{"id":"e2b63412-2589-4af7-bde1-d127f978f07b","arxiv_id":"2412.05244","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A wavelet-based tokenizer that lets an autoregressive transformer forecast quantized wavelet coefficients instead of raw values, improving accuracy and generalization on time series benchmarks.","lead":"WaveToken turns time series into tokens by decomposing them with wavelets and quantizing the coefficients, then trains a T5-style model to forecast the coefficient sequence. It reports strong accuracy and zero-shot generalization across 42 datasets while using a much smaller vocabulary than prior time series foundation models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Horizon wavelet coefficients are ill-defined without boundary handling; IDWT reconstruction from context-only plus predicted coefficients may be incoherent.","rationale":"The paper's empirical evaluation is broad, the qualitative attention analysis is a useful mechanism check, and the compact-vocabulary argument is well motivated. However, the central claim requires that forecasting in wavelet-coefficient space can be inverted into an accurate time-domain forecast. The most load-bearing gap is that the horizon coefficient sequence is not well-defined: the DWT couples boundary samples across the context/horizon cut, and the paper neither specifies the boundary conditions used nor validates oracle reconstruction through the proposed pipeline. The reader's reconstruction concern is close, but our formulation is more specific and testable. If the oracle test fails, the benchmark results are contaminated by reconstruction artifacts; if it passes, the concern is answered and the conditional acceptance stands. I do not see intentional overclaiming; the paper's own tables show Chronos-Large sometimes winning, but the aggregate-rank claim is the one that needs this reconstruction sanity check before being taken at face value.","tokens_in":47374,"tokens_out":6267,"duration_ms":66484,"concrete_test":"Implement the exact tokenizer (same wavelet family, decomposition level, quantization bins, and boundary condition used in the paper; if unspecified, test zero-padding, symmetric, and periodic). For several Benchmark I/II series of length C+H, compute the oracle horizon tokens by applying the tokenizer to the full series and extracting the last H tokens in the concatenated order. Then simulate inference: compute context tokens from x_{1:C} only, append the oracle horizon tokens, dequantize to bin centers, apply the IDWT, and measure MASE/VRSE on the horizon against the true future. Compare against the reference reconstruction obtained by applying the IDWT to the full-series coefficients.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 defines the training target as tokens z_{C+1:C+H} for the horizon and states that at inference the dequantized predicted coefficients are pushed through the IDWT to obtain a forecast. But the maximally decimated DWT is a global transform: a wavelet coefficient at a given scale and location is computed from a filter window spanning neighboring samples. The paper never specifies how z_{C+1:C+H} is obtained for training (presumably from the DWT of the full concatenated series), which boundary conditions are used (the footnote in Section 3.2 only notes that longer filters may add coefficients), or how the IDWT can produce a coherent continuation when the context coefficients come from the DWT of x_{1:C} alone rather than from the DWT of the full length-(C+H) series. Concatenating context-only coefficients with independently predicted horizon coefficients does not in general yield a valid wavelet representation of any real-valued series; the IDWT then introduces boundary artifacts that are not modeled. The central 'language of time series' claim depends on faithful invertibility of the tokenization in this autoregressive setting, yet no experiment or analysis verifies even the oracle version: using ground-truth horizon coefficients in the proposed pipeline and measuring reconstruction error. Without this, the benchmark numbers cannot be cleanly attributed to learned forecasting rather than to reconstruction artifacts.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes WaveToken, a wavelet-based tokenizer for time-series foundation models. The pipeline rescales each input series, applies a maximally decimated discrete wavelet transform (DWT), optionally thresholds the detail coefficients, quantizes all coefficients into a shared 1024-token vocabulary via Freedman-Diaconis binning, and concatenates approximation and detail coefficient tokens for autoregressive training of a T5 encoder-decoder. At inference, autoregressively sampled tokens are dequantized to bin centers and mapped back to the time domain by the inverse DWT. The method is evaluated on the Chronos benchmark collection: 15 in-domain and 27 zero-shot datasets, against statistical, task-specific deep-learning, and pretrained time-series foundation baselines, using WQL, MASE, and VRSE. The paper reports the best average rank for WaveToken across both benchmarks and all three metrics, along with ablations of vocabulary size, wavelet family, decomposition level, and thresholding, and a qualitative cross-attention analysis on trends, spikes, and non-stationary signals.","tokens_in":47576,"tokens_out":7038,"duration_ms":71176,"significance":"If the reported results hold, WaveToken would be a valuable demonstration that a compact wavelet-coefficient vocabulary can match or beat much larger scalar-quantization vocabularies and patch-based tokenizations for time-series forecasting, with particularly attractive zero-shot behavior. The evaluation is unusually thorough: raw per-dataset tables, three complementary metrics, ablations, a long-horizon extension, and attention analyses are included. The central modeling assumption, however, is the faithful invertibility of the coefficient language under the context/horizon split used at inference, and that assumption is not tested; the paper also omits the boundary conditions needed to make the reconstruction well-defined. The in-domain aggregate scores do not support the unqualified 'better accuracy' phrasing in the abstract. With these two issues addressed, the contribution would be a solid empirical advance in time-series tokenization.","major_comments":[{"comment":"Section 3.3 defines the training target as horizon coefficient tokens z_{C+1:C+H} and reconstructs forecasts by applying the IDWT to dequantized predicted coefficients, but the manuscript never specifies how the horizon coefficients are computed for training, which boundary conditions are used in the DWT/IDWT, or why the IDWT of context coefficients from x_{1:C} concatenated with independently predicted horizon coefficients is a valid wavelet representation of a length-(C+H) series. Since the maximally decimated DWT is a global filter-bank transform, coefficients at the context/horizon boundary depend on samples on both sides of the split, and an arbitrary concatenation of coefficient groups need not correspond to any real-valued series. This is load-bearing: if the IDWT introduces boundary artifacts, the benchmark numbers could reflect reconstruction bias rather than learned forecasting. A minimal fix is to add an oracle experiment that feeds ground-truth horizon coefficients through the proposed dequantization/IDWT pipeline and reports reconstruction error, and to specify the boundary mode (e.g., periodic, symmetric, zero) and the coefficient-count mapping from horizon length H to tokens at the chosen decomposition level.","section":"Section 3.3"},{"comment":"On the in-domain benchmark, WaveToken-Large's aggregate relative scores are 0.569 for WQL and 0.698 for MASE, while Chronos-Large achieves 0.564 and 0.695, respectively. The text in Section 4.2 acknowledges this exception, but the abstract and introduction state that WaveToken 'provides better accuracy than recently proposed foundation models' and 'always performs on-par or better than all other baselines.' That claim is too strong as written. I recommend restating the headline as 'competitive on WQL/MASE and better on VRSE and best average rank,' or reporting statistical significance of the aggregate differences.","section":"Benchmark I, Tables 1-2 and Figure 3"}],"minor_comments":[{"comment":"The notation q(w) = i 1{e_{i-1} ≤ w < e_i} is unusual: since q(w) should equal the bin index i when the condition holds, the indicator should be part of the definition rather than a multiplier. Please clarify.","section":"Section 3.2, quantization formula"},{"comment":"The notation z1:C is used both for the coefficient-token vector and for a temporal indexing scheme, and z_{C+1:C+H} conflates coefficient positions with original time indices; this makes the loss equation hard to parse. Please use distinct notation for coefficient indices.","section":"Section 3.3, notation"},{"comment":"The thresholding ablation changes ranking between the 1-GPU and 8-GPU settings, and the final choice (no thresholding) is made on the same evaluation benchmarks used for the main comparisons; please report these hyperparameter selections as part of a validation procedure rather than on the test benchmarks.","section":"Section 4.4 and Figure 8"},{"comment":"The aggregate relative scores are reported without confidence intervals or standard errors, although the underlying results are averaged over three seeds; adding variability estimates would strengthen the comparison.","section":"Figures 3 and 4"},{"comment":"The qualitative cross-attention analysis is suggestive but not evaluated quantitatively; consider adding a simple statistic, such as attention concentration at spike positions, so the claim that the model 'attends to the right coefficients' is testable.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The boundary-condition gap is the main technical risk in the paper. If the authors can provide the oracle reconstruction experiment and clarify the DWT/IDWT split, I would view the paper as publishable. I also recommend softening the abstract's 'better accuracy' claim. No concerns about author conduct or citation patterns beyond noting that concurrent wavelet tokenizers are cited but not compared in detail."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:2412.05244. First, it genuinely does something new: a maximally decimated DWT as the tokenizer for time series foundation models, with a 1024-token vocabulary. Second, the empirical case is strong but the paper overclaims, and there's a technical gap around boundary handling that needs airing before the strongest interpretation is taken.\n\nThe good stuff is real. The method is simple and well described: z-score scaling, single-level DWT with a biorthogonal filter, quantization into a shared vocabulary, next-token prediction on the concatenated coefficients with a T5 backbone. The empirical work is careful for the field's current standards: 42 benchmark datasets, three metrics (WQL, MASE, VRSE), raw per-dataset tables, three-seed averaging for WaveToken and Chronos, ablations on vocabulary size, wavelet family, decomposition level, and thresholding rule. The attention analysis in Section 4.3 is a nice piece of evidence that the model is actually exploiting the coefficient structure rather than ignoring it. The zero-shot results, where WaveToken-Large gets the best average rank on all three metrics, are the most compelling part.\n\nNow the soft spots, in proportion.\n\nThe biggest one is the DWT boundary issue. Section 3.3 says the model forecasts tokens for the horizon and then the IDWT turns the whole coefficient vector into a forecast. But it never says how the horizon tokens are defined for training. If they come from the DWT of the full context-plus-horizon series, then the context coefficients seen at inference (from the context alone) are not the same coefficients the model saw during training, because the DWT is a global transform with filters spanning the boundary. Concatenating context-only coefficients with independently predicted horizon coefficients does not generally give a valid wavelet representation of any real series. There is no oracle experiment: take ground-truth horizon coefficients, run the pipeline, and measure reconstruction error. That test would settle whether the benchmark numbers are forecasting performance or reconstruction artifacts. This is not a minor omission; it sits at the center of the method's premise.\n\nSecond, the abstract says \"better accuracy than recently proposed foundation models,\" but the in-domain tables show Chronos-Large beating WaveToken-Large on both WQL and MASE, and the aggregate relative scores are within a hair (0.564 vs 0.569; 0.695 vs 0.698). Without error bars or statistical tests, the small differences are plausibly noise. The zero-shot rank results are stronger, but they don't rescue the absolute claim.\n\nMinor issues: no code or weights released; key hyperparameters (vocabulary, family, level, bin range) selected on the same suite used for evaluation, with only the zero-shot split partially mitigating that; and the paper itself acknowledges slower inference.\n\nWho should read this: anyone working on tokenization for continuous signals or time series foundation models. It deserves a serious referee. I'd send it to review with a request for major revision: specify the boundary condition, add the oracle reconstruction check, and reword the performance claims. If the boundary issue resolves, this becomes a solid reference point for the field.","headline":"A genuinely new wavelet tokenizer for time series foundation models with a strong zero-shot benchmark, but the DWT boundary handling is unspecified and the in-domain superiority claim is not supported by the paper's own tables.","tokens_in":48174,"tokens_out":6104,"would_cite":true,"duration_ms":60678,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["42C40","62M10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that quantizing wavelet coefficients into a 1,024-token vocabulary gives pretrained time-series forecasters better accuracy and zero-shot generalization than scalar quantization or patching, with the best average rank…","keywords":["wavelet tokenization","time series forecasting","foundation models","discrete wavelet transform","quantization","zero-shot generalization","autoregressive transformer","T5 encoder-decoder"],"falsifier":"On a test set of high-dynamic-range series dominated by sparse sharp spikes, reconstruct the context series by quantizing its DWT coefficients to the 1,024-bin codebook, mapping to bin centers, and applying the inverse DWT; measure the relative reconstruction error. If the error is large for the very patterns the paper highlights, the fidelity assumption is broken. A second check: vary vocabulary size upward; the paper predicts accuracy flattens beyond 1,024, so a benchmark where larger codebooks keep improving would contradict the compression claim.","tokens_in":47139,"feed_emoji":"📈","tokens_out":5873,"duration_ms":50594,"temperature":0.7,"pith_summary":"The paper sets out to answer what the right discrete \"language\" is for feeding real-valued time series to a pretrained transformer. It argues that quantized wavelet coefficients form that language: a 1,024-token vocabulary that separates coarse trends from fine detail, so an autoregressive T5 model can learn to forecast coefficients from coarse to fine. If the claim holds, wavelet-based tokenization beats scalar quantization (Chronos) and patching (TimesFM, Moirai) in both accuracy and zero-shot generalization, with the best average rank on 42 datasets across three metrics. A sympathetic reading is that the paper aims to establish wavelets as a more information-efficient tokenizer for general-purpose forecasting models.","feed_headline":"Wavelet tokens beat bigger vocabularies on 42 forecasting sets","feed_subtitle":"A 1,024-token wavelet codebook gives pretrained forecasters top average rank across three metrics.","key_machinery":"The load-bearing object is the maximally decimated discrete wavelet transform (DWT), a filter-bank decomposition that maps a length-$N$ signal to $N$ coefficients split into one approximation band $\\{a_k\\}_J$ and detail bands $\\{d_k\\}_j$, $j=1,\\dots,J$, concentrating signal energy in few coefficients. Around it sits a pipeline: z-score scaling (chosen because the DWT is not translation-invariant), optional coefficient thresholding (VisuShrink, CDF, FDRC, or none—none wins in the final 8-GPU runs), quantization to bins sized by the Freedman–Diaconis rule, and concatenation in coarse-to-fine order $[a_J,d_J,\\dots,d_1]$ so the model predicts from coarser to finer resolutions. This machinery turns a continuous series into a token stream that preserves autoregressive structure within each band and gives the cross-attention a natural quadrant structure.","core_discovery":"WaveToken's central claim is that tokenizing time series in the space of time-localized frequencies—via a maximally decimated discrete wavelet transform followed by thresholding, Freedman–Diaconis binning, and coarse-to-fine concatenation—yields a compact yet expressive codebook that simplifies next-token prediction. The paper reports that WaveToken outperforms existing pretrained forecasters (including Chronos, TimesFM, Moirai, and Lag-LLama) and performs on par or better than task-specific deep learning models, while using only 1,024 tokens, one quarter of Chronos's vocabulary. It also claims best average rank across all datasets for weighted quantile loss, mean absolute scaled error, and visual relative squared error, and shows qualitative mastery of exponential trends, sparse spikes, and non-stationary multi-frequency signals that competing models miss.","pith_inferences":["Because dequantization to bin centers is a form of denoising, the model may be implicitly learning a smoothed forecast; an explicit reconstruction-error-aware training loss could push accuracy further than cross-entropy alone.","The DWT's O(N) cost means token counts need not grow with context length, so wavelet tokenization is a natural route to long-context forecasting without increasing sequence length—a direction the paper leaves open.","Continuous signals with naturally sparse high-frequency content—audio, ECG, sensor streams—are plausible next targets; the compression argument transfers whenever a wavelet basis concentrates energy.","The reported VRSE advantage may be partly baked into the design, since VRSE rewards matching the frequency-domain shape that wavelet coefficients encode; re-ranking on pointwise metrics alone would test how much of the gain is frequency-specific."],"forward_implications":["Pretrained forecasting can run with a 1,024-token codebook instead of 4,096, shrinking embedding and softmax costs without sacrificing accuracy.","Zero-shot generalization improves: WaveToken achieves the best average rank across all 42 datasets on WQL, MASE, and VRSE, including on datasets never seen in training.","The model captures exponential trends, sparse spikes, and time-varying frequencies that scalar-quantized and patch-based foundation models systematically miss.","Coarse-to-fine coefficient ordering induces interpretable attention structure: the decoder attends to approximation coefficients when predicting approximations and to detail coefficients when predicting details.","The tokenizer is architecture-agnostic enough to pair with any autoregressive transformer trained by next-token prediction, not just the T5 variant used here."],"supporting_citations":[{"why":"Provides the Chronos baseline, the T5-based training framework, the dataset collection, and the data augmentations that WaveToken reuses.","marker":"Ansari et al. (2024)"},{"why":"Supplies the T5 encoder-decoder architecture that WaveToken pairs with its tokenizer.","marker":"Raffel et al. (2020)"},{"why":"Supplies the wavelet theory and the O(N) filter-bank DWT that the tokenizer is built on.","marker":"Mallat (2009)"},{"why":"Foundational reference for the discrete wavelet transform and the biorthogonal families evaluated in ablations.","marker":"Daubechies (1992)"},{"why":"Gives the bin-size rule used to quantize wavelet coefficients into the vocabulary.","marker":"Freedman & Diaconis (1981)"},{"why":"Defines VisuShrink, the thresholding method compared in ablations and initially selected before no-thresholding won in the final setting.","marker":"Donoho (1995)"},{"why":"TimesFM, the patch-based pretrained forecaster that WaveToken is compared against; its patching tokens are the main alternative representation.","marker":"Das et al. (2023)"},{"why":"Moirai, a patch-based universal forecaster used as a baseline in both in-domain and zero-shot benchmarks.","marker":"Woo et al. (2024)"},{"why":"Defines VRSE, one of the three evaluation metrics used to compare frequency content of forecasts.","marker":"Posam et al. (2024)"},{"why":"Defines MASE, the point-forecast metric used in the benchmark aggregations.","marker":"Hyndman & Koehler (2006)"}],"fun_headline_variants":["Wavelet tokenizer shrinks vocabulary, lifts forecast accuracy","1,024 wavelet tokens outrank bigger forecast codebooks","WaveToken: frequency-space tokens beat vocab size on 42 sets","Compact wavelet codebook tops forecasting benchmarks","Wavelet tokens: small vocab, top rank across 42 datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that mapping predicted coefficient tokens to bin centers and running the inverse discrete wavelet transform reconstructs the future time series faithfully enough for accurate forecasts; the paper gives no bound on how quantization error or autoregressive coefficient errors amplify through the inverse transform, and it does not explicitly control boundary effects where the context coefficients meet the horizon coefficients.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet tokenizer shrinks vocabulary, lifts forecast accuracy","1,024 wavelet tokens outrank bigger forecast codebooks","WaveToken: frequency-space tokens beat vocab size on 42 sets","Compact wavelet codebook tops forecasting benchmarks","Wavelet tokens: small vocab, top rank across 42 datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000574,"raw_usage":{"total_tokens":2719,"prompt_tokens":959,"completion_tokens":1760,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":1679}},"tokens_in":575,"tokens_out":1760,"duration_ms":12014,"temperature":1.0,"reasoning_tokens":1679,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:49:35.086408+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a test set of high-dynamic-range series dominated by sparse sharp spikes, reconstruct the context series by quantizing its DWT coefficients to the 1,024-bin codebook, mapping to bin centers, and applying the inverse DWT; measure the relative reconstruction error. If the error is large for the very patterns the paper highlights, the fidelity assumption is broken. A second check: vary vocabulary size upward; the paper predicts accuracy flattens beyond 1,024, so a benchmark where larger codebooks keep improving would contradict the compression claim.","supporting_citations":[{"cited_title":"A Wavelet Tour of Signal Processing: The Sparse Way","cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet theory and the O(N) filter-bank DWT that the tokenizer is built on."},{"cited_title":"On the histogram as a density estimator: L 2 theory","cited_arxiv_id":null,"evidence_quote":"Gives the bin-size rule used to quantize wavelet coefficients into the vocabulary."},{"cited_title":"De-noising by soft-thresholding","cited_arxiv_id":null,"evidence_quote":"Defines VisuShrink, the thresholding method compared in ablations and initially selected before no-thresholding won in the final setting."},{"cited_title":"Difffind: Discovering differential equations from time series","cited_arxiv_id":null,"evidence_quote":"Defines VRSE, one of the three evaluation metrics used to compare frequency content of forecasts."}],"review_version":1}