{"id":"2b4beeda-52f2-4b9f-bd9e-16ae0a6187d4","arxiv_id":"2412.14175","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"BiTSA is an interactive visualization tool that wraps time-series forecasting models (DLinear, PatchTST, One-Fits-All, etc.) for building energy analytics, with a small offline benchmark on two building datasets.","lead":"This paper describes BiTSA, a dashboard that combines pre-trained time-series forecasting models with interactive visualizations to help building managers understand IoT energy data and plan ahead. A smart generalist might read it to see how forecasting research is being packaged for real building operations, though the tool's promised benefits are not yet tested.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central 'real-time insights' promise is contradicted by its own Section III-E limitation: models are pre-trained offline, so 'immediate, data-driven actions' cannot be delivered as claimed.","rationale":"The reader's weakest assumption focused on untested human decision-making and model transferability from 2-month histories to live BMS data. My concern overlaps with the reader's point about offline-learning lag, but I identify a sharper, internal logical problem: the paper's own architecture and Section III-E explicitly contradict the abstract's 'real-time insights' promise. This makes the central claim unsupported even before considering user behavior or deployment. A real-time system would require online model updates or very frequent retraining, neither of which is described. The paper could be salvaged by reframing the contribution as a dashboard with periodically updated forecasts, or by adding an online-learning component and measuring deployment latency. Since the reader already recommends REJECT and my concern strengthens that conclusion without changing the direction, the verdict remains UNCHANGED. The concrete test I propose directly checks the claimed real-time property against the implementation, settling whether the contradiction is a matter of wording or a genuine technical gap.","tokens_in":5354,"tokens_out":2294,"duration_ms":24804,"concrete_test":"Set up a local BiTSA instance (or inspect the linked GitHub repository's back-end modules) with the BTS-B dataset, feed a simulated live 10-minute BMS data stream, and measure (i) end-to-end latency from a new meter reading arriving via the BMS API to that value appearing in the dashboard, and (ii) the model weight refresh interval. If the refresh interval is not at the 10-minute data cadence or if the dashboard update period exceeds the decision horizon used in the paper (e.g., H=12 steps, 2 hours), then the system does not provide 'real-time insights' according to the authors' own architectural description and stated limitation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Abstract and Conclusion promise that BiTSA will let building managers 'take immediate, data-driven actions based on real-time insights.' This is the paper's core value proposition. However, the system description and the authors' own stated limitations undermine exactly this premise. Section II-B says the Pre-trained Models 'will be pre-trained offline using analytics-ready data' and the Results database is 'expected to be updated by each regular period'—not as a continuous stream. Section III-E then states that BiTSA's 'reliance on offline learning... often leads to lag in adapting to new data and evolving conditions' and that the tool 'may lack the ability to provide real-time insights.' These are not external critiques; they are internal admissions. If the models are pre-trained on historical 10-minute data and updated only periodically, the dashboard cannot provide genuinely real-time forecasts at the BMS data cadence. The central claim, as stated, is therefore unsupported by the paper's own architecture. No user study, deployment, or online-learning evaluation is presented to close this gap. The forecasting benchmark, while internally consistent, addresses model accuracy on historical data, not the real-time operational benefit promised to building managers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes BiTSA, an interactive visualization tool for building energy analytics, which integrates several time-series forecasting models (DLinear, PatchTST, Informer, iTransformer, and One-Fits-All) behind a browser-based dashboard. The authors report an offline evaluation of these models on two building IoT datasets (BTS-B and BLDG59) using MAE, MSE, and SMAPE, and they argue that the tool will close the gap between research and practice by letting building managers take immediate, data-driven actions from real-time insights. The paper also includes a demo video, architecture diagrams, and UI screenshots. The central claim is that the dashboard empowers non-expert building managers to act on forecasting outputs in real time.","tokens_in":5519,"tokens_out":4861,"duration_ms":51615,"significance":"If validated, BiTSA could make advanced time-series forecasting accessible to building managers without data-science expertise, and the use of standard public datasets and multiple baseline models is commendable. However, the current manuscript does not evaluate the tool itself: the forecasting benchmarks address model accuracy on static historical data, not the promised operational benefit of the dashboard, and the paper's own stated limitations in Section III-E contradict the abstract's real-time claims. The paper's contribution is therefore a design proposal plus a model benchmark, not an empirically supported claim of manager empowerment. The reproducible benchmarking and public demo video are strengths, but they do not substantiate the core value proposition.","major_comments":[{"comment":"The abstract and conclusion claim that BiTSA enables building managers to 'take immediate, data-driven actions based on real-time insights,' but the system architecture described in Section II-B relies on offline pre-training and a Results database 'updated by each regular period,' and Section III-E explicitly admits that 'reliance on offline learning... often leads to lag in adapting to new data and evolving conditions' and that BiTSA 'may lack the ability to provide real-time insights.' This is a load-bearing internal contradiction. The authors must either implement an online-learning or streaming-update mechanism and evaluate it, or substantially revise the real-time and 'immediate' claims throughout the paper to accurately describe periodic offline analytics.","section":"Abstract, Conclusion, Section II-B, Section III-E"},{"comment":"The paper's central value proposition is that the visual dashboard empowers building managers to make better decisions, yet there is no evaluation of that claim. The benchmark results in Tables II-III evaluate forecasting models on historical data only, and the screenshots in Figures 3-4 are illustrative. No user study, deployment case, task-based usability test, or even a concrete worked example demonstrating how a manager would translate a forecast into an operational decision is provided. Without such evidence, the assertion that the tool 'enables' managers to take immediate actions is an unsupported claim rather than a demonstrated result.","section":"Tables II-III, Section III-D, Figures 3-4"},{"comment":"The title and framing use 'Time Series Foundation Model' and 'pre-trained models' loosely. Of the five models listed in Section III-B, only One-Fits-All uses a pretrained GPT-2 backbone; DLinear, PatchTST, Informer, and iTransformer are trained from scratch on two months of building data, as described in Section III-C. 'Pre-trained' in Section II-B appears to mean 'trained offline beforehand,' not 'pretrained on a large general corpus.' This ambiguity obscures what is actually being evaluated and should be clarified either by narrowing the claim or by adjusting the terminology.","section":"Title, Section III-B"}],"minor_comments":[{"comment":"The conclusion that DLinear is 'most robust' and One-Fits-All excells in relative accuracy is based on averaged metrics with overlapping standard deviations across horizons and channels; no statistical significance tests or per-horizon breakdowns are shown, so the ranking may not be reliable.","section":"Section III-D"},{"comment":"The sentence 'A more detailed result is illustrated in [10]' delegates the core experimental detail to a companion paper by the same authors. Either provide the detailed results in this paper or make it explicit that [10] is an unreviewed preprint and clarify how it can be accessed by reviewers.","section":"Section III-D text"},{"comment":"The note 'The best score is in red' will be lost in black-and-white printing; please use boldface or a symbol to mark the best values. The figures should also state in the captions that they are screenshots of the proposed interface rather than results from a live deployment.","section":"Tables II-III, Figures 3-4"},{"comment":"The paper mixes future tense ('will provide,' 'will be pre-trained') with present-tense claims ('The tool enables building managers'). Since the system appears to be a design proposal, make the tense consistent and clearly distinguish implemented components from planned ones.","section":"Throughout"},{"comment":"The term 'real-time' is used without a precise definition. Given that the data is resampled to 10-minute intervals and results are updated 'each regular period,' clarify the intended meaning of 'real-time' and any latency bounds relative to the BMS API.","section":"Introduction, Section II"},{"comment":"There are several minor typographical issues: 'BuildingTimeSeriesAnalytics' should be 'Building Time Series Analytics,' 'i.e,' should be 'i.e.,' and 'BTS-A/C [1] are not yet public' should be 'is not yet public.'","section":"Text quality"}],"recommendation":"major_revision","confidential_remarks":"The manuscript leans heavily on the authors' own unpublished companion work [10] for detailed results and on their own dataset paper [1]; the review would benefit from direct access to [10]. The absence of any user or deployment evaluation, together with the internal contradiction between the abstract's real-time promises and Section III-E's limitation, makes the current version a weak candidate for acceptance. However, the core ideas are not irreparable: a substantial revision that consistently frames BiTSA as an offline analytics tool (or adds an online-learning component and a concrete usability evaluation) could bring the claims into line with the evidence. I would ask the editor to ensure that companion works are made available to reviewers if a revision is invited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou should know that this paper is a system demo with a large gap between its headline promise and its own stated limitations. The abstract and conclusion say BiTSA allows building managers to take 'immediate, data-driven actions based on real-time insights,' but Section III-E admits that offline learning leads to lag and that the tool 'may lack the ability to provide real-time insights.' That internal contradiction is the main thing to take away.\n\nTo be fair, the paper does some things well. The dashboard design is clearly described with screenshots, and the architecture is sensible: frontend, backend, data ingestion, preprocessing, pretrained models, results database. The offline benchmark compares five standard models (DLinear, PatchTST, Informer, iTransformer, One-Fits-All) on two public datasets (BTS-B and BLDG59), reporting mean and standard deviation. The authors honestly state their limitations, which is more than many system papers do. The demo video is public.\n\nThe soft spot is exactly the real-time claim. The system is designed for periodic updates, not live streaming, and no user study or deployment is reported to test whether managers can actually convert forecasts into better decisions. The model tables are also condensed from the companion paper [10], so the only truly new artifact here is the visualization wrapper. That is a legitimate contribution for a demo track, but not for a full research paper as framed.\n\nThe citation pattern is fine; the authors lean on their own companion work, but that is expected in this kind of project report. No parameter fitting is used to manufacture the headline result.\n\nWho benefits from this? Someone building energy dashboards might find the architecture a useful reference. A reader looking for a new forecasting method will not. As a research claim, the paper overreaches; as a demo, it is acceptable.\n\nMy recommendation: I would not desk reject this, but I would send it to reviewers with the expectation that the authors either add a user study or deployment evaluation, or rewrite the abstract and conclusion to match the actual system capability. If they choose the second path, it becomes a reasonable demo paper.\n\nBest,\n[Your name]","headline":"A designed dashboard that overpromises real-time insights in its abstract, contradicted by its own limitations section, but the architecture and honest model benchmark give it value as a demo.","tokens_in":6140,"tokens_out":2264,"would_cite":false,"duration_ms":21520,"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":"The paper argues that wrapping pre-trained time-series forecasting models in an interactive visualization tool lets building managers act directly on energy forecasts, bridging research and operations.","keywords":["building energy analytics","time series forecasting","foundation model","interactive visualization","building management system","IoT sensor data","channel-independent models"],"falsifier":"Deploy BiTSA against a live BMS stream for a month without retraining and compare forecast MAE and SMAPE with the offline results from BTS-B and BLDG59; if live errors rise well above the reported range (SMAPE around 43–50%) or if a manager acting on the forecasts does not reduce energy use relative to a control period, the real-time-insight claim is falsified.","tokens_in":5080,"feed_emoji":"🏢","tokens_out":9792,"duration_ms":87450,"temperature":0.7,"pith_summary":"The paper's central claim is that the main obstacle to using advanced energy-forecasting models in real buildings is not accuracy but usability: IoT data is too complex for building managers to interpret. To close that gap, it presents the design of BiTSA, an interactive visualization platform that wraps pre-trained time-series models—DLinear, PatchTST, Informer, iTransformer, and a GPT-2-backed foundation model called One-Fits-All—behind a browser dashboard where managers can inspect forecasts for individual sensors. To support the design, the paper benchmarks these models on two real building datasets and reports that channel-independent models generally outperform channel-dependent ones, with DLinear best on absolute error and One-Fits-All best on percentage error. A sympathetic reader would take the contribution as a workflow design plus an initial model comparison; the paper itself concedes that offline learning makes the system lag changing building conditions. The payoff, if the design works, is that non-expert staff can make proactive energy decisions from live visualizations.","feed_headline":"BiTSA turns AI forecasts into building-management decisions","feed_subtitle":"If BiTSA works, building managers can act on 2-hour to 1-week energy forecasts without touching raw data.","key_machinery":"The carrying object is the BiTSA pipeline itself, whose named components are the front-end UI, authentication, data ingestion, preprocessing, statistics extractor, pre-trained models, and results database. The load-bearing comparison inside it is channel-independent versus channel-dependent forecasting: channel-independent models are trained separately per sensor series, while channel-dependent models learn cross-sensor dependencies. The 'time series foundation model' of the title is One-Fits-All, a pretrained GPT-2 backbone repurposed for forecasting; it carries the claim that a general pretrained model can be competitive on building data when made accessible through a simple dashboard. What this machinery does is replace manual data wrangling and model expertise with an automated pipeline and visual filtering, so the remaining user action is reading forecasts and responding.","core_discovery":"On the paper's own terms, BiTSA is an end-to-end analytics pipeline: cloud databases from a Building Management System feed IoT time series through authentication, data ingestion, preprocessing (10-minute resampling with polynomial imputation), statistical extraction, and pre-trained forecasting models; results are stored and rendered in an interactive dashboard. The experiments train each built-in model on one day of history (144 10-minute steps) and evaluate horizons from 12 to 1008 steps, i.e., two hours to one week ahead, on the BTS-B and BLDG59 datasets. The finding that channel-independent models (DLinear, PatchTST, One-Fits-All) beat channel-dependent models (Informer, iTransformer) is offered as evidence that per-sensor modeling fits heterogeneous building IoT data. The paper then maps this to decisions: DLinear for absolute accuracy such as cost budgeting, One-Fits-All for percentage accuracy such as efficiency benchmarking, all delivered through a visual interface that lets a manager select sensors from a dropdown.","pith_inferences":["Our inference: a randomized user study comparing decision quality with BiTSA versus a conventional static dashboard would separate the interface benefit from the forecast benefit; this test is not in the paper.","Our inference: the offline-learning lag points to an obvious extension the paper does not develop—an online fine-tuning loop or drift detector that retrains when forecast error rises—which would make the real-time-insight claim testable.","Our inference: because the reported SMAPE values are around 43–50%, a simulation study that feeds forecast errors into typical HVAC or demand-response decisions would show whether the accuracy is actually actionable; the paper does not establish an actionability threshold."],"forward_implications":["If the design works as described, a building manager can select any sensor from the Analytics page and see multi-horizon forecasts without data-science support, which is the paper's central usability claim.","The benchmark implies DLinear should be the default choice when absolute error matters, such as budgeting for energy cost, and One-Fits-All when percentage accuracy matters, such as efficiency benchmarking.","The consistent advantage of channel-independent models is a direct argument that per-sensor forecasting is sufficient for heterogeneous building IoT data, so more complex cross-channel transformers may be unnecessary.","Because models are pre-trained on two months of 10-minute data and stored in the system, the tool is designed for periodic batch updates rather than stream learning; the paper acknowledges this creates a lag.","The reported practical guidance is that DLinear suits absolute-unit predictions and One-Fits-All suits relative-error or percentage-based sustainability tracking."],"supporting_citations":[{"why":"Supplies the BTS-B building timeseries dataset used to simulate a BMS-controlled building in the evaluation.","marker":"[1]"},{"why":"Supplies the BLDG59 building dataset, the second evaluation environment for the built-in forecasting models.","marker":"[2]"},{"why":"DLinear is the channel-independent baseline that achieves the best absolute-error metrics in the comparison.","marker":"[5]"},{"why":"PatchTST is the channel-independent transformer baseline whose per-sensor modeling supports the main comparison.","marker":"[6]"},{"why":"Informer is the channel-dependent transformer baseline that the paper contrasts with channel-independent models.","marker":"[7]"},{"why":"iTransformer is the channel-dependent transformer baseline included in the benchmark.","marker":"[8]"},{"why":"One-Fits-All, the GPT-2-backed foundation model, produces the best relative-error results and motivates the title's foundation-model claim.","marker":"[9]"},{"why":"The companion study provides the detailed per-horizon results that back the aggregated tables in this paper.","marker":"[10]"}],"fun_headline_variants":["BiTSA: AI forecasts become building decisions at a glance","BiTSA bridges AI forecasting and building management","BiTSA gives building managers AI-powered energy insights","BiTSA turns raw sensor data into manager-ready forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that building managers will turn dashboard forecasts into correct operational actions, and that models trained on two months of offline 10-minute data will keep producing trustworthy forecasts on live building data; the paper does not test either link.","fun_headline_variants_meta":{"raw":{"variants":["BiTSA: AI forecasts become building decisions at a glance","BiTSA bridges AI forecasting and building management","BiTSA gives building managers AI-powered energy insights","BiTSA turns raw sensor data into manager-ready forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000967,"raw_usage":{"total_tokens":4075,"prompt_tokens":869,"completion_tokens":3206,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":3141}},"tokens_in":485,"tokens_out":3206,"duration_ms":21039,"temperature":1.0,"reasoning_tokens":3141,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:54:59.614615+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy BiTSA against a live BMS stream for a month without retraining and compare forecast MAE and SMAPE with the offline results from BTS-B and BLDG59; if live errors rise well above the reported range (SMAPE around 43–50%) or if a manager acting on the forecasts does not reduce energy use relative to a control period, the real-time-insight claim is falsified.","supporting_citations":[{"cited_title":"Prabowo, X","cited_arxiv_id":null,"evidence_quote":"Supplies the BTS-B building timeseries dataset used to simulate a BMS-controlled building in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BLDG59 building dataset, the second evaluation environment for the built-in forecasting models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DLinear is the channel-independent baseline that achieves the best absolute-error metrics in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Informer is the channel-dependent transformer baseline that the paper contrasts with channel-independent models."},{"cited_title":"Exploring Capabilities of Time Series Foundation Models in Building Analytics","cited_arxiv_id":"2411.08888","evidence_quote":"The companion study provides the detailed per-horizon results that back the aggregated tables in this paper."}],"review_version":1}