acceptodds
Under review as a conference paper at ICLR 2027

SPLICE: Latent Diffusion over JEPA Embeddings for Conformal Time-Series Inpainting

Abstract

Long gaps in time series are difficult to reconstruct reliably, and in power systems imputed load values inform dispatch and planning. Existing generative imputers are accurate but provide no finite-sample reliability guarantees under changing conditions. We introduce SPLICE (Self-supervised Predictive Latent Inpainting with Conformal Envelopes), which couples a JEPA encoder, a conditional latent bridge, an hourly-conditioned decoder, and Adaptive Conformal Inference (ACI). Controlled ablations indicate that long-horizon fidelity is driven more by the pre- dictive representation than by generative complexity: under a matched bridge and decoder, JEPA outperforms reconstructive VAE and contrastive TS2Vec encoders, and once the representation is fixed a 5-step flow sampler matches or improves on a 50-step DDIM sampler at a 5−10× speedup. The same ordering is recovered on twelve hydrological catchments. Across thirteen load datasets at 91-day gaps, SPLICE attains the lowest mean Load-only MSE (0.056), winning 9 of 12 non-degenerate datasets against seven established baselines, and the best mean CRPS (0.182, 7.8% below the strongest competitor), while ACI holds 93–95% empirical coverage where static calibration under-covers by up to 7.5 percentage points. A pooled JEPA encoder transfers to four held-out datasets with only bridge and decoder adaptation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised representation learning, time-series imputation, uncertainty quantification, conformal prediction, energy systems, joint-embedding predictive architectures (JEPA), flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_C931OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes SPLICE, a modular framework for long-horizon time series imputation. It combines a JEPA-based predictive encoder, a conditional latent bridge (supporting deterministic, noise-perturbed, or generative variants), an hourly-conditioned decoder, and adaptive conformal inference (ACI) for distribution-free uncertainty calibration. Experiments across multiple datasets demonstrate that SPLICE achieves lower errors.

Critical strengths
  1. The integration of JEPA as a predictive latent backbone for long-gap time-series inpainting appears novel to me.
  2. Overall, the experiments is solid, including those presented in the appendix.
  3. The code is provided in the Supplementary Material, which facilitates reproducibility.
Critical weaknesses
  1. The manuscript is difficult to read and understand from a human reader’s perspective due to its writing and organization. First, neither the Introduction nor the overall narrative of the paper clearly explains why JEPA is suitable for the problem at hand or what it can address that previous generative machine learning approaches cannot. The same concern applies to the other components of the proposed framework. Some extracted tables/figures in the text appear corrupted; while likely an artifact of PDF parsing, it occasionally interrupts readability. Is Figure 1 AI-generated? The font is so small that it is illegible. Even worse, the figure contains excessive whitespace around it, making the layout visually unappealing. Is it appropriate to devote so much space in the main text to describing the setup while relegating many important experiments to the appendix? This seems to be a misplaced prioritization. Furthermore, the experiments, particularly those in the appendix, lack sufficient discussion of key findings and analysis of observed phenomena. ... I do not believe the manuscript, in its current form, meets the acceptance threshold for ICLR, at least for its presentation.
  2. The proposed method also lacks novelty and reads more like a simple stacking of existing machine learning techniques than a genuinely new contribution. As a point of comparison, I am curious why the related work only covers studies published before 2025, as this makes it difficult to assess the method's distinctiveness relative to existing approaches.
  3. While the proprietary datasets are operationally meaningful, heavy reliance on them makes independent verification difficult; only four public datasets are included in the main table. Statistical tests indicate differences vs strong baselines (SAITS/BRITS) are not significant at p<0.05; more seeds or broader public benchmarks could strengthen claims of superiority.
  4. The imputation comparison omits some recent strong generative imputers, especially for algorithms after 2025.
  5. How αt\alpha_t maps into updated band quantiles from ensemble samples is not fully specified; more algorithmic clarity would help reproducibility and theoretical alignment. The coverage guarantee is marginal/long-run; no conditional or within-window guarantees are offered, which may limit operational interpretations for fixed gaps. The inference time claim (“6.5–9.0 ms for a 91-day gap”) appears implausibly low given sampling and decoding; this may be a units typo (likely seconds) and warrants correction.
Questions for authors

See Critical Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer C931 (1/2)

Thank you for your review, and for finding the use of JEPA as a predictive latent backbone for long-gap inpainting novel, the experiments solid, including those in the appendix, and the provided code helpful for reproducibility. We answer each point below; all changes to the PDF are listed in our general comment.

W1. Readability, why JEPA, and Figure 1

Why JEPA. Filling a 91-day gap means inferring unobserved days from the days around them, and JEPA is trained on this exact task. A predictor infers the embeddings of masked days from the observed context, with the loss computed directly in latent space (Section 4.1, Eq. 1). The latent therefore retains what is predictable from context: load regime, daily and weekly periodicity, and response to weather. Hourly detail is restored from covariates by the hourly-conditioned decoder (Section 4.4). Masked reconstruction (PatchTST) optimizes input fidelity, and contrastive learning (TS2Vec) optimizes invariance between views. Neither objective targets the prediction of missing segments.

We test this choice against both alternatives, with the same bridge architecture:

  • against a reconstructive β-VAE pipeline, JEPA lowers MSE on 9 of 9 datasets (Table 14);
  • against a contrastive TS2Vec latent, with bridge and decoder fixed, JEPA has a lower MSE on 9, 9 and 10 of 12 datasets at three seeds, and TS2Vec shifts the daily peak on 8 of 11 datasets (Appendix U);
  • the same ordering holds on twelve hydrological catchments (Section 6.5, Table 3).

Section 4.1 now opens with this rationale, and Section 8 states the resulting design principle.

JEPA also gives the pipeline the structure of a world model (Section 1). The bridge rolls out predicted latent states instead of sampling observations directly. This is where operational or physical constraints could be imposed, which we leave for future work.

Presentation. Figure 1 is now redrawn as a vector graphic at its printed size, which increases its text from about 3 pt on the page to 7–8 pt. The empty areas are removed, and the figure shows the 64-dimensional interface shared by the modules. The three-seed encoder swap has moved from the appendix into Section 6.1 (new Table 1).

W2. Novelty, and related work before 2025

We isolate, through controlled module swaps, which component drives reconstruction accuracy in long-gap inpainting, and build an operational pipeline on that finding. We swap each module individually while keeping the rest of the pipeline fixed:

  • Latent space: JEPA is compared with a β-VAE pipeline under the same bridge architecture (12.2% lower MSE, Table 14) and with a contrastive TS2Vec latent (lower on 9, 9 and 10 of 12 datasets at three seeds);
  • Sampler: with the latent space fixed, 5 flow-matching steps beat 50 DDIM steps on 12 of 13 datasets (8.7% lower MSE, Table 23);
  • Domain: both encoder and sampler orderings reproduce on twelve CAMELS-SPAT catchments (Table 3);
  • Transfer and reuse: a frozen pooled encoder transfers to four unseen domains, with only the bridge and decoder adapted (Section 7).

These results give the design rule of Section 8: learn the predictive state space first, then use the cheapest sampler that gives the required diversity. Our setting is contiguous 91-day outages, not randomly missing values. The pipeline provides online conformal intervals whose coverage holds under the deployable protocol: 0.949 on gaps never used before scoring (see our reply to 6FkT, W3). To our knowledge, no previous imputer combines a deep generative model with ACI (Table 10).

Recent work. Section 2 now covers work from 2025–2026:

  • V-JEPA 2 (Assran et al., 2025), which uses a JEPA latent as a world model for prediction and planning, the framing of our Section 1;
  • LeJEPA (Balestriero & LeCun, 2025) and VISReg (Wu et al., 2026), which replace the variance–covariance regularizer of Eq. 1 with distribution-matching terms and could be used directly in our encoder objective;
  • the imputation survey of Wang et al. (IJCAI 2025);
  • TimeMixer++ (ICLR 2025), which we also added as a baseline (W4).

TS-JEPA (Ennadir et al., 2025) was already discussed in the submitted Section 2.

AuthorsOpenReview ↗

Response to Reviewer C931 (2/2)

W3. Proprietary data, four public datasets, p > 0.05

Section 6.2 and Appendix P state that SPLICE is not significantly better than SAITS, BRITS or PatchTST. A one-sided Wilcoxon test over 12 datasets gives p = 0.102, 0.102 and 0.117, respectively. SPLICE is significantly better than ImputeFormer (p = 0.021) and than CSDI, TimesNet and TimeMixer++ (p < 0.001).

Public data. Four of the 13 datasets in Table 2 (Table 1 in the submitted version) are public: three UCI Electricity feeds and ETTh1. Section 6.5 also replicates the three controlled swaps on twelve catchments from the public CAMELS-SPAT dataset (Knoben et al., 2025). The target is hourly discharge, the covariates are ERA5 weather variables, and the pipeline is unchanged. The same orderings hold (Table 3):

  • JEPA beats TS2Vec on 9 of 12 catchments and the VAE latent on 10 of 12;
  • 5-step flow matching beats 50-step DDIM on 8 of 12.

Both encoder comparisons keep their ordering under hydrological criteria: six of ten criteria, including KGE and NSE, favor JEPA (Appendix F). The catchments were chosen before any training, on data completeness and hydrological diversity only (Table 7). CAMELS-SPAT is now listed alongside the other public datasets in the Data Availability appendix (Appendix H) and in the reproducibility statement.

W4. No imputers after 2025

We added TimeMixer++ (Wang et al., ICLR 2025), a multiscale and multiresolution architecture. It is trained through PyPOTS with the same splits, gap masks, three seeds and training protocol as the revised TimesNet column.

  • Mean result: its mean Load-only MSE over the twelve datasets is 0.298, against 0.056 for SPLICE.
  • Per dataset: SPLICE is lower on 11 of 12 datasets (one-sided Wilcoxon, p < 0.001).
  • Exception: RIInd1014, where TimeMixer++ reaches 0.149 against 0.202 for SPLICE; Seasonal Naïve is best there.

Overall win counts do not change: SPLICE is best on 9 of 12 datasets. TimeMixer++ is now in Table 2, the MAPE table (Table 15), Appendix P and the baseline figure (Figure 7).

W5. From α_t to the bands; marginal guarantee; 6.5–9.0 ms

At step t, the conformal quantile q̂_t is the empirical quantile at level ⌈(n+1)(1−α_t)⌉/n of the n pooled pointwise calibration scores. The band is [q_0.025 − q̂_t, q_0.975 + q̂_t], where q_0.025 and q_0.975 are the quantiles of the generated ensemble. Section 4.3 now states this. As noted in the limitations (Section 8), the guarantee is marginal and holds in the long run.

We now measure timing per 91-day gap using a script included in the supplementary material (RTX A4000, median over the nine datasets):

  • Deterministic pipeline: 2.8 ms.
  • Single trajectory: 32 ms for FM-A (5 steps), 0.36 s for DDIM (50 steps).
  • Ensembles for the intervals: 20 DDIM samples drawn in one batch take 1.1 s per gap, and the 50 calibration samples take 2.5 s. With FM-A, the same two figures are 0.11 s and 0.26 s.

Table 13 (Table 12 in the submitted version) now reports these values, listing single trajectory and ensemble separately.

We hope these answers address your concerns, and we are happy to clarify any remaining point during the discussion.

Thank you for the detailed response. I have updated my score to reflect the clarifications. However, I believe some problems, particularly regarding the writing, remain unresolved, and I therefore still lean negative.

AuthorsOpenReview ↗

Response to the follow-up

Thank you for considering our response and for updating your score. Could you please point us to the specific sections or passages that you still find difficult to follow? In the revised PDF, we have added the motivation for the predictive latent space in Section 4.1, redrawn Figure 1, and moved the encoder comparison to Section 6.1. We would be happy to address any specific passages you identify in a new revision of the PDF before the end of the discussion period, time permitting, and otherwise in the final version.

Reviewer_RfeJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper tries to reconstruct long stretches of missing data. The pipeline includes predictive JEPA embeddings, a conditional latent bridge, an hourly weather/calendar-conditioned decode, and adaptive conformal calibration. The main empirical result is that the proposed pipeline achieves the lowest mean load-only MSE on 9 of the 12 datasets when compared to seven existing baselines.

Critical strengths

The paper provides a practical, modular architecture with promising performance gains. The pipeline separates representation learning, latent prediction, decoding, and calibration in a straightforward manner. This offers potentially useful engineering contribution for long-horizon, covariate assisted data reconstruction.

Another strength of the paper is its broad investigation of representations and computational tradeoffs, which includes encode swaps, decoder studies, sampler comparisons, and transfer experiments. The thorough experimentation could yield helpful design guidance beyond the particular load application at the core focus of the paper.

Critical weaknesses

The reliability guarantees do not appear to be supported for the implemented method and deployment setting. Static and adaptive calibration are evaluated on different periods. Most reported ACI coverages remain below the stated 95% coverage floor. Hourly versus window-level update units are unresolved. As guaranteed reliability is a central stated contribution, these issues need to be resolved.

Experimental attribution and reporting inconsistencies also weaken the design principal. The legacy VAE uses 32 dimensions versus JEPA’s 64 dimensions. The appendix indicates that the reconstruction comparisons also use different encoders. Ts2Vec comparison are single-seed. Some of the stated improvement seem to conflict with the tables.

Questions for authors
  1. How is Proposition 1 valid with interior clamping? Please identify the exact implemented update.
  2. During a real 91-day outage, when are true missing targets available for ACI updates, and is the update unit an hour or a window?
  3. Do ACI’s improvements persist when static CQR and ACI are evaluated on identical timestamps with identical base predictions and information access?
  4. Which exact configurations support the principal accuracy, CRPS, interval-width, and sampler win-count claims?
3 comments
AuthorsOpenReview ↗

Response to Reviewer RfeJ (1/2)

Thank you for your detailed review, and for noting the practical, modular architecture and the broad investigation of representations and computational tradeoffs, which could give design guidance beyond load data. We answer each point below; all changes to the PDF are listed in our general comment.

Q1 "How is Proposition 1 valid with interior clamping? Please identify the exact implemented update."

The implemented update is α_{t+1} = α_t + γ(α − err_t) with γ = 0.01, followed by a clamp to [0.001, 0.999]. Here err_t denotes the fraction of miscovered points in window t, a rate in [0, 1] rather than a binary indicator. The clamp is never active in any run (α_t remains within [0.013, 0.059]), so the published sequence is exactly the unclamped recursion. Summing the recursion gives the identity

(1/T) Σ_t err_t − α = (α_1 − α_{T+1}) / (γT),

where T is the number of online windows (106–236; Appendix A now states this unit). This identity reproduces the observed shortfall on all nine datasets. For RIInd1014, for example, (0.05 − 0.0157)/(0.01 · 234) = 0.0147, which equals 0.0646 − 0.05. The worst-case bound (1 − 2ε)/(γT) evaluates to 0.42–0.94 at these horizons and is therefore loose here, as Appendix A now notes.

Q2 "During a real 91-day outage, when are true missing targets available for ACI updates, and is the update unit an hour or a window?"

The update is performed per window, as in the code above, and Eq. 8 now says so. During a real outage the labels arrive only once the gap ends, so a deployable update can only use completed gaps. Under this restriction, mean coverage is 0.937 (range 0.916–0.959) over all online windows. On gaps that do not overlap any calibration gap and whose labels were never used before scoring, coverage reaches 0.949 (revised Table 12, formerly Table 11; see also our reply to 6FkT, W3).

Q3 "Do ACI’s improvements persist when static CQR and ACI are evaluated on identical timestamps with identical base predictions and information access?"

In the submitted Table 11, static CQR was evaluated on the last window, whereas ACI was evaluated online over the remaining windows. The revised Table 12 compares all protocols on the same online windows with the same samples (mean coverage / mean width over the nine datasets):

all online windowsgaps never used before scoring
CQR, frozen α0.920 / 1.530.922 / 1.53
CQR recalibrated with completed gaps0.930 / 1.660.939 / 1.76
ACI updated with completed gaps0.937 / 1.850.949 / 2.05
ACI, published (immediate)0.942 / 1.760.948 / 1.81

Given the same information, ACI comes closest to the 0.95 target (0.949, against 0.939 for recalibrated CQR on never-used gaps), with wider intervals (2.05 vs 1.76).

Coverage below 95%

Under ACI, 8 of the 9 datasets fall below 0.95 in Table 12, by at most 0.016. The identity derived in Q1 accounts for the entire shortfall. The text no longer says "93–95%"; it gives the observed range, 93.4–95.4%, and Table 12 lists the per-dataset values.

AuthorsOpenReview ↗

Response to Reviewer RfeJ (2/2)

Q4 "Which exact configurations support the principal accuracy, CRPS, interval-width, and sampler win-count claims?"

Table numbers below refer to the revised PDF (submitted Tables 1 and 11 are now Tables 2 and 12). All four claims use the same JEPA encoder (d = 64) for a given dataset. They differ in the bridge, the sampling and the evaluation windows:

ClaimBridge → decoderSamplingSeeds / windowsMetric
Point accuracy (Table 2, 9 of 12)deterministic masked-transformer bridge → hourly decoderone deterministic passseeds 42, 43, 44; last 91-day gapLoad MSE
CRPS 0.182 (5 of 9)same seed-42 model as Table 220 members: decoder on bridge latents + N(0, 0.15²) noiseseed 42; same gapCRPS on Load, scaled to [0, 1] on the gap
Coverage / width (Table 12)diffusion bridge → base frame decoderDDIM, 50 steps, η = 0; 50 samples per calibration window, 20 per online window; ensemble quantiles 0.025 / 0.975first half of the windows calibrate, the rest are online; α = 0.05, γ = 0.01coverage, mean width
Sampler counts (Table 23)generative bridge → decoderDDIM 50 steps vs flow matching, Euler 5 steps (FM-A, FM-C)seed 42; 212–472 held-out windows per datasetall-feature gap-frame MSE

Table 23 reports three-way wins: FM-A is best on 10 of 13 datasets, FM-C on 3 and DDIM on none. Pairwise, FM-A beats DDIM on 12 of 13 datasets (all but UES_NH_Med), and FM-C beats DDIM on 8 of 13, with one tie (MT_001). Section 6.4 and Appendix S now report these pairwise counts.

Attribution: latent dimension, TS2Vec seeds, contradictions

  • Dimension. Retrained at d = 64 (seed 42, as in Table 14), the β-VAE pipeline still has a higher MSE than JEPA+Bridge on all 9 datasets. JEPA+Bridge is 6.8% lower on average, compared to 12.2% at d = 32. TS2Vec already uses d = 64, matching JEPA .
  • Encoder swap seeds. We re-ran both arms of the encoder swap (JEPA and TS2Vec, with identical bridge, decoder and evaluation) at seeds 43 and 44 on the twelve load datasets. JEPA has a lower gap-MSE on 9, 9 and 10 of the 12 datasets at seeds 42, 43 and 44. At the median, TS2Vec's gap-MSE is 4.4%, 19.0% and 22.2% higher. JEPA is lower at all three seeds on 7 datasets, whereas TS2Vec is lower at all three seeds on none. These results are in the new Table 1 (Section 6.1).
  • WCMA1011lig. On this dataset ACI widens the bands. On the same windows (revised Table 12), static CQR has a width of 2.616 and coverage of 0.950, while ACI has 2.911 and 0.954. We removed the sentences in Sections 6.3 and 8 of the submitted version that described a tightening.

We hope these answers address your concerns, and we are happy to clarify any remaining point during the discussion.

AuthorsOpenReview ↗

Follow-up on Author Response

Thank you again for your detailed review. As the discussion period ends on November 18, we wanted to ask whether our response has addressed your questions, particularly regarding the exact ACI update and the identity explaining the coverage shortfall in Q1, as well as the comparison between CQR and ACI over the same windows in Q3 and Table 12. We would be happy to clarify any remaining points before the deadline.

Reviewer_6FkTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces SPLICE (Self-supervised Predictive Latent Inpainting with Conformal Envelopes), which couples a JEPA encoder, a conditional latent bridge, an hourly-conditioned decoder, and Adaptive Conformal Inference (ACI).

Critical strengths
  1. The setting is more practical compared to random missing values

  2. overall framework design is reasonable and each module is relative indepenent to study effects clearly

Critical weaknesses
  1. there is no good independent testing set. early stopping setting looks not so stable.You should split strictly by time into train / validation / calibration / test, then re-run the main table (Table 1) and the coverage table (Table 11).

  2. baseline is not well trained. For example, MSE 6.66 for TimesNet, and this seems to it even did not coverage

  3. ACI coverage rate is a little cheating. In practice, how can I know the correct coverage rate. You should freeze α and measure coverage on gaps whose true values were never used for adjustment.

Questions for authors

I think the experiments are not so solid, you should make it really practical before publication.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6FkT

Thank you for your review, and for noting that the setting is more practical than randomly missing values and that the modules are independent enough to study the effect of each one clearly. We answer each point below; all changes to the PDF are listed in our general comment.

W1. Independent test set and early stopping

The gaps are excluded from all training losses. Their only link to training is the choice of the early-stopping epoch: the validation region used for that choice can overlap with the gap (Appendix B). Appendix W (Table 32, formerly Table 30) already reports the stricter setting you suggest. There, the validation window ends before the gap starts, so no test label is used at any stage. Compared with the standard protocol on the same runs, the median change in Load gap-MSE is 0.0%:

  • six of the 12 datasets are unchanged to the last reported digit;
  • the remaining datasets move in both directions (UES_NH_Med −12.0%, ETTh1 +2.5%, and WCMA1011lig +50.6% on a short series). No dataset switches from a win to a loss or the reverse. We address calibration (Table 12, formerly Table 11) in W3 below.

W2. Under-trained baselines (TimesNet 6.66)

We retrained TimesNet with early stopping on a validation window (the last 91 training days), with three seeds and its hyperparameters unchanged, and updated Table 2 (Table 1 in the submitted version), the MAPE and CRPS tables and the corresponding figures.

  • Performance: on MT_150, TimesNet now reaches a Load-only MSE of 2.01 ± 0.10 (SPLICE: 0.005), and its mean over the twelve datasets is 0.303 (SPLICE: 0.056). On the four public datasets, TimesNet stays behind SPLICE, and also behind BRITS on MT_320 and ETTh1.
  • Ranking and significance: win counts and significance do not change (SPLICE vs TimesNet: p < 0.001). SAITS (0.064), BRITS (0.068) and PatchTST (0.073) remain the closest competitors, against 0.056 for SPLICE (Table 2, three seeds).

W3. "ACI coverage rate is a little cheating"

The coverage does not depend on reusing labels. We ran the test you suggested on the same windows and samples as the submitted Table 11, using a single sampling pass shared by four protocols. The published protocol reproduces Table 11, with a maximum deviation of 0.0001 in coverage. Averaged over the nine datasets:

  • Frozen α after calibration: 0.920 (range 0.869–0.950). The adaptive update is what brings coverage to the target.
  • ACI updated only with gaps that ended before the current gap starts, which is the information available during a real outage: 0.937 over all online windows. On gaps that do not overlap any calibration gap and whose labels were not used before scoring (18–149 per dataset), coverage is 0.949 for a 0.95 target, and 8 of 9 datasets are at 0.93 or above. Delayed updates produce wider intervals (2.05 vs 1.81 on the same gaps, normalized units), as expected when the level is adjusted using older information. The revised Table 12 (Section 6.3; protocol in Appendix B) reports both protocols.

Q. "Make it really practical"

The revision adds the evidence required for practical deployment, using only information available during a real outage:

  • Calibration without future labels. Updating ACI only with completed gaps gives 0.949 coverage (target 0.95) on gaps whose labels were never used before scoring (W3, Table 12).
  • Model selection without test labels. Placing the validation window before the gap gives a 0.0% median change in Load gap-MSE (W1, Table 32).
  • Runtime and training cost. On a single RTX A4000, imputing a 91-day gap takes 2.8 ms with the deterministic pipeline. A 20-sample interval ensemble takes 106 ms with 5-step flow matching and 1.08 s with the DDIM sampler used in Table 12. Training takes 8–16 min per dataset (Table 13).
  • Evaluation setting. Evaluated on contiguous outages in nine operational utility feeds and four public datasets, with gap lengths of 7, 30 and 91 days (Appendix O). Errors in physical units are in Appendix D.
  • Reuse. A pooled encoder serves unseen datasets, with only the bridge and decoder adapted (Section 7; few-shot curves over seeds in Appendix Y).

Experimental rigor. Table 2 reports three seeds with standard deviations. TimesNet is retrained with early stopping (W2), TimeMixer++ (Wang et al., ICLR 2025) is added as a baseline, and the encoder swap is repeated at three seeds (new Table 1).

Deployment status. The pipeline has also been packaged for internal use at an electric utility. The package covers training, an automatic acceptance test against a reference, a model registry, and gap filling from operational exports. Details that would identify the operator are withheld to preserve anonymity.

We hope these answers address your concerns, and we are happy to clarify any remaining point during the discussion.

AuthorsOpenReview ↗

Follow-up on Author Response

Thank you again for your review. As the discussion period ends on November 18, we wanted to ask whether our response has addressed your concerns, particularly regarding (i) the coverage achieved when ACI is updated using only completed gaps (0.949 on gaps not previously used for calibration, as reported in Table 12) and (ii) the results obtained with the retrained TimesNet model. We would be happy to clarify any remaining points before the deadline.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment

We thank the reviewers for their careful reading. We have uploaded a revised PDF with changes in blue. All changes are summarized below. The new Table 1 shifts the numbering of the later tables: for example, submitted Tables 1, 11, 12 and 30 are now Tables 2, 12, 13 and 32. Our replies use the revised numbers.

  • Conformal protocol (6FkT W3, RfeJ Q2–Q3). Table 12 (Section 6.3) now compares four calibration protocols on the same online windows and samples:

    • CQR with frozen α;
    • the published ACI;
    • ACI updated only with completed gaps;
    • CQR recalibrated with the same information.

    We also report coverage on gaps whose labels were never used before scoring (protocol in Appendix B).

  • Proposition 1 and the ACI update (RfeJ Q1–Q2, C931 W5). We revised the text to match the implementation. The update uses the miscoverage rate of each window, and T counts windows (Eq. 8, Appendix A).

  • Coverage wording (RfeJ). We replaced the "93–95%" range with the observed range, 93.4–95.4%; Table 12 lists the per-dataset values.

  • Baselines (6FkT W2, C931 W4).

    • Retrained TimesNet with early stopping on a validation window (the last 91 training days), and updated Table 2, the MAPE and CRPS tables (Tables 15–16) and the figures.
    • Added TimeMixer++ (ICLR 2025) to Table 2, the MAPE table, Appendix P and the baseline figure (Figure 7).
    • All conclusions on wins and significance remain unchanged.
  • Table 2. Evaluated all methods with three seeds on the four public datasets, as the caption states, and updated Appendix P. All conclusions on wins, ranking and significance remain unchanged.

  • Presentation and related work (C931 W1–W2).

    • Added the motivation for the predictive latent space at the beginning of Section 4.1.
    • Redrew Figure 1 as a vector figure at its printed size.
    • Moved the three-seed encoder swap from the appendix to Section 6.1 (new Table 1).
    • Expanded Section 2 to discuss V-JEPA 2, LeJEPA, VISReg and a recent imputation survey.
  • Bands and timing (C931 W5). Defined the band and the conformal quantile in Section 4.3, and separated single-trajectory and 20-sample ensemble runtimes in the timing table (Table 13).

  • Other corrections.

    • Updated Table 14 and Figure 6 to report averages over all evaluation windows.
    • Removed the sentences in Sections 6.3 and 8 that described a tightening on WCMA1011lig.
    • Corrected the win counts of DDIM against flow matching in Section 6.4 and Appendix S (12 and 8 of 13).
    • Added CAMELS-SPAT to the Data Availability appendix (Appendix H) and the reproducibility statement.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.