{"id":"17c7669a-df32-4617-a841-5e8036d65258","arxiv_id":"2505.23017","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Combining a learned Koopman linearization with a learned Kalman filter inside a VAE produces a probabilistic forecaster that beats existing methods on most tested short- and long-horizon datasets.","lead":"K2VAE is a new forecasting model that turns messy historical data into a simple linear pattern, then uses a learned Kalman filter to sharpen future predictions and their uncertainty. It reports better accuracy and speed than current methods on long-horizon energy, traffic, and weather data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Full per-horizon tables do not support the blanket 'outperforms SOTA' claim: K2VAE loses to CSDI on Weather-L at L=96/192 and to GRU-MAF/Trans-MAF/CSDI on Exchange-L at L=96, so the headline rests on the L=720 subset.","rationale":"I read the paper as an empirical systems paper: the load-bearing condition for the abstract is that the reported experiments justify 'outperforms state-of-the-art methods.' That condition is weakest in the full per-horizon tables, not in the L=720 summary. The paper deserves credit for releasing code and data, reporting ablations, and providing efficiency numbers, and a conditional acceptance is consistent with those strengths. The concern I raise does not overturn the verdict; it sharpens the condition: the acceptance should require either an aggregate significance analysis or a qualified claim. I partially agree with the reader because the self-observation Kalman update is also a real weakness in the mechanism story, but I would not place it ahead of the empirical support issue, since the empirical claim could survive even if the Kalman interpretation is nonstandard. No misconduct is implied; the issue is presentation and verification of the headline claim.","tokens_in":36604,"tokens_out":8723,"duration_ms":79696,"concrete_test":"Aggregate Tables 9 and 10 across all four long-term horizons per dataset, compute mean CRPS and NMAE ranks for K2VAE vs each baseline, and run a paired Wilcoxon signed-rank test over the 9 datasets; do the same for the short-term settings in Table 2. Also move the L=96/192/336 columns into the main text instead of the appendix. If K2VAE is not significantly better than CSDI on the long-term aggregate, or if the conclusion depends on excluding Weather-L and Exchange-L losses, the abstract's 'outperforms SOTA' claim must be weakened or qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: K2VAE 'outperforms state-of-the-art methods in both short- and long-term PTSF.' The main long-term table (Table 3) reports only L=720, where K2VAE is best on most datasets. The complete per-horizon results in Table 9 contradict a uniform reading. On Weather-L, K2VAE CRPS is 0.080±0.007 at L=96 and 0.079±0.009 at L=192, while CSDI achieves 0.068±0.008 and 0.068±0.006; at L=336 the two are tied within error (0.082±0.010 vs 0.083±0.002). On Exchange-L at L=96, K2VAE (0.031±0.002) is worse than GRU MAF (0.026±0.010), Trans MAF (0.028±0.002), and CSDI (0.028±0.003). Short-term Table 2 also shows CSDI with lower CRPS on Electricity-S (0.051±0.001 vs 0.053±0.002). The abstract's claim is therefore only defensible as an average-over-benchmark claim, but no average, rank, or significance test is reported. Highlighting only L=720 makes the headline non-falsifiable as stated. The KalmanNet self-observation issue identified by the reader is a legitimate mechanism-level concern, but the first-order question is whether the complete empirical record supports the empirical claim, and that is where the support is thinnest.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"K^2VAE is a VAE-based probabilistic forecaster that patches multivariate series, projects tokens into a learned Koopman measurement space, fits a linear Koopman operator (local eDMD plus a global learnable part), and then runs a learnable Kalman filter/update to produce refined states and covariance matrices that serve as a variational posterior. A decoder maps samples to an isotropic Gaussian predictive distribution. The paper claims state-of-the-art performance in both short- and long-term PTSF and reports efficiency gains over diffusion/flow baselines, with experiments on ProbTS benchmarks, 5-seed runs, ablations, and qualitative interval plots.","tokens_in":37038,"tokens_out":5139,"duration_ms":50951,"significance":"The empirical contribution is potentially useful: a one-step generative forecaster that is lightweight and shows strong results on several datasets is a practically meaningful addition to LPTSF. Strengths include the external ProbTS benchmark, 5-seed error bars, full ablation tables, efficiency measurements, and a public code/data repository. However, the two load-bearing supports for the claims—the convergence theorem and the blanket 'outperforms SOTA' statement—are not established by the presented evidence. The theoretical result relies on an unproved assumption and a special-case construction, and the full per-horizon tables contain clear counterexamples to the unqualified empirical claim. These issues are correctable in revision, but they are substantive enough to require reworking the paper's claims and proof.","major_comments":[{"comment":"The statement that K^2VAE 'outperforms state-of-the-art methods in both short- and long-term PTSF' is not supported by the complete per-horizon results. In Table 9, Weather-L at L=96 gives K^2VAE CRPS 0.080±0.007 against CSDI 0.068±0.008, and at L=192 0.079±0.009 against 0.068±0.006. On Exchange-L at L=96, K^2VAE (0.031±0.002) is worse than GRU MAF (0.026±0.010), Trans MAF (0.028±0.002), and CSDI (0.028±0.003). Even in the short-term Table 2, CSDI has lower CRPS on Electricity-S (0.051±0.001 vs 0.053±0.002). No average rank, aggregated CRPS, or significance test is reported, so the main-table emphasis on L=720 does not establish the blanket claim. Please qualify the claim or add an appropriate aggregate comparison with uncertainty.","section":"Abstract, §4.2, Table 9/10"},{"comment":"Theorem 3.2 (and the corresponding Theorem A.2 in the appendix) is not proven. The proof begins by assuming U→0 without any argument that the training objective drives the Integrator residual to zero; Section 3.1.2 asserts this, and the assertion is not analyzed. More importantly, the 'general case' merely states that A can be treated as a fine-tuned Koopman operator, and convergence is only demonstrated for the special parameter selection H→I, A→0, R→0, under which Eq. (20) becomes z_k = x_hat^H_k by construction. This does not establish that the KalmanNet state transition converges to the Koopman operator for the learned matrices used in practice. Please either provide a genuine proof under stated assumptions or explicitly relabel this as a limiting-case remark rather than a theorem.","section":"Appendix A.2, Theorem 3.2"},{"comment":"The observation model in Eq. (16) defines o_k = H z_k + v_k, but the update in Eq. (20) uses x_hat^H_k, the KoopmanNet's own forecast, as the observation. This is not an observation of the true future state, so the standard Kalman-filter justification for state correction does not apply; the update is a learned alignment between two model outputs. Since the paper attributes the reduction of error accumulation to this Kalman mechanism, the mechanism needs a formal justification or a controlled experiment (e.g., replacing the update with a generic learned correction, or ablating the Kalman structure while keeping parameters) to show what the Kalman form actually contributes.","section":"§3.1.2, Eqs. (16)–(20)"}],"minor_comments":[{"comment":"There are many typographical errors ('excell', 'consitute', 'roubustness', 'weightsumed', 'foreacasting', 'sytem', 'distengle', 'pionner') that should be corrected during copyediting.","section":"Throughout"},{"comment":"The metric is inconsistently written as 'CPRS' instead of 'CRPS'.","section":"Appendix C.3"},{"comment":"The code repository for TimeGrad is listed as the PatchTST repository, which appears to be an incorrect URL; please verify all baseline links.","section":"Table 8"},{"comment":"Several long-term baseline cells are marked '-' (e.g., CSDI on Electricity-L and Traffic-L in Table 9); please state explicitly why these runs are missing and whether the omitted baselines change the comparisons that support the conclusion.","section":"Table 9"}],"recommendation":"major_revision","confidential_remarks":"This is a technically solid empirical paper with a serious overclaim in the abstract. The full appendix tables show the model is not uniformly better than CSDI across short- and long-term horizons. I would encourage the editors to require the authors to either restrict the claim or present a statistically grounded aggregate evaluation. The theoretical section should be downgraded to a remark unless a valid proof is provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Patrick—\n\nThis is a serious engineering contribution. The architecture—Koopman linearization via eDMD plus a learnable global correction, a learnable Kalman filter inside a VAE, all as a one-step generative forecaster—is new and well-suited for long-horizon probabilistic forecasting. The experimental effort is substantial: five seeds per run, ablations for every module, efficiency measurements, and released code and data. On the most demanding long-horizon setting (L=720), K2VAE beats all baselines on most datasets, and its memory/compute footprint is genuinely attractive (0.094GB vs 1.411GB for CSDI on Electricity-96).\n\nThat said, the paper overclaims. The abstract says K2VAE “outperforms state-of-the-art methods in both short- and long-term PTSF,” but the full per-horizon tables tell a more nuanced story. On Weather-L, CSDI has lower CRPS at L=96 (0.068 vs 0.080) and L=192 (0.068 vs 0.079); on Exchange-L at L=96, GRU-MAF, Trans-MAF, and CSDI all beat K2VAE. The main text only highlights L=720, where K2VAE wins most often, and no average ranks or significance tests are reported. So the blanket claim isn’t falsifiable as stated.\n\nThe theoretical analysis is the weakest part. Theorem 3.2 essentially assumes what it needs to prove: it sets U→0 and then cherry-picks H→I, A→0, R→0 so the Kalman update reduces to the Koopman prediction. That shows consistency, not convergence. Theorem 3.1 is a standard Joseph-stabilized update and is fine. The Kalman mechanism itself is also conceptually off: Eq. 20 treats the model’s own prediction as the observation, so the filter is aligning one model output with another rather than correcting with real data. It may still work as a learned refinement, but the “reducing error accumulation” story is not the classical Kalman one. The paper also misses the KalmanNet literature, which would sharpen the novelty claim about the Kalman component.\n\nOverall, this is a paper worth engaging with. The architecture is novel, the empirical study is broad, and the reproducibility is good. The flaws are fixable: tone down the headline claim, add per-horizon significance or rank-based comparisons, rewrite Theorem 3.2 honestly, and clarify what the KalmanNet is actually doing. That’s exactly the kind of revision a serious referee should request, not a reason to desk-reject. I’d send it to review.","headline":"Solid engineering paper with a novel Koopman-Kalman VAE for probabilistic forecasting; the empirical claim is overstated and the convergence theorem is circular, but the architecture and experimental breadth deserve proper review.","tokens_in":37547,"tokens_out":4402,"would_cite":true,"duration_ms":42562,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A variational autoencoder that linearizes its latent dynamics with a Koopman operator and filters them with a Kalman recursion claims state-of-the-art probabilistic forecasting on both short and long horizons at lower cost.","keywords":["probabilistic time series forecasting","long-term forecasting","variational autoencoder","Koopman operator","Kalman filter","error accumulation","uncertainty quantification","generative models"],"falsifier":"Retrain K²VAE on a long-horizon benchmark with the update step $z_k = \\hat{z}_k + K_k(\\hat{x}^H_k - H\\hat{z}_k)$ replaced by $z_k = \\hat{z}_k$ (or with the previous refined state as the observation) while keeping the control input, the KoopmanNet, and all training settings identical; if long-horizon CRPS is essentially unchanged, the Kalman observation mechanism is not what produces the reported gains. A second check is to measure the empirical coverage of the 95% prediction intervals on held-out long-horizon test sets, since the paper visualizes intervals but never scores their calibration.","tokens_in":36417,"feed_emoji":"📈","tokens_out":14414,"duration_ms":127083,"temperature":0.7,"pith_summary":"The paper claims that long-term probabilistic time series forecasting fails because the nonlinear, non-stationary dynamics of real series make generative models both inaccurate and expensive as the prediction horizon grows. To fix this, K²VAE maps patched series tokens through a learned measurement function into a space where their evolution is approximately a linear dynamical system governed by a Koopman operator, then runs a learned Kalman filter over that system so that each forecast step refines the state and carries a covariance forward as the uncertainty of the prediction. The design gives the variational autoencoder's latent space a concrete meaning: the posterior mean is the filtered trajectory and the posterior covariance is the process uncertainty of the linear system. The authors report that this one-step generative model beats eleven baselines on most of the tested short- and long-term datasets, with the largest gains on long horizons, while using far less inference time and GPU memory than diffusion- or flow-based models. If the claim holds, probabilistic forecasting can stay accurate and cheap out to horizons of hundreds of steps instead of collapsing as the horizon grows.","feed_headline":"Koopman-Kalman VAE tops long-term forecasters with one sampling step","feed_subtitle":"Claims sharper probabilistic scores on 96–720 step horizons while using less memory than diffusion baselines","key_machinery":"The load-bearing object is the pairing of a Koopman operator with a Kalman recursion inside a variational autoencoder, so that the latent space becomes the state space of a filtered linear system. The KoopmanNet provides the linearization: a learnable measurement function $\\psi$ maps patch-embedded tokens into measurement space, and the transition operator $\\mathbf{K} = K_{\\mathrm{loc}} + K_{\\mathrm{glo}}$ is assembled from the one-step eDMD formula $K_{\\mathrm{loc}} = X_{\\mathrm{fore}}(X_{\\mathrm{back}})^\\dagger$ plus a learnable global correction, so iterating $\\mathbf{K}$ on the first token produces the whole context reconstruction and horizon forecast in one pass. The KalmanNet performs the refinement: its process model predicts $\\hat{z}_k = A z_{k-1} + B u_k$ and $\\hat{P}_k = A P_{k-1} A^\\top + Q$, its observation model produces a Kalman gain $K_k = \\hat{P}_k H^\\top (H \\hat{P}_k H^\\top + R)^{-1}$, and the update $z_k = \\hat{z}_k + K_k(\\hat{x}^H_k - H\\hat{z}_k)$ blends the linear forecast with the integrated nonlinear residual; the covariance update is rewritten in a numerically stable dual form to preserve positive definiteness. The refined states and covariances directly define the variational posterior, giving the VAE latent semantics, and a proven convergence statement says the KalmanNet's transition reduces to the Koopman operator as the residual control input vanishes.","core_discovery":"On the paper's own terms, the discovery is that probabilistic forecasting can be recast as Bayesian filtering over a linearized measurement space, and that this removes both the accuracy loss and the cost blow-up that appear when the forecast horizon extends. The KoopmanNet learns a measurement function $\\psi$ that projects patch-embedded tokens into a space where consecutive tokens advance under a Koopman operator $\\mathbf{K} = K_{\\mathrm{loc}} + K_{\\mathrm{glo}}$, with $K_{\\mathrm{loc}}$ fit in closed form by one-step extended dynamic mode decomposition and $K_{\\mathrm{glo}}$ learned; iterating $\\mathbf{K}$ from the first token generates the context reconstruction and the horizon forecast. Because this learned linearization is imperfect, the authors call the system 'biased' and pass it through the KalmanNet, which integrates the nonlinear residual as a control input and treats the Koopman forecast as the observation. Iterated predict-and-update steps then yield refined states $z_k$ and covariance matrices $P_k$, which define the variational posterior $Q(Z|X) = \\mathcal{N}(Z', P)$; the decoder acts as the inverse measurement function and outputs a Gaussian target distribution. The authors report that the resulting model surpasses all compared baselines in most settings, including a 7.3% CRPS reduction over the best short-term baseline and a 20.9% reduction over the best long-term baseline, and their ablations show that both the KoopmanNet and the KalmanNet are needed for the gains.","pith_inferences":["An editorial extension: the same predict–update recursion could run online after deployment, assimilating newly arriving observations into the state without retraining, a streaming-forecasting use the paper does not discuss.","The paper displays 95% confidence intervals but does not score their calibration; an editorially suggested check is measuring empirical coverage on the long-horizon test sets, which would separate genuine uncertainty quantification from a well-shaped predictive distribution.","The filtered-state representation is portable: because every token carries a mean and a covariance, the same encoder could be reused for imputation, anomaly detection, or scenario generation, not only for forecasting.","A cheap diagnostic suggested by the architecture is freezing the Kalman gain at zero while keeping the residual control input, to test how much of the long-horizon gain is filtering versus residual integration."],"forward_implications":["One-step generation replaces iterative sampling: because the whole horizon is drawn from a single Gaussian posterior, inference time and memory scale with the model size rather than with the number of denoising or flow steps, which is what the efficiency experiments measure.","Long horizons become tractable for probabilistic use: the reported advantage over point-forecasting models equipped with distributional heads grows as the horizon extends to 96–720 steps, the regimes used in energy, traffic, and public-health planning.","The latent space gains semantics: with the posterior mean as the filtered state trajectory and the posterior covariance as process uncertainty, the model outputs per-token prediction intervals as a by-product of the filtering recursion.","The ablations indicate that linearization and filtering are complements: removing either the KoopmanNet or the KalmanNet degrades scores substantially, and the mixed operator $K_{\\mathrm{loc}} + K_{\\mathrm{glo}}$ is what keeps long-horizon training numerically stable."],"supporting_citations":[{"why":"Supplies the founding theory that nonlinear dynamics can be represented by a linear operator on measurement functions.","marker":"Koopman (1931)"},{"why":"Establishes the Koopman-operator linearization in the large that the KoopmanNet is built to approximate.","marker":"Lan & Mezić (2013)"},{"why":"Provides the dynamic mode decomposition whose one-step variant yields the closed-form local operator $K_{\\mathrm{loc}}$.","marker":"Schmid (2010)"},{"why":"Supplies the Kalman predict–update recursion that the KalmanNet mimics and stabilizes.","marker":"Welch (1995)"},{"why":"Defines the variational autoencoder objective (ELBO) that K²VAE optimizes with its filtered latent space.","marker":"Kingma & Welling (2014)"},{"why":"Koopa, the Koopman-based point forecaster that K²VAE extends from deterministic to probabilistic forecasting and compares against as a baseline.","marker":"Liu et al. (2023)"},{"why":"PatchTST, the patching source and the point-forecasting baseline that the paper reports beating on long horizons.","marker":"Nie et al. (2023)"},{"why":"CSDI, the diffusion baseline that defines the paper's headline short-term comparison.","marker":"Tashiro et al. (2021)"}],"fun_headline_variants":["Koopman-Kalman VAE cuts long-term CRPS by 20.9%","Outperforms diffusion with one sampling step: Koopman-Kalman VAE","Less memory, one step: Koopman-Kalman VAE beats diffusion on long horizons","Koopman-Kalman VAE: linearize then filter for accurate long-term forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that feeding the KoopmanNet's own forecast back into the Kalman update as though it were a measurement of the true future actually corrects the state and curbs error accumulation; in a standard Kalman filter such correction is justified only when the observation comes from the real system, so the claimed mechanism treats one model output as evidence about the world.","fun_headline_variants_meta":{"raw":{"variants":["Koopman-Kalman VAE cuts long-term CRPS by 20.9%","Outperforms diffusion with one sampling step: Koopman-Kalman VAE","Less memory, one step: Koopman-Kalman VAE beats diffusion on long horizons","Koopman-Kalman VAE: linearize then filter for accurate long-term forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001491,"raw_usage":{"total_tokens":6035,"prompt_tokens":1041,"completion_tokens":4994,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":4898}},"tokens_in":657,"tokens_out":4994,"duration_ms":34394,"temperature":1.0,"reasoning_tokens":4898,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:55:53.697663+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain K²VAE on a long-horizon benchmark with the update step $z_k = \\hat{z}_k + K_k(\\hat{x}^H_k - H\\hat{z}_k)$ replaced by $z_k = \\hat{z}_k$ (or with the previous refined state as the observation) while keeping the control input, the KoopmanNet, and all training settings identical; if long-horizon CRPS is essentially unchanged, the Kalman observation mechanism is not what produces the reported gains. A second check is to measure the empirical coverage of the 95% prediction intervals on held-out long-horizon test sets, since the paper visualizes intervals but never scores their calibration.","supporting_citations":[],"review_version":1}