{"id":"3ca77a8f-e413-4deb-919c-cc8e2f91dff2","arxiv_id":"2411.15716","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Fed-TREND uses server-generated synthetic time series data, distilled from client and global model updates, to reduce the accuracy gap in heterogeneous federated forecasting.","lead":"This paper introduces Fed-TREND, a way to improve federated learning for time series forecasting by having the server create small synthetic datasets from the models that clients upload. The synthetic data helps devices learn from each other and improves forecast accuracy on eight public datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Untested causal control: the reported gains may come from extra synthetic data, not from trajectory matching.","rationale":"The reader's weakest-assumption analysis points toward MTT transfer, and I partially agree, but the deeper issue is causal attribution. The paper's ablations show that Dct and Dgt each help relative to having no synthetic data; they do not show that optimized synthetic data helps relative to any synthetic data. Because Eq. (7) uses only endpoints and does not report the inner-loop step count Lc, the optimization is not even specified precisely enough to re-implement unambiguously. This strengthens the reader's CONDITIONAL verdict: the empirical claims are plausible and internally consistent, but the central mechanism needs a random-synthetic control and multi-seed reporting before the paper can be accepted as demonstrating that trajectory matching is what drives the improvements. There is no suggestion of bad faith; this is a purely evidential gap in the presented argument.","tokens_in":59,"tokens_out":9899,"duration_ms":164799,"concrete_test":"On Electricity and State-ILI, replace Dct and Dgt with fixed random time series of identical shapes, generated once with a fixed seed and scaled to each dataset's training mean and standard deviation, and rerun the Table III pipeline with all other hyperparameters unchanged, including the 10-round update schedule and the 300-iteration construction loop (or with construction replaced by regenerating random series). Repeat both Fed-TREND and the control with at least 10 seeds and report mean ± std. If the random-synthetic control attains MSE within 50% of Fed-TREND's improvement over FedAvg on either dataset, the trajectory-matching component is not the driver of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central explanation is that Dct and Dgt improve federated forecasting because MTT trajectory matching captures representative client and global distributions. That causal step is the least secure part of the argument. Equation (7) optimizes Dct from only the start and end models of each Lct-long interval, discarding all intermediate checkpoints; the inner-loop length Lc is not reported and no regularization is described. Endpoint-only matching is therefore underdetermined, and the usual MTT guarantee of matching the whole training trajectory does not transfer automatically to time series. The experiments cannot detect this failure: Table III and the ablations compare Fed-TREND only with no-synthetic baselines, never with a control that uses random or data-independent synthetic series of the same size and schedule. Since Dct is mixed into local training and Dgt adds server-side fine-tuning, Fed-TREND has extra training signal relative to FedAvg. Some margins are small (USWeather MSE 0.45444 vs 0.44036; Solar 0.32251 vs 0.31457), and all numbers are single-run. If a random-synthetic control closes most of the gap, the observed gains come from extra synthetic data or regularization, not from the trajectory-matching mechanism that motivates the method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Fed-TREND, a server-side synthetic-data augmentation framework for cross-device federated time series forecasting. In each federated round, the server stores client model updates and aggregated global models; every Lct or Lgt rounds it constructs two small synthetic datasets using the MTT trajectory-matching objective: Dct is built from pairs of start/end client model checkpoints (with a consistency mask) and distributed to clients for mixing into local training, while Dgt is built from global model trajectory segments and used to finetune the aggregated global model. The method is designed to be compatible with any forecasting backbone and any standard FL aggregation protocol. Experiments compare Fed-TREND against FedAvg, FedProx, FedDyn, Elastic, FedHEAL, and DynaFed on eight time series datasets with four forecasting models (DLinear, LightTS, TSMixer, iTransformer), and include ablations, hyperparameter sensitivity studies, integration with other FL frameworks, and a differential-privacy experiment.","tokens_in":20844,"tokens_out":4346,"duration_ms":42980,"significance":"If the central mechanism is confirmed, Fed-TREND is a practically attractive contribution: it addresses data heterogeneity without altering the federated protocol, offloads the synthetic-data construction burden to the server, and the reported results span a wide range of datasets, backbones, and FL frameworks. The ablation study and the compatibility experiments are useful evidence that each component contributes to the gains. However, the paper's central explanatory claim—that trajectory matching, rather than the mere presence of additional synthetic training data, drives the improvements—is not isolated experimentally, and the empirical evidence is weakened by single-run reporting with no error bars or significance tests. The strengths are the breadth of evaluation and the clean integration with existing FL algorithms, but the current evidence is not yet sufficient to fully support the strong conclusions made in Section V-E.","major_comments":[{"comment":"The central causal claim that Dct and Dgt improve federated forecasting because MTT trajectory matching captures representative client and global distributions is not isolated experimentally. Every comparison in Table III and the ablations in Table V is against a baseline with no synthetic data at all; there is no control that injects same-size, same-schedule random or data-independent synthetic series into local training and server-side finetuning. Since Dct adds training signal to each client and Dgt adds server-side finetuning, Fed-TREND has strictly more training data than FedAvg. If a random-synthetic control closes most of the gap, the reported gains would be attributable to extra data or regularization rather than to the trajectory-matching mechanism that motivates the method. Please add such controls—for example, Gaussian-noise series or randomly shuffled real windows—with the same sizes, update intervals, and construction iterations, at least for the main DLinear results and the Table V ablations.","section":"§IV-B, Eq. (7); §V-E, Table V"},{"comment":"All results are single-run and no error bars, multiple seeds, or significance tests are reported. Several margins are small (e.g., Table IV Solar Energy with TSMixer: 0.27527 vs. 0.27441 MSE; Table III USWeather: 0.45444 vs. 0.44036 MSE), which is inconsistent with the claim in Section V-E of improvement 'with a large margin.' The paper should report mean and standard deviation over at least three seeds, and preferably a paired significance test, for the headline comparisons and the ablation study, so that the consistency and magnitude of the gains can be assessed.","section":"§V-E, Table III; §V-F, Table IV"},{"comment":"The synthetic-data construction procedure is underspecified in ways that affect reproducibility and the validity of Eq. (7). The inner-loop length Lc used to train Wstart on Dsyn in Eq. (6) and in Algorithm 1 Line 34 is never defined or reported; the distance metric d(·,·) is not specified; and the consistency-mask mechanism is described only as 'will be masked' in the text, with no exact formula for how the mask enters Eq. (7). Since the paper argues that the consistency direction is important, the exact masking rule must be stated. These details are needed for readers to reproduce the method and to check whether the endpoint-only optimization in Eq. (7) is faithful to the MTT procedure.","section":"§IV-B, Eq. (6)–(7); Algorithm 1, Lines 29–36"}],"minor_comments":[{"comment":"The text says the authors implement 'LDP with the Laplace mechanism' but then states that noise is sampled from N(0, λ²I), which is Gaussian noise. Please clarify which distribution was actually used and, if Laplace noise was intended, correct the formula.","section":"§V-I"},{"comment":"The column header 'Iliness' should be 'Illness'.","section":"Table V"},{"comment":"The pseudocode checks t%Lct==0 at both Line 11 and Line 19, and the ordering of appending client updates, aggregating, and reconstructing Dct within the same round is ambiguous. Please clarify the exact round indices and ordering.","section":"Algorithm 1"},{"comment":"Figure 6 reports only the ETTh1 dataset, while the text says 'similar trends are observed across the other seven datasets'; either include those results or state the limitation explicitly.","section":"§V-H4"},{"comment":"The claim that 'in all cases, Fed-TREND consistently outperforms FedAvg by a significant margin' is only supported by plots; numerical values should be provided, as the difference appears small on some datasets.","section":"§V-H3"},{"comment":"There are several typos and infelicities, e.g., 'Notebly' in Section III-A and 'severing various applications' in Section I; a careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the idea is relevant, but the main empirical claim currently rests on comparisons against no-synthetic baselines only. The missing random-synthetic control is the key load-bearing issue; if the authors add such a control and report multi-seed statistics, the contribution would be substantially strengthened. I do not see a novelty-disclosure problem with the related work cited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: Fed-TREND is a clean, well-motivated addition to federated time series forecasting. The idea—use MTT trajectory matching to synthesize two auxiliary datasets, one from client model updates for local training and one from global trajectories for server-side fine-tuning—is genuinely new in this setting. The experimental footprint is broad: eight datasets, four forecasting backbones, five FL baselines, and an LDP compatibility check. Every comparison shows improvement, and the ablation supports each component's contribution. That is real work, and it deserves a careful referee.\n\nSoft spots, in order of importance. First, the causal story is under-tested. The paper attributes the gains to Dct and Dgt capturing representative client/global distributions via trajectory matching. But the experiments never compare against a control where the synthetic data are random or data-independent series of the same size and update schedule. Since Dct is mixed into local training and Dgt adds server-side fine-tuning, Fed-TREND has strictly more training signal than FedAvg. On several datasets the margins are thin (Solar with TSMixer: 0.27527 to 0.27441 MSE; USWeather: 0.45444 to 0.44036). Without a random-synthetic control, the central mechanism is not isolated from 'extra data helps'—a mundane but publication-worthy alternative. This is the main reason I would not accept the empirical claims as-is.\n\nSecond, the trajectory matching itself is done on the cheap: Eq. 7 only matches start and end checkpoints per interval, discarding intermediate dynamics that MTT normally uses; the inner-loop length Lc is never reported. For time series, where distributions drift over time, endpoint-only matching is even less clearly justified. This is a missing detail, not necessarily a flaw, but it reinforces the need for the control.\n\nThird, single-run results, no error bars, no significance tests. Some improvements are large and consistent, so I am not accusing the authors of cherry-picking, but the evidence would be much stronger with repeated seeds.\n\nMinor: the privacy analysis in Sec. IV-E1 is about one paragraph. The authors say privacy is 'consistent with traditional FL' because synthetic data is generated on the server. That is plausible, but synthetic data derived from model updates is then sent back to clients, and the inversion risk is not analyzed. A short discussion would help.\n\nBottom line: the paper is serious, the idea is useful, and the experiments are broad. The missing control and missing seeds are fixable. I'd send it to review and ask for those before acceptance.","headline":"Useful, well-motivated method for federated time series forecasting with broad empirical gains; main caveat is that no control disentangles trajectory matching from simply adding extra synthetic data.","tokens_in":21320,"tokens_out":2216,"would_cite":true,"duration_ms":20924,"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":"Fed-TREND generates two synthetic datasets from client and global model trajectories, and this server-side augmentation consistently outperforms federated learning baselines, matching or beating centralized training on six of eight time…","keywords":["federated learning","time series forecasting","data heterogeneity","synthetic data","dataset distillation","trajectory matching","model aggregation","cross-device learning"],"falsifier":"Replace the trajectory-matching objective for Dct and Dgt with randomly initialized or fixed synthetic sequences of the same size and schedule, keeping everything else identical; if Fed-TREND's advantage over FedAvg on the eight reported datasets persists, the trajectory-matching mechanism is not what carries the result.","tokens_in":20349,"feed_emoji":"📈","tokens_out":5639,"duration_ms":47239,"temperature":0.7,"pith_summary":"Federated time series forecasting fails when client data are heterogeneous: each device may track different variables, or the same variable under different temporal patterns, so simple model averaging produces a weak global model. This paper tries to close that gap with Fed-TREND, a server-side augmentation that builds two small synthetic datasets from model trajectories and feeds them back into training. The first synthetic set captures the representative distribution of all clients' model updates and is mixed into each client's local data; the second captures the long-term dynamics of the aggregated global model and is used to fine-tune the global model after aggregation. If the reported results hold, a modest amount of synthetic data can reduce most of the accuracy loss caused by heterogeneity, and on six of eight datasets the federated model even outperforms centralized training.","feed_headline":"Synthetic data lets federated forecasting beat centralized training","feed_subtitle":"Two distilled datasets carry client and global knowledge across devices, closing most of the heterogeneity gap.","key_machinery":"The load-bearing mechanism is trajectory matching, a dataset-distillation technique in which a small synthetic dataset is learned so that a model trained on it from a stored checkpoint reproduces a later recorded checkpoint after the same number of steps. Fed-TREND applies this to two trajectory banks: Tct stores the start and end client model updates within each interval, and Tgt stores the sequence of aggregated global models. The synthetic datasets Dct and Dgt are the trainable input-output pairs that minimize the distance between reproduced and recorded checkpoints, with the Dct loss masked to parameters whose updates move consistently across rounds, and with the whole construction carried out on the server at fixed intervals.","core_discovery":"The paper's claim is that data heterogeneity in cross-device federated time series forecasting can be addressed without changing the federated protocol, by having the central server distill two types of synthetic data from the models it already receives. Synthetic dataset Dct is optimized from the start and end checkpoints of each client's local update within a window, using only parameters whose update direction is consistent, and is then sent back to clients to augment their local training. Synthetic dataset Dgt is optimized from the trajectory of aggregated global models and is used to fine-tune the global model after each aggregation. Across eight datasets and four forecasting backbones, the resulting framework, Fed-TREND, consistently outperforms FedAvg, FedProx, FedDyn, Elastic, FedHEAL, and DynaFed, and on Electricity, Traffic, Solar Energy, ETTh1, ETTh2, and USWeather it reports lower error than centralized training.","pith_inferences":["The consistent-update masking implies the method bets on consensus directions: if heterogeneity is driven by conflicting but informative client-specific signals, masking out inconsistent updates could discard precisely the information that matters for generalization.","The same two-level trajectory distillation could be applied to other federated tasks beyond time series, since neither the formulation nor the trajectory banks use temporal structure; a natural test is whether the gains replicate for tabular or graph-structured federated data.","The finding that federated training with synthetic augmentation beats centralized training on six datasets suggests the synthetic data act partly as a regularizer; a testable extension is measuring whether the advantage shrinks as local epochs or client participation rates increase.","Because Dgt is built from global model dynamics, its quality depends on the aggregation rule being stable; switching to a different aggregator may require re-tuning the update interval Lgt to keep the long-term trajectory meaningful."],"forward_implications":["If Fed-TREND is correct, server-side synthetic augmentation is a general drop-in remedy for heterogeneity: plugging it into FedAvg, FedProx, FedDyn, Elastic, FedHEAL, or DynaFed improves each baseline on every dataset tested.","The method transfers across forecasting architectures, improving DLinear, LightTS, TSMixer, and iTransformer relative to plain federated training, which suggests the synthetic data carry model-agnostic temporal knowledge.","Because the synthetic datasets are small (20 input-output pairs refreshed every 10 rounds), the additional communication cost is less than 30KB per client over 80 rounds, making the mechanism practical for bandwidth-limited devices.","Fed-TREND remains effective when local differential privacy noise is added to model updates, so the privacy properties of the base federated protocol are not sacrificed.","On the two most heterogeneous datasets, State-ILI and Country-Temp, Fed-TREND does not reach centralized performance but still records the best results among all federated baselines."],"supporting_citations":[{"why":"Supplies the trajectory-matching distillation objective that Fed-TREND uses to construct both synthetic datasets.","marker":"[12]"},{"why":"Defines the FedAvg baseline and aggregation rule that Fed-TREND extends and compares against.","marker":"[9]"},{"why":"Provides the idea of focusing on consistently updated parameters, which Fed-TREND adopts when masking its Dct construction loss.","marker":"[48]"},{"why":"The closest prior work, using global model trajectories to generate synthetic data, which Fed-TREND distinguishes by adding client-side Dct and by updating its synthetic data over time.","marker":"[54]"},{"why":"Supplies DLinear, the default forecasting backbone used in the main experiments and ablations.","marker":"[6]"}],"fun_headline_variants":["Fed-TREND: synthetic data beats central training in federated forecasting","Two synthetic datasets fix heterogeneity in federated time series","Server-distilled synthetic data lifts federated forecasting accuracy","Synthetic data closes heterogeneity gap in federated forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire gain rests on the assumption that a small synthetic dataset distilled from only the start and end checkpoints of each client's local update, via trajectory matching, captures enough of every client's data distribution to improve other clients' local training once mixed with real data.","fun_headline_variants_meta":{"raw":{"variants":["Fed-TREND: synthetic data beats central training in federated forecasting","Two synthetic datasets fix heterogeneity in federated time series","Server-distilled synthetic data lifts federated forecasting accuracy","Synthetic data closes heterogeneity gap in federated forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1561,"prompt_tokens":982,"completion_tokens":579,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":512}},"tokens_in":598,"tokens_out":579,"duration_ms":5113,"temperature":1.0,"reasoning_tokens":512,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:58:15.610205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the trajectory-matching objective for Dct and Dgt with randomly initialized or fixed synthetic sequences of the same size and schedule, keeping everything else identical; if Fed-TREND's advantage over FedAvg on the eight reported datasets persists, the trajectory-matching mechanism is not what carries the result.","supporting_citations":[{"cited_title":"Fair federated learning under domain skew with local consistency and domain diversity,","cited_arxiv_id":null,"evidence_quote":"Provides the idea of focusing on consistently updated parameters, which Fed-TREND adopts when masking its Dct construction loss."},{"cited_title":"Dynafed: Tackling client data heterogeneity with global dynamics,","cited_arxiv_id":null,"evidence_quote":"The closest prior work, using global model trajectories to generate synthetic data, which Fed-TREND distinguishes by adding client-side Dct and by updating its synthetic data over time."}],"review_version":1}