hybrid quantum-classical model","weather dataset","amplitude embedding","linear baselines","time series forecasting"],"falsifier":"Re-run QuLTSF and the Linear, NLinear, and DLinear baselines on the Weather dataset using the same code pipeline, same 70/20/10 data split, same normalization, and multiple random seeds; if QuLTSF does not consistently achieve lower MSE and MAE than all three linear baselines across the four prediction horizons, the paper's central claim fails.","tokens_in":7630,"feed_emoji":"🌦️","tokens_out":3668,"duration_ms":32458,"temperature":0.7,"pith_summary":"The paper tries to establish that adding a variational quantum circuit as a hidden layer to a simple linear forecasting model improves long-term time series forecasting. On the widely used Weather dataset, the proposed QuLTSF model reports lower mean squared error and mean absolute error than the classical linear models Linear, NLinear, and DLinear, as well as several transformer-based models, across four prediction horizons. If the claim holds, it would show that quantum machine learning can add practical value in a forecasting task where simple linear models had already outperformed much more complex transformers. The authors position this as the first application of quantum machine learning to long-term time series forecasting.","feed_headline":"Quantum layer improves long-term weather forecasts over linear models","feed_subtitle":"QuLTSF, a hybrid quantum-classical model, reports lower MSE and MAE than six baselines on the Weather benchmark.","key_machinery":"The central object is the QuLTSF architecture itself: a hybrid quantum-classical model in which a variational quantum circuit (a hardware-efficient ansatz composed of trainable single-qubit gates and fixed entangling CNOT gates) sits between two classical linear layers. Amplitude embedding converts the classical linear layer's output into a quantum state, and the expectation values of Pauli-Z observables on each qubit form the quantum layer's output, which a final linear layer maps to the forecast. This design directly extends the simple Linear model of Zeng et al. by inserting a parameterized quantum feature map, and all parameters of the quantum and classical layers are trained jointly with an Adam optimizer to minimize MSE.","core_discovery":"QuLTSF is a simple three-part architecture: a classical input layer maps each univariate series of length L to a 2N-dimensional vector, amplitude embedding encodes that vector into the state of N=10 qubits, a hardware-efficient ansatz with K=3 layers applies parametrized single-qubit rotations and fixed CNOT entangling gates, and the Pauli-Z expectation values are passed through a classical output layer to produce the T-step forecast. With look-back length L=336 and prediction lengths T in {96, 192, 336, 720}, the authors report that QuLTSF outperforms all six baselines—Linear, NLinear, DLinear, FEDformer, Autoformer, and Informer—in every setting. Additional experiments with varying sequence lengths for T=96 and T=720 show QuLTSF consistently beating the three classical linear baselines. The paper also notes that in two settings QuLTSF is comparable to or slightly better than the more recent PatchTST and MTST transformer models.","pith_inferences":["Editorial inference: The paper does not report error bars or multiple-seed statistics, so a natural next step is to re-run QuLTSF and the classical baselines under an identical preprocessing pipeline to see whether the MSE gap is stable or within run-to-run noise.","Editorial inference: If the quantum layer's benefit is real, the likely mechanism is that the variational circuit supplies a rich nonlinear feature map that the purely linear model lacks; this suggests testing the same architecture on other high-dimensional regression tasks where linear baselines are strong.","Editorial inference: The choice of ansatz, the number of qubits, and the number of circuit layers may strongly affect the result, so a systematic ansatz search over the Weather dataset could reveal whether the reported advantage is tied to this specific circuit or generalizes across circuit designs."],"forward_implications":["If the reported results are correct, a relatively small 10-qubit variational circuit can improve forecasting accuracy over purely classical linear models on at least one standard benchmark.","The claimed improvements appear across all tested prediction lengths, suggesting the benefit is not limited to one specific horizon.","QuLTSF's performance being comparable to or better than PatchTST and MTST in limited settings suggests a quantum hidden layer may be competitive with more elaborate classical preprocessing and attention architectures.","The paper opens the possibility of applying hybrid quantum-classical models more broadly to long-term forecasting problems in weather, finance, and other domains."],"supporting_citations":[{"why":"Provides the Linear, NLinear, and DLinear baselines and the transformer baseline numbers quoted in Table 1, which the central claim directly extends and beats.","marker":"(Zeng et al., 2023)"},{"why":"Defines long-term time series forecasting as prediction length greater than 48 and supplies the Informer baseline.","marker":"(Zhou et al., 2021)"},{"why":"Supplies the Autoformer baseline and the historical context of transformer-based models for long-term forecasting.","marker":"(Wu et al., 2021)"},{"why":"Supplies the FEDformer baseline used in the comparison table.","marker":"(Zhou et al., 2022)"},{"why":"Provides the PennyLane software framework used to implement and jointly train the hybrid quantum-classical model.","marker":"(Bergholm et al., 2018)"},{"why":"Supplies the amplitude embedding method used to encode classical data into the quantum circuit.","marker":"(Schuld and Petruccione, 2021)"},{"why":"Supplies the hardware-efficient ansatz concept used as the variational quantum circuit in QuLTSF.","marker":"(Simeone, 2022)"}],"fun_headline_variants":["QuLTSF: quantum model beats linear baselines on weather forecasting","Quantum ML forecast edges out linear models on weather data","Hybrid quantum model tops linear forecasts in long-term weather","QuLTSF: quantum advantage in weather time-series forecasting","Quantum layer lowers error in long-term weather predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that the baseline numbers quoted from Zeng et al. (2023) were produced under the same data split, normalization, and evaluation protocol used for QuLTSF, since the baselines were not re-run by the authors.","fun_headline_variants_meta":{"raw":{"variants":["QuLTSF: quantum model beats linear baselines on weather forecasting","Quantum ML forecast edges out linear models on weather data","Hybrid quantum model tops linear forecasts in long-term weather","QuLTSF: quantum advantage in weather time-series forecasting","Quantum layer lowers error in long-term weather predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000453,"raw_usage":{"total_tokens":2280,"prompt_tokens":949,"completion_tokens":1331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1249}},"tokens_in":565,"tokens_out":1331,"duration_ms":9602,"temperature":1.0,"reasoning_tokens":1249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:48:30.073923+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run QuLTSF and the Linear, NLinear, and DLinear baselines on the Weather dataset using the same code pipeline, same 70/20/10 data split, same normalization, and multiple random seeds; if QuLTSF does not consistently achieve lower MSE and MAE than all three linear baselines across the four prediction horizons, the paper's central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FEDformer baseline used in the comparison table."},{"cited_title":"and Petruccione, F","cited_arxiv_id":null,"evidence_quote":"Supplies the amplitude embedding method used to encode classical data into the quantum circuit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the hardware-efficient ansatz concept used as the variational quantum circuit in QuLTSF."}],"review_version":1}