{"id":"1ada0c7d-b766-46a9-a16a-93ec172b9d23","arxiv_id":"2412.11171","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LTG decomposes series into trend and seasonal parts, learns latent factors with a conditional beta-VAE and domain regularization, and feeds them to a forecaster to improve generalization to unseen domains.","lead":"This paper introduces LTG, a method that learns shared and domain-specific patterns from time series data to help forecasting models predict on unseen domains. It is relevant for businesses that forecast web traffic, sales, or energy use in new regions or platforms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decoder conditioning on DomID may not survive encoder-only test-time inference; ablation w/o Cond degrades, but the mechanism is unverified.","rationale":"The reader's weakest assumption is that the domain-specific information injected through the decoder remains useful in the latent vector once the decoder is removed at test time. This is exactly the concern I identify as most load-bearing: the mechanism of the paper hinges on it. I agree with the reader's identification and the CONDITIONAL verdict. The paper has genuine strengths: a clean plug-in framework, five datasets, ablations including the relevant LTG(w/o Cond) and LTG(w/o Spe) variants, and a self-acknowledged limitation in Section 6. However, the central claim of superiority is not established because the reported gains are small, lack significance testing, and are partially inconsistent (LTG(w/o Cond) outperforms LTG on Web-traffic Q(0.5) with DeepAR). I do not see an internal logical contradiction; the issue is that the proposed mechanism is empirically unverified. A concrete probe of domain information in the inference-time latent vectors would settle whether the mechanism works as claimed. This does not change the reader's verdict (CONDITIONAL), because the concern is addressable and the core framework may still be useful even if the domain-specific explanation is softened.","tokens_in":19199,"tokens_out":1587,"duration_ms":13561,"concrete_test":"Measure the domain-specific information actually present in the test-time latent vectors. Concretely: train a linear probe (or nearest-centroid classifier) on zspecific from held-out training domains and evaluate on test domains; if probe accuracy is near chance, the domain-specific signal does not survive in the encoder output, undermining the claimed mechanism. Additionally, compare the full LTG against a variant that feeds only zshared to the forecasting decoder (already partially reported as LTG(w/o Spe) in Table 4) but with the same latent pretraining; if the gap between LTG and LTG(w/o Spe) is within noise on most datasets, the DomID conditioning contributes little beyond the shared latent representation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that LTG achieves improved generalization via latent vectors that encode both domain-shared and domain-specific temporal dependencies, produced by a Conditional β-VAE whose decoder is conditioned on DomID (Eq. 5). At test time the authors state that 'we only need the encoder to generate the latent vectors' (Section 4.2), so the DomID-conditioned decoder is discarded. In a VAE, the encoder posterior q(z|x) is trained jointly with the decoder p(x|z, DomID); removing the decoder changes the training objective at inference and there is no guarantee that the learned z retains domain-specific information. In fact, because the encoder is domain-agnostic (DomID is not an encoder input), the information-theoretic pressure is for z to encode only the shared reconstruction residual, with domain specifics pushed into the decoder's DomID conditioning. The domain regularization loss (Eq. 9) explicitly encourages zspecific to differ across domains, but this is a soft loss and its effect on the final forecast is not quantified. The ablation LTG(w/o Cond) in Table 4 removes the DomID decoder conditioning and shows some degradation on Web-traffic (e.g., 0.1635 vs 0.1717 Q(0.5) for DeepAR is actually better without conditioning on one metric), which is inconsistent and does not robustly demonstrate that the proposed mechanism is load-bearing. Crucially, no experiment verifies that the learned z vectors, as used at inference time, actually carry domain-specific information that improves forecasting on unseen domains beyond the shared latent parts.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Latent Temporal Generalization (LTG), a plug-in framework for domain generalization in probabilistic time series forecasting. LTG first decomposes each input series into trend-cyclical and seasonal components, models each component with a conditional beta-VAE whose decoder is conditioned on a domain identifier, and adds a domain-regularization term that encourages part of the latent vector to be shared across domains and the remainder to be domain-specific. The resulting latent vector is concatenated with the raw input through a linear layer and fed into an arbitrary forecasting decoder, which is trained in a second stage while the VAE encoders are fine-tuned. Experiments on five datasets (web traffic, retail, finance, energy) compare LTG applied to DeepAR, WaveNet, DLinear, and GPT4TS against IDGM, Cedar, and UniTime. The paper reports improved range accuracy in several configurations, especially with WaveNet, and includes ablations, t-SNE latent-space visualizations, a forecast-horizon sensitivity study, and a runtime analysis.","tokens_in":19483,"tokens_out":7629,"duration_ms":67743,"significance":"If the main empirical claim were fully established, LTG would be a useful and modular contribution: it is architecture-agnostic, requires no label information, and adds only a lightweight pretraining stage. The experimental design is broad, covering four base forecasters and five real-world datasets, and the two-stage training scheme is simple to implement. However, the claim that LTG 'achieves the overall best performance' is currently supported only by selected comparisons of means over five seeds; several configurations show performance degradation, and no significance tests or aggregate rankings are reported. The central mechanism, namely that domain-specific information is preserved in the latent vector after the DomID-conditioned decoder is removed, is not verified, and the latent-space visualization confirms an objective that is explicitly enforced by Equation (9), so it does not independently validate disentanglement. These gaps make the evidence for the core claim conditional rather than conclusive.","major_comments":[{"comment":"The paper's headline claim that LTG achieves the overall best range accuracy is not substantiated by the reported results. On Web-traffic, LTG degrades both base models: DeepAR Q(0.5) increases from 0.155 +/- 0.030 to 0.172 +/- 0.014 and Q(mean) from 0.122 +/- 0.018 to 0.179 +/- 0.038; DLinear Q(0.5) increases from 0.152 +/- 0.003 to 0.172 +/- 0.032 and Q(mean) from 0.119 +/- 0.008 to 0.145 +/- 0.040. The Power-cons dataset, where LTG shows its largest gains, has only three domains total, so the 20% test split leaves one test domain per random split. No significance tests, confidence intervals, or win-rate/rank summaries are reported, so it is unclear whether the differences over five seeds support the conclusion. Please add paired significance tests across seeds, report per-configuration wins and losses, and discuss the small number of test domains.","section":"Section 5.2, Tables 2 and 3"},{"comment":"The proposed decoder-conditioning mechanism is not verified at inference time. The text states that 'we only need the encoder to generate the latent vectors' at test time, and the encoder is not given DomID; in a VAE the decoder conditions z through reconstruction, so the model can satisfy its training objective by moving domain-specific information into the decoder and leaving z largely shared. The domain-regularization loss in Eq. (9) softly encourages zspecific to differ across domains, but no quantitative measurement (e.g., a domain classifier trained on zspecific, or mutual information between zspecific and DomID) demonstrates that the encoder-only latent vectors actually carry domain-specific information. The ablation LTG(w/o Cond) is inconsistent with the proposed mechanism: removing DomID from the decoder improves DeepAR on Web-traffic Q(0.5) from 0.1717 to 0.1635, while degrading some Stock-volume metrics. The authors' own limitation statement in Section 6 that latent vectors are used only as additional input knowledge further suggests the mechanism is not load-bearing. Please provide a direct test of the information content of zspecific as used at inference.","section":"Section 4.2 and Table 4"},{"comment":"The t-SNE visualizations are presented as evidence that domain-shared and domain-specific components are separated, but Equation (9) explicitly imposes this separation: it minimizes pairwise distances among zshared samples and maximizes pairwise distances between zspecific samples from different domains. The figure therefore confirms the regularizer rather than providing independent evidence. The disentanglement claim in the abstract and contributions should be supported with a quantitative measure (e.g., MIG or DCI on the learned latents, or an evaluation of whether zspecific predicts domain labels and zshared does not).","section":"Section 5.4.1 and Equation (9)"},{"comment":"The pretraining objective as written is ambiguous. Equation (6) includes an L2 reconstruction term for (xhat_t + xhat_s) - x, and Equations (7)-(8) add log p_theta_t(xt|zt,DomID) and log p_theta_s(xs|zs,DomID) as separate quantities. If the log p terms are Gaussian log-likelihoods, they are themselves reconstruction terms up to scaling, so the objective double-counts the same fit; if they are intended to be the ELBO reconstruction terms, then Equation (6) duplicates them. This makes the exact training signal in Stage 1 (Section 4.5) unclear and is an obstacle to reproducing the method. Please rewrite the objective to specify one reconstruction term and state how the L2 term and the log-likelihood terms are combined.","section":"Equations (6)-(8)"}],"minor_comments":[{"comment":"Several entries in Table 2 use a comma instead of a period as the decimal separator (e.g., '0.016, .001'), and the spacing of the mean and standard deviation is inconsistent across columns; please unify the formatting.","section":"Table 2"},{"comment":"The definition of the split index as index = alpha * dz should specify how non-integer values are handled (e.g., floor or round), since alpha values such as 0.25 with odd dimension dz are used in the hyperparameter search.","section":"Section 4.3"},{"comment":"The TimeLLM rows in Table 5 are missing Q(0.5) values for Favorita-cat and Stock-volume; please explain why those entries are absent or provide the values.","section":"Appendix E.4, Table 5"},{"comment":"There are typographical issues, including 'Contidional' in the Section 4.5 heading and inconsistent spacing in 'V AE' and 'beta-V AE' in the abstract and introduction.","section":"Section 4.5 and Abstract"},{"comment":"The phrase 'reverse instance norm' should be 'reversible instance normalization', and the description should state whether the normalization is applied to the input of the VAE, the forecasting decoder, or both.","section":"Appendix C"},{"comment":"No code is released, and the paper does not mention a plan to do so; given the number of hyperparameters and the two-stage training procedure, releasing code or a detailed pseudocode would substantially improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible empirical skeleton and the framework is within scope for a machine learning venue, but the central mechanism is currently under-supported. The strongest evidence comes from WaveNet and Power-cons, and the Power-cons result rests on very few test domains. The t-SNE analysis as presented is partly circular because it confirms an objective explicitly enforced by Eq. (9). In revision, I would look for a direct test of whether zspecific carries domain information, a significance-aware summary of wins and losses, and a clarified VAE objective. I would not reject the paper outright, but I would not accept it in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate extension of the prior Cedar line, not a breakthrough. It combines decomposition, a conditional beta-VAE with the decoder conditioned on domain IDs, and a pairwise domain regularization loss. The combination is new for time series forecasting DG, the implementation is clear, and they test on five real datasets, two of which are new to this problem. When it works, it works well: WaveNet+LTG is substantially better than WaveNet+Cedar on Web-traffic and Power-cons. The paper is honest about where it is weak, including a limitation paragraph saying the latent vectors are only additional input knowledge and may not always help.\n\nOne soft spot is that the central claim of \"overall best performance\" is not fully supported by the tables. On Web-traffic, DeepAR+LTG and DLinear+LTG are worse than the base models, and on several datasets the differences are tiny. There are no significance tests, so I cannot tell which gaps are noise. The paper says it ran 5 seeds; it should report error bars or paired tests. No code is released, which matters for a plug-in claim.\n\nThe mechanism story is the other soft spot. The decoder is conditioned on DomID during VAE pretraining, but at test time only the encoder produces latents. The authors' own ablation, LTG(w/o Cond), removes that conditioning and gives mixed results: on Web-traffic it actually improves DeepAR. The t-SNE in Section 5.4.1 shows shared/specific separation, but Eq. 9 explicitly enforces that separation, so the visualization is partly circular. They never quantify disentanglement or directly test whether the domain-specific part of the latent carries forecasting value at inference. The stress-test note is right to flag this. It is not fatal, because stage-2 fine-tuning of the encoder on the forecasting loss can reshape the latents, but the paper should probe this directly.\n\nI disagree with the reader on one point: the use of known components is not a flaw by itself, and the paper does cite the prior beta-VAE and CVAE work. The novelty is modest but real. I also think the paper is more self-aware than the reader gives it credit for; the conclusion's limitation paragraph is a genuine caveat, not a hedge.\n\nBottom line: this deserves a serious referee and likely a conditional accept after revision. I would ask for significance testing, code, disentanglement metrics, and an experiment that either verifies or falsifies the claim that domain-specific latent components improve forecasting on unseen domains. I would cite it if I worked on time series DG, mainly because it is a clean statement of an architecture others will build on.","headline":"Reasonable plug-in for time series DG with mixed evidence; the decoder-conditioning mechanism needs direct verification before the 'overall best' claim can hold.","tokens_in":20023,"tokens_out":2919,"would_cite":true,"duration_ms":27072,"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":"A plug-in framework called LTG learns latent temporal factors and reports better forecasting accuracy on unseen domains than state-of-the-art domain-generalization methods.","keywords":["time series forecasting","domain generalization","variational autoencoder","latent space","disentanglement","conditional VAE","domain shift","probabilistic forecasting"],"falsifier":"Train LTG, then at test time replace every latent vector with a version from which the supposed domain-specific half has been removed (or shuffled across domains); if quantile loss on unseen domains does not change, the domain-specific mechanism carries no information and the claimed separation is not occurring. A companion check: permute the DomID fed to the decoder during pretraining and measure whether downstream forecasting accuracy drops.","tokens_in":18965,"feed_emoji":"📈","tokens_out":7354,"duration_ms":59070,"temperature":0.7,"pith_summary":"Time series forecasting models often fail when asked to predict on new but related domains, such as a website or store never seen during training. This paper argues that existing domain-generalization methods align feature distributions or use label information but ignore the latent factors that actually govern temporal dependencies. It introduces LTG, a plug-in framework that learns these latent factors with a conditional β-variational autoencoder on trend and seasonal components, splitting the latent vector into shared and domain-specific parts. The paper reports that LTG improves range accuracy on unseen domains over strong baselines across five real-world datasets, and that it can be attached to different base forecasters with modest added training cost.","feed_headline":"Plug-in VAE module improves forecasts on unseen domains","feed_subtitle":"Learns shared and domain-specific temporal patterns; tested on web, retail, finance, and energy data.","key_machinery":"The load-bearing object is the Conditional $\\beta$-VAE: an autoencoder whose encoder estimates a latent posterior $q_\\phi(z|x)$ for each decomposed component, and whose decoder reconstructs the component from $z$ together with a one-hot domain identifier $\\mathrm{DomID}$. Because the domain ID enters only the decoder, the encoder stays domain-agnostic and the latent vector must carry both shared and domain-specific information; a domain-regularization loss then splits $z$ into a shared part and a specific part by forcing shared parts of all samples close together and specific parts of different domains apart. A two-stage procedure pretrains the VAE and then fine-tunes its encoder while training any base forecaster on the augmented input $x' = \\mathrm{Concat}(z,x)W + b$.","core_discovery":"On the paper's own terms, the central discovery is that modeling temporal dependencies in a disentangled latent space—rather than only matching feature distributions across domains—is what lets a forecaster generalize to unseen domains. LTG's conditional β-VAE learns latent vectors whose shared part encodes domain-invariant patterns and whose specific part encodes per-domain behavior, and feeding both to a forecasting decoder improves range accuracy (quantile loss) in most settings. The paper argues that prior approaches neglect these latent temporal dependencies, and demonstrates the point by showing improved performance across web traffic, retail sales, stock volume, and power consumption data.","pith_inferences":["If the domain signal does persist in the latent vector despite being absent at test time, LTG is effectively learning a domain code; a direct comparison against simply concatenating the one-hot domain ID to the forecaster input would isolate how much of the gain is from the VAE machinery versus the ID itself.","The hyperparameters β and the split fraction α jointly control how much information the specific half can hold; an automatic selection of α per dataset could make LTG more robust, since the Web-traffic result suggests the optimal split varies.","Because latent factors are only used as extra input, LTG's gains are bounded by the base model's capacity to use them; the paper's own stated limitation opens the door to using the latent factors to condition the forecaster's parameters or training objective instead.","The method is univariate only; extending to multivariate series would require deciding whether to share latent factors across channels, which is a natural next test."],"forward_implications":["LTG can be bolted onto DeepAR, WaveNet, DLinear, and GPT4TS, and in most configurations it lowers quantile loss on unseen domains compared with the base model alone.","The domain-regularized latent space separates domains visually: t-SNE plots show overlapping shared parts and separated specific parts across test domains.","Two-stage training is necessary; training the whole pipeline end-to-end worsens results, so the latent factors must be established before the forecaster uses them.","Using only shared latent vectors can outperform the full model on some datasets, indicating that domain-specific factors are not always beneficial.","LTG's pretraining is cheaper than the gradient-matching baseline IDGM, and total runtime stays close to the base model."],"supporting_citations":[{"why":"Supplies the problem formulation, the common-patterns assumption, three of the five datasets, and the Cedar baseline that LTG must outperform.","marker":"[11]"},{"why":"Introduces the β-VAE whose disentanglement of latent factors LTG extends with domain conditioning.","marker":"[22]"},{"why":"Provides the variational autoencoder framework the conditional β-VAE builds on.","marker":"[31]"},{"why":"Defines DeepAR, the probabilistic base forecaster and the negative log-likelihood forecasting objective.","marker":"[44]"},{"why":"Defines WaveNet, the convolutional base forecaster used in the experiments.","marker":"[40]"},{"why":"Provides the gradient-matching domain-generalization baseline IDGM.","marker":"[45]"},{"why":"Defines DLinear, the linear base forecaster used to test LTG as a plug-in.","marker":"[56]"},{"why":"Defines GPT4TS, the LLM-based base forecaster used to test LTG's compatibility.","marker":"[62]"}],"fun_headline_variants":["Disentangled latent spaces boost domain-general forecasting","Plug-in conditional beta-VAE improves cross-domain forecasts","Mining latent dependencies for time series domain generalization","Domain-invariant forecast model via disentangled latent factors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest load-bearing premise is that conditioning the decoder on the domain ID during pretraining actually injects useful domain-specific information into the latent vector, even though the ID is absent at test time and only the encoder is used; if the domain signal lives entirely in the decoder, the forecasting gains would have to come from the shared latent part alone.","fun_headline_variants_meta":{"raw":{"variants":["Disentangled latent spaces boost domain-general forecasting","Plug-in conditional beta-VAE improves cross-domain forecasts","Mining latent dependencies for time series domain generalization","Domain-invariant forecast model via disentangled latent factors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000334,"raw_usage":{"total_tokens":1852,"prompt_tokens":941,"completion_tokens":911,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":850}},"tokens_in":557,"tokens_out":911,"duration_ms":7441,"temperature":1.0,"reasoning_tokens":850,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:13:34.228277+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LTG, then at test time replace every latent vector with a version from which the supposed domain-specific half has been removed (or shuffled across domains); if quantile loss on unseen domains does not change, the domain-specific mechanism carries no information and the claimed separation is not occurring. A companion check: permute the DomID fed to the decoder during pretraining and measure whether downstream forecasting accuracy drops.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the problem formulation, the common-patterns assumption, three of the five datasets, and the Cedar baseline that LTG must outperform."},{"cited_title":"Higgins, L","cited_arxiv_id":null,"evidence_quote":"Introduces the β-VAE whose disentanglement of latent factors LTG extends with domain conditioning."},{"cited_title":"Salinas, V","cited_arxiv_id":null,"evidence_quote":"Defines DeepAR, the probabilistic base forecaster and the negative log-likelihood forecasting objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines GPT4TS, the LLM-based base forecaster used to test LTG's compatibility."}],"review_version":1}