{"id":"9134e5a4-41e1-4c45-ad60-f995c6c27ca5","arxiv_id":"2411.15743","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Freq-Synth trains zero-shot forecasters on harmonic sine waves generated from the target sampling rate and beats real-data training on 6 of 8 benchmarks.","lead":"This paper finds that time series forecasting models learn poorly when training data contains many frequencies and fail to generalize to unseen frequencies. It proposes Freq-Synth, a lightweight synthetic data generator that builds sine waves from the target sampling rate's fundamental frequency and harmonics, and reports better zero-shot and few-shot forecasts than training on real data in most benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Freq-Synth's claimed 6/8 advantage may depend on test-informed selection of the fundamental-frequency anchor and harmonic count, undermining the 'only sampling rate' zero-shot claim.","rationale":"The reader's weakest assumption concerned spectral dominance (Sec. 4.2), which the authors acknowledge fails for Exchange (App. B.4). That is a real limitation but is already hedged in the 6/8 claim. My concern targets a more load-bearing ambiguity: the sampling-rate-to-anchor mapping is not a unique, pre-specified rule, and the harmonic count appears tuned on the evaluation benchmark. If these choices are test-informed, the central empirical claim is unverified. The paper contains a plausible mechanism and useful analysis, but the current manuscript lacks code, error bars, and a fixed protocol. Therefore, I recommend UNVERDICTED rather than leaving the reader's CONDITIONAL: the claim may hold, but as written it is not verifiable.","tokens_in":829,"tokens_out":666,"duration_ms":89431,"concrete_test":"Fix an a priori anchor rule before seeing test data, e.g., daily fundamental for all sub-daily sampling rates and weekly for daily sampling rates, and fix h=1 (or select h on a holdout set that shares no datasets with Tab. 1). Rerun Table 1 with this fixed protocol, report per-seed results, and repeat for each alternative natural anchor (weekly/monthly). If no single fixed rule preserves the reported 6/8 advantage, or if the reported numbers sit near the best over anchors, the central claim is not supported as a zero-shot result.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim is that Freq-Synth needs only the target sampling rate (Sec. 4.2, App. B.1). However, a sampling rate does not uniquely determine a fundamental frequency: e.g., 15-minute data could anchor to daily (1/96), weekly (1/672), or monthly (1/2880); hourly data to 1/24, 1/168, or 1/720. App. B.1 gives only a 'closest common frequency' heuristic, and Tab. 4 says the anchor 'depends on dataset.' Meanwhile, the harmonic count h is varied in Fig. 5/App. C.1 with MSE on the evaluation datasets as the selection criterion. Thus, the reported 6/8 win over real data is consistent with a pipeline that uses the evaluation periodogram twice: once to choose the anchor and once to choose h. The real-data baseline has no analogous tuning freedom. The paper's own failures on Exchange and Weather show that anchor choice is consequential. Without a fixed a priori rule, the experiment does not establish that Freq-Synth, given only a sampling rate, beats real-data training in a genuine zero-shot protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a Fourier-analysis view of zero-shot and few-shot time series forecasting, introducing the notions of frequency confusion and frequency generalization, and then presents Freq-Synth, a synthetic data generator that creates sinusoidal training series from the target dataset's sampling rate. The method is evaluated by training six forecasting models (TTM, Timer, UniTime, Moment, GPT4TS, PatchTST) on Freq-Synth data versus a large real-data training pool, on eight LTSF benchmarks, and is compared with other synthetic generators (TimesFM, ForecastPFN, KernelSynth). The authors report that Freq-Synth outperforms real-data training in 6/8 datasets on average MSE/MAE, helps in few-shot fine-tuning, and is substantially cheaper to generate than competing synthetic baselines.","tokens_in":27586,"tokens_out":5461,"duration_ms":48842,"significance":"If the central claims hold, Freq-Synth would be a meaningful practical contribution: a lightweight, easy-to-code synthetic generator that needs only the target sampling rate could reduce the data and compute requirements of zero-shot and few-shot forecasting. The paper's strengths include the breadth of the evaluation (six model families, eight benchmarks, horizon lengths 96-720), the inclusion of a code snippet and an explicit generation-time comparison (0.1 s vs. 138 min for KernelSynth), and the diagnostic experiments on frequency confusion and frequency generalization, which are clearly explained and potentially useful for future work. However, the zero-shot protocol as reported leaves room for target-dependent tuning of the fundamental-frequency anchor and the harmonic count, and one baseline (TimesFM) is a self-implementation; these issues currently limit the strength of the headline 6/8 claim.","major_comments":[{"comment":"The central zero-shot claim that Freq-Synth needs only the target's sampling rate is not established by the current protocol. Table 4 defines the fundamental frequency as 'depends on dataset, see B.1', and App. B.1 describes only a 'closest common frequency' heuristic without specifying a unique mapping; for example, 15-minute data could be anchored to 1/96, 1/672, or 1/2880, and hourly data to 1/24, 1/168, or 1/720. Because the anchor is chosen per dataset, the comparison in Table 1 may use target-frequency information beyond the sampling rate, information that the real-data baseline does not receive. Please provide a fixed, a priori rule that maps each sampling rate to exactly one fundamental frequency, or restrict the claim accordingly.","section":"Sec. 4.2, App. B.1, Table 4"},{"comment":"The harmonic count h is effectively selected using the evaluation benchmarks. The main protocol uses three harmonic groups (h=1,2,3), but Fig. 5 reports MSE on ETTm1, ETTh2, ETTh1, Traffic, Weather, and Electricity as a function of h, and the choice h=3 is justified by those test-set results. Since the real-data baseline has no analogous test-informed tuning, this does not demonstrate a genuine zero-shot advantage. Please specify a fixed h, or a validation-based selection rule, before reporting the 6/8 comparison, and preferably show sensitivity to h with the chosen rule.","section":"Sec. 5.1, App. C.1, Fig. 5"},{"comment":"The TimesFM synthetic-data baseline is a self-implementation: App. D.3 states 'we provided results based on our implementation as the original implementation is not available.' The reported 12.7% MSE reduction over TimesFM in the known-sampling-rate block rests on this unvalidated baseline. Please either use the official TimesFM checkpoint, or clearly mark the comparison as an approximation and provide enough implementation detail to reproduce and assess it.","section":"Sec. 5.2, App. D.3"},{"comment":"The three-seed averages are reported without standard deviations or confidence intervals. Several per-dataset margins are small (e.g., ETTh2 and ETTm2 in Table 1), and the 6/8 claim is an aggregate statement; without variance information the reader cannot assess whether the observed wins are within noise. Please report seed-wise results or error bars for the main comparisons.","section":"Tables 1-3"},{"comment":"The assumption that the target distribution is dominated by a single fundamental frequency and its harmonics is acknowledged to fail on Exchange and Weather, and the method's failures on those datasets are explicitly documented. However, the abstract and conclusion state the method improves zero-shot forecasting generally. The stated scope should be narrowed to datasets satisfying the single-fundamental assumption, or the method should be evaluated on a broader set of multi-dominant datasets to determine where the approach breaks down.","section":"Sec. 4.2, App. B.4, App. E.1"}],"minor_comments":[{"comment":"The method is sometimes called 'Synth-Freq' and sometimes 'Fq-Synth' while the introduction uses 'Freq-Synth'; please use one name consistently throughout.","section":"App. B.2, Sec. 5.2, Table 7"},{"comment":"The x-axis labels ('0-0.5', '0.5-0.9', '0.9-', 'first choice second choice 3rd choice') are not explained in the caption; please clarify how the periodogram PCC bins and the 'choice' categories are defined.","section":"Fig. 2"},{"comment":"The phase formula 'tan^-1(bj/aj)' should specify the quadrant-aware arctangent (atan2) or restrict the phase to a branch, since the standard arctangent does not recover the correct quadrant for all amplitude signs.","section":"Eq. (2)"},{"comment":"Table 4 sets n=50,000 and A'=5, while Listing 1 uses n=250 and A_avg=1; please state which parameter settings produce the reported results and harmonize the code snippet with the table, or explain that the listing is only an illustrative skeleton.","section":"Table 4 vs. Listing 1"},{"comment":"The sentence 'Freq-Synth trains on a fraction (i.e.,1/14) of the data ForecastPFN, TimesFM, and Synth-Freq use' should state precisely which data volumes are compared and what '1/14' refers to; as written, the comparison is ambiguous.","section":"Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible and potentially useful contribution, but the zero-shot framing needs to be tightened. The main risk is not the method itself but the evaluation protocol: the fundamental-frequency anchor and the harmonic count appear to be selected with knowledge of the evaluation benchmark, and the TimesFM baseline is a self-implementation. These issues are fixable within the manuscript's scope, so I would not reject, but the claims should be scaled back or the protocol should be made strictly a priori. I would also ask the authors to report variance, since several comparisons are close."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here is my read of arXiv:2411.15743.\n\nThe headline claim—synthetic harmonic data built from the target's fundamental frequency beats real-data training in zero-shot forecasting—is probably true for a meaningful class of benchmarks, but the paper's 'only sampling rate' framing is not fully supported by the experiments.\n\nWhat's new: the paper names two failure modes, frequency confusion and frequency generalization, and shows empirically that both foundation and non-foundation forecasters suffer from them. That analysis is the most useful part. The Freq-Synth generator itself is a simple, easy-to-code heuristic: build a pool of sines at the fundamental frequency and its harmonics, then sum random subsets into multivariate series. The paper also shows a real efficiency win—generating a million points takes under a second, versus minutes or hours for TimesFM, ForecastPFN, and KernelSynth. And it evaluates broadly: six models, eight benchmarks, zero-shot and few-shot, plus comparison against other synthetic generators.\n\nThe soft spots are real but fixable. First, no error bars are reported; three seeds are run but only averages are shown. Second, the mapping from sampling rate to fundamental frequency is underdetermined. A 15-minute series could anchor to a daily, weekly, or monthly cycle; App. B.1 offers a 'closest common frequency' heuristic, but the choice is still a judgment call that seems made with knowledge of the benchmark. Third, the harmonic count h is varied in Fig. 5 with MSE on the evaluation datasets as the selection criterion, and the main results use h=3. That is a leak: the real-data baseline has no analogous tuning freedom. Fourth, the TimesFM synthetic-baseline result comes from a self-implementation, which is a smaller concern since ForecastPFN and KernelSynth are also compared.\n\nThe stress-test note is right that the claimed 6/8 advantage is conditional. If the authors fixed the anchor rule and h a priori—say, always use the daily harmonic for sub-daily sampling, and h=3—the results would be much more convincing. As written, the zero-shot protocol is not fully clean.\n\nWho should read this? People working on synthetic data for forecasting, and anyone building zero-shot or small-model forecasting pipelines. It deserves peer review; it's a solid empirical paper with a useful conceptual contribution. But I'd ask for a revision that closes the tuning leak, reports variance, and releases code.\n\nRecommendation: engage, with the caveat that the main result should be re-verified under a stricter protocol.","headline":"Plausible and useful empirical claim—synthetic harmonic training can beat real data—but the 'only sampling rate' framing is weakened by test-informed tuning of the anchor and harmonic count.","tokens_in":28043,"tokens_out":4784,"would_cite":true,"duration_ms":41198,"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":"Synthetic sine waves beat real data for zero-shot forecasts","keywords":["zero-shot forecasting","few-shot learning","synthetic data generation","Fourier analysis","periodogram","frequency generalization","frequency confusion","time series forecasting"],"falsifier":"Train Freq-Synth on a target series whose periodogram shows two or more comparable dominant peaks not related by integer multiples (e.g., a strong daily cycle plus a strong weekly cycle) and measure zero-shot error against real-data training: if the single-fundamental assumption is load-bearing, Freq-Synth should fail to match real data on such series, just as it already fails on the Exchange dataset.","tokens_in":27121,"feed_emoji":"📈","tokens_out":6437,"duration_ms":56107,"temperature":0.7,"pith_summary":"The paper argues that deep forecasting models—both large foundation models and ordinary networks—struggle in the same systematic way: they get worse when the training set contains many frequencies beyond the target's (frequency confusion), and they fail to generalize to frequencies they never saw during training (poor frequency generalization). To exploit this diagnosis, the authors propose Freq-Synth, a synthetic-data generator that, given only the sampling rate of the target series, builds training examples from sums of sine waves whose frequencies are the fundamental and its harmonics. Trained on just 5,000 synthetic samples, several models achieve lower error than when trained on roughly a thousand times more real data, beating real-data training on 6 of 8 zero-shot benchmarks and improving few-shot fine-tuning as well. If this holds, zero-shot forecasting no longer demands vast corpora or huge foundation models.","feed_headline":"Synthetic sine waves beat real data for zero-shot forecasts","feed_subtitle":"A generator using only the target's sampling rate trains models that beat real-data training on 6 of 8 benchmarks.","key_machinery":"The load-bearing tool is the scaled periodogram, $P(\\omega_j) = A_j^2$, which the paper uses to quantify how much of a signal's variance sits at each frequency, along with the two definitions it introduces: frequency confusion (performance degrades when the training set includes unrelated frequencies) and frequency generalization (ability to forecast unseen frequencies). The generative machinery is the pool of harmonics: given a fundamental frequency $\\bar{\\omega}$ estimated from the target's sampling rate, Freq-Synth samples frequencies from $\\Omega = \\{\\bar{\\omega}, 2\\bar{\\omega}, \\ldots, h\\bar{\\omega}\\}$, draws amplitudes from an exponential distribution and phases uniformly, and forms each variate as the sum of $l$ randomly chosen sines. This construction makes the training spectrum deliberately sparse and aligned with the presumed target spectrum, avoiding the confusion caused by extra frequencies.","core_discovery":"The central claim is that the factor governing effective learning in zero-shot time series forecasting is frequency alignment: a model transfers well when the training data shares the target's dominant periodic components, and poorly when the training data introduces unrelated frequencies or omits the target's fundamental. The paper demonstrates this with controlled sine-wave experiments and periodogram correlations, then turns the diagnosis into a generator: Freq-Synth constructs a pool of sinusoids at the target's fundamental frequency and its harmonics, with random amplitudes and phases, and sums a random subset per variate to form multivariate series. Because it needs only the sampling rate of the target (mapped to a common natural frequency), the method is parameter-light and fast. Empirically, training on Freq-Synth data outperforms training on real data in 6/8 benchmark datasets in zero-shot evaluation, reduces MSE by roughly 10–20% in few-shot fine-tuning across three tested models, and outperforms other synthetic generators while using a fraction of the data.","pith_inferences":["If frequency confusion is as widespread as the paper suggests, the common practice of mixing many heterogeneous datasets for pretraining may actively hurt forecasting performance; a natural extension is to filter or reweight pretraining corpora by periodogram similarity to the deployment frequencies.","The closest-common-frequency heuristic for mapping sampling rates to fundamentals could be replaced by a direct periodogram estimate on a handful of target samples, which would extend Freq-Synth to irregularly sampled or non-natural frequencies, as the paper hints in its appendix.","The 6/8 success rate may be sensitive to the evaluation suite: applying the same protocol to a broader set of benchmarks with diverse spectral shapes would clarify whether the reported advantage is a general law or specific to this set of eight datasets."],"forward_implications":["Zero-shot forecasting can be performed with a tiny, fast-to-generate synthetic corpus, removing the need to collect and clean large real-world training sets.","Non-foundation models trained on Freq-Synth data can match or beat foundation models on several benchmarks, suggesting compute-heavy pretraining may not be necessary for these tasks.","Frequency alignment, not domain identity, is the main driver of transfer success in time series, so dataset selection for training should prioritize periodogram similarity over sector labels.","Few-shot fine-tuning also benefits: pre-training on synthetic data then fine-tuning on 10% of the target data outperforms fine-tuning from real-data pretraining on the tested models."],"supporting_citations":[{"why":"Introduces the periodogram used throughout to quantify frequency content of signals.","marker":"Schuster, 1898"},{"why":"Provides the Fourier representation of discrete time series that underlies the harmonic construction.","marker":"Shumway & Stoffer, 2000"},{"why":"ForecastPFN is the main synthetic-only zero-shot baseline that Freq-Synth must outperform.","marker":"Dooley et al., 2024"},{"why":"TimesFM supplies both a synthetic-data baseline and a pre-trained model used in the frequency generalization analysis.","marker":"Das et al., 2024"},{"why":"KernelSynth provides the Gaussian-process synthetic data baseline compared in the unknown-sampling-rate setting.","marker":"Ansari et al., 2024"},{"why":"The Monash archive is the source of real training data used in the zero-shot comparison against synthetic data.","marker":"Godahewa et al., 2021"}],"fun_headline_variants":["Frequency alignment beats data volume in zero-shot forecasts","Synthetic sine waves outperform real data for zero-shot prediction","Freq-Synth: Sampling rate alone improves zero-shot forecasting","Why frequency matching matters in time series forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the target series is dominated by a single fundamental frequency, estimated from its sampling rate, and that its harmonics; if the series has a broad or multi-dominant spectrum, the synthetic data cannot represent it and the claimed advantage disappears.","fun_headline_variants_meta":{"raw":{"variants":["Frequency alignment beats data volume in zero-shot forecasts","Synthetic sine waves outperform real data for zero-shot prediction","Freq-Synth: Sampling rate alone improves zero-shot forecasting","Why frequency matching matters in time series forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1366,"prompt_tokens":966,"completion_tokens":400,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":337}},"tokens_in":582,"tokens_out":400,"duration_ms":4328,"temperature":1.0,"reasoning_tokens":337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:56:21.692211+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Freq-Synth on a target series whose periodogram shows two or more comparable dominant peaks not related by integer multiples (e.g., a strong daily cycle plus a strong weekly cycle) and measure zero-shot error against real-data training: if the single-fundamental assumption is load-bearing, Freq-Synth should fail to match real data on such series, just as it already fails on the Exchange dataset.","supporting_citations":[],"review_version":1}