{"id":"9af6a38f-a6f7-43cc-a221-a4da1c322bb1","arxiv_id":"2412.08906","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FFTS is a federated pretraining framework with a timescale-aware mixture-of-experts module that trains a time series foundation model from scratch across heterogeneous, non-shared datasets.","lead":"This paper trains time series foundation models with federated learning, keeping each dataset on its own client and sharing only model updates. The resulting model is claimed to match or beat centralized training on forecasting, imputation, and anomaly detection while preserving data privacy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The no-leakage guarantee for downstream evaluation is unverifiable because the 18-dataset pretraining list is in the omitted appendix; if any downstream set was pretrained on, all generalization claims collapse.","rationale":"The paper's contribution is an FL pretraining method that yields a TSFM with 'superior generalization capabilities' across forecasting, imputation, and anomaly detection. The strongest claim, as stated in the abstract, is about generalization. The precondition for any generalization claim is that evaluation datasets were not seen during pretraining. The authors assert this, but the appendix containing the pretraining dataset list is absent from the arXiv v1. The full text names only Traffic and Electricity as concrete pretraining datasets, and those are explicitly excluded from downstream evaluation because they were in pretraining. The downstream sets—ETT, Weather, ILI, M4, SMD, MSL, SMAP, SWaT, and PSM—are standard public benchmark corpora, so overlap with an 18-dataset collection is a real possibility rather than a paranoid speculation. Because code is public, this is easily checkable. This concern is more load-bearing than the ATM ablation question raised by the reader: if ATM contributed nothing, the method would still be FedAvg with an extra regularization term, and the FL-side comparisons (FedAvg competitive, FFTS slightly better) would still stand as a partial result. If leakage occurred, even the FedAvg numbers would be invalid for the generalization claim. I therefore do not change the conditional verdict but would make the condition explicit: the no-leakage check must be passed before accepting the abstract's generalization claim. The reader's rationale already mentioned the unverifiable no-leakage claim, although their designated weakest assumption was the ATM; I mark partial agreement.","tokens_in":15875,"tokens_out":4364,"duration_ms":44959,"concrete_test":"Download the released code from https://github.com/shengchaochen82/FFTS (or the camera-ready appendix) and enumerate the 18 client datasets from the pretraining configuration and dataloader. Check exact overlap with the downstream evaluation sets: ETTh1, ETTh2, ETTm1, ETTm2, Weather, ILI, M4, SMD, MSL, SMAP, SWaT, and PSM. If any downstream set appears as a pretraining client, the no-leakage claim is false and Tables 4–9 lose their evidentiary force; if no overlap exists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FFTS yields superior generalization from federated pretraining depends critically on the assertion (Experiments, Federated Pretraining Setups, and in the downstream setups) that 'All downstream datasets were excluded from the pre-training phase to prevent data leakage.' The full list of 18 pretraining datasets is deferred to an omitted Appendix B, so this exclusion cannot be checked from the manuscript. The downstream benchmarks are exactly ETT (four variants), Weather, ILI, M4, SMD, MSL, SMAP, SWaT, and PSM—common public sets that could plausibly appear in a collected pretraining corpus. If any of them is among the 18 pretraining clients, the zero-shot and few-shot numbers in Tables 4–9 are not generalization but memorization, and the 'federated pretraining replaces centralized pretraining' conclusion is unsupported. This is load-bearing because it is the only gate between the pretraining corpus and the evaluation of the strongest claim. The manuscript itself flags the omission ('see Appendix for details', 'More about pretraining can be found at Appendix B'), so this is an acknowledged missing support rather than a manufactured issue. Until the list is disclosed or the released code/config is inspected, the evidence for the headline claim is conditional.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FFTS, a federated learning framework for pretraining time series foundation models from heterogeneous datasets. Each dataset is treated as an independent client that trains a local encoder-only Transformer with patch embedding, an Adaptive Trend-awareness Module with four fixed timescale experts, and a masking-based reconstruction objective. A heterogeneous knowledge alignment regularization term is added to both local and global objectives, and a unified adaptation head is used for downstream forecasting, imputation, and anomaly detection. The authors evaluate FFTS on standard benchmarks, comparing it with FedAvg, FedProx, pFedMe, a centralized variant FFTS-Cen, and several task-specific and LLM-based baselines.","tokens_in":16194,"tokens_out":5588,"duration_ms":58534,"significance":"If the claims hold, the paper makes a useful contribution: it provides an alternative to centralized pretraining for time series foundation models, preserves data locality, and includes a code release. The evaluation covers three major tasks with many baselines, and the ablation study attempts to isolate the effect of the proposed module and regularizer. The main limitations are evidentiary: the pretraining dataset list is deferred to an omitted appendix, making the no-leakage guarantee unverifiable; the headline improvement number is inconsistent with the reported table; and the differences from the centralized baseline are small and reported without error bars. These issues are load-bearing for the central claim that federated pretraining can replace centralized pretraining for TSFMs, but they are fixable with additional disclosure and analysis.","major_comments":[{"comment":"The text states that FFTS surpasses Time-LLM by an averaged 9.52%, but the table averages are 0.497 for Time-LLM and 0.473 for FFTS, which is a relative improvement of about 4.8% and an absolute difference of 0.024. Please correct this number and state explicitly whether reported percentages are relative or absolute improvements.","section":"Time Series Forecasting, Main Results (Table 4)"},{"comment":"The guarantee that all downstream datasets were excluded from pretraining cannot be checked: the list of 18 pretraining datasets is deferred to an omitted Appendix B. This is not a cosmetic issue because Table 4 excludes Traffic and Electricity specifically due to their presence in pretraining, so overlap with other common public downstream sets (ETT variants, Weather, ILI, M4, SMD, MSL, SMAP, SWaT, PSM) is a live risk. Please provide the full pretraining corpus and an explicit overlap check with all downstream benchmarks.","section":"Experiments, Federated Pretraining Setups and Downstream Setups"},{"comment":"The reported differences between FFTS and FFTS-Cen are small (Table 4: 0.473 vs 0.475; Table 5: 0.366 vs 0.369; Table 8: 0.034 vs 0.035; Table 9: 88.10 vs 87.69), and no error bars, number of seeds, or significance tests are reported. The claims that FFTS outperforms centralized pretraining in most cases and is superior to FL baselines need uncertainty quantification before they can be accepted.","section":"Tables 4-9 (FFTS versus FFTS-Cen)"},{"comment":"The global objective in Eq. (5) is written as a weighted sum of local objectives plus a regularization term involving local and global ATM parameters, but the server does not train model parameters in the described protocol. It is unclear how the server-side ATM regularization is computed and optimized. Please specify the exact server update rule or revise the formulation to match the actual algorithm.","section":"Heterogeneous Knowledge Alignment, Eq. (5)"}],"minor_comments":[{"comment":"There are several typographical errors: 'FedAvd' should be 'FedAvg', 'shwon' should be 'shown', 'resuls' should be 'results', 'Knowledege' should be 'Knowledge', and 'statics' should be 'statistics'.","section":"Throughout"},{"comment":"The client index is introduced as k in the text but the sums use i; please align the notation consistently.","section":"Preliminary, Eq. (1) and Eq. (5)"},{"comment":"The claim that regularization 'mitigates frequent global updates, thus lowering bandwidth requirements' is not supported by any experiment or quantitative analysis; the communication cost per round is the same as in standard FedAvg. Please remove or substantiate this claim.","section":"Discussion, Computational/Communication Efficiency"},{"comment":"The model is described as a 'foundation model', but no parameter count or model scale is reported. Since scale is often part of what distinguishes foundation models, please report model size and training cost.","section":"Abstract and Title"},{"comment":"Several percentage improvements (4.1%, 3.87%, 14.7%, 1.02%) are reported without specifying the baseline average or the formula used; please state whether these are relative or absolute and provide the corresponding standard deviations.","section":"Tables 5, 6, 8, 9 and accompanying text"},{"comment":"The privacy discussion claims that regularization reduces overfitting risk and that ATM reduces the data footprint in updates, but these are not formal privacy guarantees and are not quantified. Consider softening the wording or adding measurements such as gradient inversion attack evaluations.","section":"Discussion, Privacy Guarantee"}],"recommendation":"major_revision","confidential_remarks":"The header note that this paper has been accepted by AAAI and that the complete version will be updated after camera-ready, combined with the omitted Appendix B, makes the current arXiv manuscript difficult to evaluate fully. I would ask the editor to ensure the final version includes the full pretraining dataset list, the complete result tables, and the code configuration. Also, the 'first exploration' claim should be checked against Time-FFM (Liu et al. 2024a), which is cited in the related work but not positioned as prior work on federated foundation models for time series forecasting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: this is the first paper I know of that trains a time series foundation model from scratch in a federated way, and the idea is sensible. The components—patch embedding, encoder-only Transformer, Mixture-of-Experts over timescales, masking, proximal alignment—are all known, but the combination, and specifically the choice to treat each dataset as a client rather than concatenating them, is a genuine new configuration. If the empirical claims hold, it matters for privacy-sensitive domains.\n\nThe paper does a fair amount of work well: experiments across forecasting, imputation, and anomaly detection, with ablations and hyperparameter sensitivity. The framework ablation (FFTS-A and FFTS-B) gives some evidence that both the ATM and the alignment term contribute. The writing is clear enough to reproduce the method.\n\nNow the soft spots, in proportion.\n\nFirst, the headline number is wrong. The text says FFTS surpasses Time-LLM by 9.52% on long-term forecasting; Table 4 gives Time-LLM 0.497 and FFTS 0.473, a relative improvement of about 4.8%. Discrepancies like that make a reader wonder what else in the tables was computed the same way.\n\nSecond, no error bars or significance tests anywhere. The differences between FFTS and the centralized FFTS-Cen are small—roughly 0.4% to 2.9% relative across tasks—so the word \"superior\" is stronger than the evidence. The pattern is consistent, so I don't think it's noise, but I can't rule it out.\n\nThird, and most important: the no-leakage claim cannot be checked. The paper states that all downstream datasets were excluded from pretraining, but the list of 18 pretraining datasets is deferred to an omitted Appendix B. The downstream sets are common public benchmarks, so this is not a hypothetical concern. If any of them appear in pretraining, the zero-shot and few-shot results are not generalization. The stress-test note is right: this is load-bearing. The code is public, so a referee can inspect configs, but the manuscript itself is missing the evidence.\n\nFourth, the comparison set is thin for a foundation model paper. Time-LLM and GPT4TS are LLM adapters; there are no numbers against Moirai, TimesFM, or Chronos. That weakens the \"state of the art\" claim, though it doesn't undermine the core federated-pretraining idea.\n\nBottom line: this is a solid new combination, honestly presented in most places, with one concrete arithmetic inconsistency and one unverifiable leakage assertion. It deserves a serious referee, but the revision should fix the 9.52% claim, add error bars, and disclose the pretraining corpus.\n\nFor a reading group: yes. I'd bring it to prompt a discussion about what counts as evidence in TSFM papers.","headline":"First federated from-scratch TSFM training, with a sensible method but a checkable leakage gap and an inflated headline number.","tokens_in":16691,"tokens_out":2583,"would_cite":true,"duration_ms":26244,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Training time series foundation models under federated learning—each dataset kept at its own client—generalizes better than centralized pretraining on the same data.","keywords":["federated learning","time series foundation model","cross-domain heterogeneity","adaptive trend-awareness module","mixture of experts","masked reconstruction","zero-shot forecasting","anomaly detection"],"falsifier":"Run the same federated pretraining with the ATM disabled but the alignment regularization kept intact: if the resulting model matches FFTS on forecasting, imputation, and anomaly detection, the timescale experts are not carrying the reported gain. Conversely, removing the regularization while keeping the ATM would show whether the alignment term is the essential component.","tokens_in":15703,"feed_emoji":"📈","tokens_out":4506,"duration_ms":47002,"temperature":0.7,"pith_summary":"The paper argues that fusing heterogeneous time series into a single centralized corpus hurts the training of time series foundation models, because the same numerical pattern can mean different things in different domains. It proposes FFTS, a federated learning scheme in which each dataset stays on its own client, trains a local encoder-only Transformer, and only shares model parameters with a server. The paper claims that the resulting global model transfers better to unseen forecasting, imputation, and anomaly detection tasks than centralized training on the same data, and even beats task-specific state-of-the-art baselines. This matters because it suggests organizations can collaboratively pretrain strong time series models without pooling raw data, which is both a privacy advantage and a practical path to larger training corpora.","feed_headline":"Federated training beats centralized data pooling for time-series models","feed_subtitle":"Keeping datasets in silos yields stronger forecasting, imputation, and anomaly detection than fusing them into one corpus.","key_machinery":"The load-bearing component is the Adaptive Trend-awareness Module (ATM), a mixture-of-experts-style block appended to each Transformer layer with four fixed timescale experts for second, minute, hour, and day patterns. A gating network decomposes each representation into trend and seasonal components, computes softmax timescale weights, activates the top-k experts, and fuses their outputs through a feed-forward layer. The second mechanism is heterogeneous knowledge alignment: an ATM-specific regularization term $\\lambda\\|\\Theta_T - \\hat{\\Theta}_T\\|^2$ is applied both locally and on the server to pull each client's ATM parameters toward the global ATM parameters. A unified masking strategy with geometrically distributed mask segments forces local models to reconstruct masked time points rather than memorize domain-specific patterns. Together these pieces let each client preserve its unique data characteristics while the regularization aligns shared cross-timescale trends in the global model.","core_discovery":"On its own terms, the paper's central claim is that a time series foundation model pretrained with federated learning, where each domain is an independent client, achieves superior zero- and few-shot generalization compared with centralized pretraining that fuses the same datasets. FFTS reports the best average long-term forecasting error across ETT, Weather, and ILI benchmarks, surpassing the LLM-based Time-LLM by 9.52% on average in the fine-tuned setting, by 4.1% in few-shot settings, and by 3.87% in zero-shot transfer. For imputation, the FFTS-pretrained model reduces MSE by 14.7% relative to GPT4TS, and for anomaly detection it reaches an average F1 of 88.10%, beating GPT4TS by 1.02 percentage points and the centralized FFTS-Cen baseline on most tasks. The paper interprets these results as evidence that federated pretraining can replace centralized pretraining for time series foundation models while preserving data privacy.","pith_inferences":["The fixed set of four timescale experts may not be the actual source of gain: a variant that keeps the alignment regularization but replaces the ATM with a simpler linear projection would isolate whether trend decomposition matters, and the paper's current evidence for the ATM is visual only.","The approach should transfer to other settings where token meaning is context-dependent, such as multisensor IoT streams, where local training with an alignment penalty may beat cross-source fusion whenever identical numeric patterns mean different things in different contexts.","A data-driven alternative would learn timescale partitions per client or per layer; if that matches or beats the fixed second/minute/hour/day split, then the specific choice of four timescales is a convenience rather than a discovered property of time series.","The strong FedAvg baseline results suggest that the largest benefit may come from not fusing heterogeneous corpora at all; testing on more datasets and more clients could clarify how much the ATM adds beyond that effect."],"forward_implications":["Federated pretraining can replace centralized pretraining for time series foundation models when data holders cannot share raw data, without giving up accuracy.","A single FFTS-pretrained model transfers across forecasting, imputation, and anomaly detection using one lightweight adaptation head, so the same pretrained weights serve multiple downstream tasks.","The federated model beats the same-architecture centralized baseline on most benchmarks, implying that statistical heterogeneity across domains is not only an obstacle but can act as a useful regularizer.","Standard federated algorithms designed for image-style heterogeneity, such as FedProx and pFedMe, underperform on heterogeneous time series, so time series heterogeneity needs domain-specific alignment mechanisms.","Even vanilla FedAvg produces a competitive time series foundation model, supporting federated learning as a general pretraining strategy for this modality."],"supporting_citations":[{"why":"Supplies the FedAvg aggregation baseline and the core federated learning setting used throughout the paper.","marker":"[McMahan et al. 2017]"},{"why":"Provides the Mixture-of-Experts structure that the Adaptive Trend-awareness Module is inspired by.","marker":"[Fedus, Zoph, and Shazeer 2022]"},{"why":"Supplies the patch embedding strategy used to turn time series segments into tokens.","marker":"[Nie et al. 2022]"},{"why":"Supplies the masking strategy with geometrically distributed mask segments for reconstruction-based pretraining.","marker":"[Zerveas et al. 2021]"},{"why":"Provides the GPT4TS baseline and the experimental protocols for imputation and anomaly detection.","marker":"[Zhou et al. 2023]"},{"why":"Provides the Time-LLM baseline and the benchmark setup for long-term and few-shot forecasting.","marker":"[Jin et al. 2023]"},{"why":"Supplies the FedProx baseline and the proximal-term idea that the heterogeneous knowledge alignment regularization extends.","marker":"[Li et al. 2020]"},{"why":"Motivates the broader direction of combining federated learning with foundation model training.","marker":"[Zhuang, Chen, and Lyu 2023]"}],"fun_headline_variants":["Federated beats centralized for time-series foundation models","Privacy-preserving training wins for time-series AI","Siloed data outperform pooled corpora in time-series tasks","Federated pretraining tops centralized for forecasting and more","Federated learning improves forecasting, imputation, and anomaly detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Adaptive Trend-awareness Module's fixed set of four timescale experts and its gating network extract cross-domain trend patterns that improve the global model; if that fails, FFTS collapses to FedAvg with a proximal penalty.","fun_headline_variants_meta":{"raw":{"variants":["Federated beats centralized for time-series foundation models","Privacy-preserving training wins for time-series AI","Siloed data outperform pooled corpora in time-series tasks","Federated pretraining tops centralized for forecasting and more","Federated learning improves forecasting, imputation, and anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000943,"raw_usage":{"total_tokens":4027,"prompt_tokens":944,"completion_tokens":3083,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":3002}},"tokens_in":560,"tokens_out":3083,"duration_ms":23463,"temperature":1.0,"reasoning_tokens":3002,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:26:44.493063+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same federated pretraining with the ATM disabled but the alignment regularization kept intact: if the resulting model matches FFTS on forecasting, imputation, and anomaly detection, the timescale experts are not carrying the reported gain. Conversely, removing the regularization while keeping the ATM would show whether the alignment term is the essential component.","supporting_citations":[],"review_version":1}