acceptodds
Under review as a conference paper at ICLR 2027

Parallax: Correcting Time Series Forecasts from Cross-Resolution Disagreement

Abstract

In long-term time series forecasting, a fine-resolution forecast, when aggregated, can disagree with a forecast produced directly at a coarser resolution for the same future interval. This disagreement is available at prediction time and carries information about fine-resolution forecast errors. We present Parallax, a post-hoc, model-agnostic method that uses cross-resolution disagreement to correct the fine-resolution forecasts of frozen backbones, whether pretrained foundation models or task-specific deep forecasters trained on the target dataset. Parallax decomposes the disagreement into the aggregated fine residual and the coarse-forecast error, reducing forecast correction to residual estimation, and fits an affine corrector whose moments combine a resolution-structured covariance, built from a nested aggregation chain, with empirical moments estimated from historical forecast–target pairs. At test time, Parallax uses only the fine- and coarse-resolution forecasts, requiring neither future observations nor any updating to the backbone. Evaluated on 13 benchmark datasets with 7 task-specific models and 5 pretrained foundation models, Parallax reduces MSE by 11.8%-26.4% across task-specific backbones and by 33.2% on average across foundation models.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, Cross-resolution disagreement, Post-hoc forecast correction, Time series foundation models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_b1izOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes Parallax, a method for correcting time series predictions. It starts from the observation that averaged fine predictions can differ from coarse predictions. The method uses this difference to correct the fine predictions. It fits a linear correction using past predictions and their true values. It mixes error statistics based on time scales with those calculated directly from the same past records. The time scales and mixing weight are selected using validation data. The results show improvements across different models, including foundation models and models that already use multiple time scales.

Critical strengths
  • The method improves different types of forecasting models, including strong baselines, foundation models, and models that already use multiple time scales. I believe showing that a simple correction can help these different models is a practical contribution of this work.
  • Ablations show that it helps to consider how reliable the coarse predictions are. Using more details from past errors also brings further gains. These results support the design and help explain why each part is included.
Critical weaknesses
  • Parallax improves predictions through a separate correction step, which requires additional fitting and inference time. Its improvement therefore comes with a time cost. If a foundation model could learn during pretraining to make averaged fine predictions agree with coarse predictions, this extra correction step might be less necessary.
  • Parallax keeps the correction fixed after fitting. In a streaming environment, error patterns may change, making the fixed correction less useful. Updating it requires additional fitting time, creating a trade-off between keeping the correction useful and its update cost. Although Appendix E acknowledges this risk, the paper does not evaluate how often updates are needed or whether their gains justify the extra cost.
Questions for authors

The weaknesses above are my main concerns. Addressing them would have the greatest impact on my assessment. I also have the following questions:

  • How much past data do the fine and coarse models use? If the coarse model sees a longer period of data, can you separate the benefit of this extra information from the benefit of Parallax? Does Parallax still help when the fine model sees the same period of past data?
  • Do the reported running times include preparing the inputs, making all fine and coarse predictions, and applying the correction? How long does it take to update the correction module? This would help assess whether the method can be used in streaming settings and how often it can be updated.
  • How much past data is needed to fit a reliable correction? Do the results change much when less or more past data is used?
  • Foundation models can be used across datasets, but Parallax still fits a correction for each dataset. Appendix D.5 shows that several mixing weights work well with LightGTS. Does this also hold for the other foundation models? Such a comparison would help clarify how much dataset-specific tuning is needed and could be useful for developing more general post-hoc methods.
5 comments
AuthorsOpenReview ↗

Response to the Reviewer b1iz (W1 Part 1)

We sincerely thank you for the constructive comments, which have helped us to improve our paper. We provide new experiments for each point, and all results have been added to the revised paper (in blue-colored text).

W1 (Part 1): Parallax improves predictions through a separate correction step, which requires additional fitting and inference time. Its improvement therefore comes with a time cost.

We agree that the Parallax correction adds time, and we have now measured it for each stage, compared with δ-Adapter (ICLR 2026) and RGMR (ICML 2026) (see Table R1 below, full results in Table 18, Appendix D.6, Page 43). The additional time is small and in return Parallax achieves the largest MSE reductions among the 3 correctors. For each window, Parallax takes 1.01 to 3.07 times as long as the backbone alone, while reducing MSE by up to 64.8%.

Table R1. Cost and accuracy of the post-hoc correctors at H=96, with LightGTS, Toto-2.0-4m, and Toto-2.0-22m on ETTm1, Weather, and Electricity. Time per window covers: coarse-input preparation + all fine and coarse forecasts + the correction, relative to the backbone alone. Fitting time includes the forecasts for fitting and chain selection, and storage excludes the backbone. Negative ΔMSE means improvement over backbone.

CorrectorTime per window ratio between Corrector and Backbone aloneFitting time, forecasts includedStorage (MB)ΔMSE
δ-Adapter (ICLR 2026)1.00 to 1.0493.1 s to 2,681.6 s, GPU training3.81 to 127.40−0.6% to −13.4%
RGMR (ICML 2026)56 to 1,9428 h to 10 days0.60 to 19.75−11.9% to +14.5%
Parallax1.01 to 3.0754.1 s to 1,803.1 s, no gradients0.23 to 10.97−32.7% to −64.8%

For task-specific backbones, the coarse models add a one-time training cost, which we report in our answer to the Reviewer C45Z, Q3. We summarize the results, including training, inference, and storage costs of Parallax for 3 foundation models and for OLinear, in Table R2 below.

Table R2: Costs (additional training time, storage, time per window, and correction time) added by Parallax over ETTm1, Weather, and Electricity at H=96. Per-window time covers coarse-input preparation, all fine and coarse forecasts, and the correction time. Models are trained on GPU, and the corrector is fitted on CPU. Foundation models have no additional training stage for coarse levels. For OLinear, the additional training time is relative to training the backbone alone. Negative ΔMSE means improvement over backbone alone.

BackboneAdditional models to trainAdditional training timeAdditional storage (MB): coarse models + correctorTime per window (ms): backbone → ParallaxCorrection time (ms)ΔMSE
LightGTSnonenone0 + 0.23 to 10.979.70 to 12.12 → 10.45 to 21.580.19 to 0.20−32.7% to −64.0%
Toto-2.0-4mnonenone0 + 0.25 to 10.4814.18 to 14.53 → 14.39 to 21.360.18 to 0.20−37.8% to −64.8%
Toto-2.0-22mnonenone0 + 0.24 to 9.4920.20 to 22.23 → 20.46 to 68.230.19 to 0.22−37.1% to −63.9%
OLinearone per coarse resolution+26% to +112%, once64.36 to 74.82 + 0.25 to 10.971.42 to 1.93 → 8.23 to 10.150.18 to 0.20−6.9% to −11.2%

As seen in Tables R1 and R2 above (see full results in Table 18, Appendix D.6, Page 43), the correction itself is almost free (about 0.2 ms against 8.23 to 68.23 ms for the whole window in Table R2), so nearly all of the time is spent on producing the forecasts. Parallax builds the cross-resolution disagreement from these forecasts and uses it to reduce MSE by up to 64.8% on the 3 datasets.

AuthorsOpenReview ↗

Response to the Reviewer b1iz (W1 Part 2, W2)

W1 (Part 2): If a foundation model could learn during pretraining to make averaged fine predictions agree with coarse predictions, this extra correction step might be less necessary.

We agree that such pretraining could reduce the disagreement, but agreement does not guarantee a smaller error. By Eq. (1), the disagreement $ \mathbf{d}_k = (\hat{\mathbf{y}}^{(k)} - \mathbf{A}_k\mathbf{y}) - (\mathbf{A}_k\hat{\mathbf{y}}^{(1)} - \mathbf{A}_k\mathbf{y})$, so dk=0\mathbf{d}_k=\mathbf{0} can mean that the two views make the same error on every block mean, not that this error is small (Section 3.1). For example, if the coarse forecast and the averaged fine forecast both predict 10 for a block whose true mean is 12, they agree but both are off by 2. Our results also show that although reconciliation methods enforce this agreement on the same forecasts, they increase MSE on up to 13 of the 20 dataset-backbone pairs, while Parallax improves all 20 (see Table 15, Appendix D.3.2, Page 36-38). Parallax also improves backbones that process several resolutions internally by up to 26.4% and recent foundation models by up to 37.6% in average MSE (see Section 4.1, Tables 1 and 11, Page 6 and 30). Thus, the correction remains useful for models that already learn across scales.

W2: Parallax keeps the correction fixed after fitting. In a streaming environment, error patterns may change, making the fixed correction less useful. Updating it requires additional fitting time, creating a trade-off between keeping the correction useful and its update cost. Although Appendix E acknowledges this risk, the paper does not evaluate how often updates are needed or whether their gains justify the extra cost.

We evaluated both how often updates are needed and whether they justify their cost on LightGTS (H=96, 7 datasets). Every H, 4H or 16H test windows, we refit the Parallax corrector on the most recent data, using the same number of windows as the training set, reusing their forecasts with the chain and λ fixed. We compare it with δ-Adapter (ICLR 2026), which is updated by one gradient step per window. As shown in Table R3 below, Parallax fitted once already beats this online δ-Adapter, and refitting every 16H windows costs at most 86.8 ms per test window.

Table R3: Updating Parallax and δ-Adapter on LightGTS (H=96). "Fitted once" uses the Parallax corrector fitted on the training set for the whole test set. With "refit every kH" (k=1, 4, or 16), we fit a new corrector after every kH test windows on the most recent data using the same number of windows as the training set and use it for the next kH test windows. Negative ΔMSE means improvement over LightGTS alone. Update time is the CPU time of one Parallax refit divided by the number of test windows between two refits.The sampling interval of each dataset is given in parentheses.

DatasetΔMSE, δ-Adapter onlineΔMSE, Parallax fitted onceΔMSE, Parallax refit every HΔMSE, Parallax refit every 4HΔMSE, Parallax refit every 16HParallax update time per test window (ms): refit every H / 4H / 16H
ETTh1 (1 h)−0.47%−13.41%−13.79%−13.76%−13.32%8.5 / 2.1 / 0.5
ETTh2 (1 h)−0.29%−4.92%−4.28%−4.04%−4.52%8.7 / 2.2 / 0.5
ETTm1 (15 min)−0.97%−64.02%−64.19%−64.21%−64.13%46.7 / 11.7 / 2.9
ETTm2 (15 min)−0.89%−26.19%−25.82%−25.86%−25.91%47.8 / 11.9 / 3.0
Weather (10 min)−8.84%−32.75%−34.01%−33.99%−33.99%135.3 / 33.8 / 8.5
Electricity (1 h)−13.99%−49.88%−50.13%−50.05%−49.92%1,389.5 / 347.4 / 86.8
Solar-Energy (10 min)−11.52%−84.15%−84.70%−84.73%−84.69%640.5 / 160.1 / 40.0

On the studied benchmarks, a Parallax correction fitted once stays useful over time. In the standard setting, the corrector is already kept fixed over the whole test set, as in the "fitted once" column of Table R3. To check for a decline within this period, we take 3 settings of Table R3 (Parallax fitted once, Parallax refit every H and the online δ-Adapter), split their test windows into 4 consecutive quarters in time order, and on each quarter compare the MSE of each setting with that of LightGTS alone on the same windows (see Figure 10, Appendix D.6, Page 44). Parallax fitted once beats the online δ-Adapter in every quarter and still reduces MSE by 4.2% to 78.5% in the last one. Refitting every H windows changes this by at most 2.1 points, so what Parallax learns stays valid over time.

AuthorsOpenReview ↗

Response to the Reviewer b1iz (Q1, Q2)

Q1: How much past data do the fine and coarse models use? If the coarse model sees a longer period of data, can you separate the benefit of this extra information from the benefit of Parallax? Does Parallax still help when the fine model sees the same period of past data?

In short, yes, the benefit of the longer history can be separated from that of Parallax, and Parallax still helps even when the coarse forecasts see less history than the fine model.

  • “How much past data do the fine and coarse models use?”

The fine model sees L = 96 time steps (36 on ILI), as in the standard benchmarks. With foundation models, the same frozen model forecasts each coarse resolution k from 96 block means, covering 96k original steps. With task-specific models, an independent coarse model is trained for each k, and its lookback is chosen on validation from 96 and max(4, ⌊96/k⌋) block means. We have clarified this protocol in Appendices C.1 and C.4.

  • “If the coarse model sees a longer period of data, can you separate the benefit of this extra information from the benefit of Parallax? Does Parallax still help when the fine model sees the same period of past data?”

Yes to both. We rerun Parallax on LightGTS with the same frozen model at every resolution and limit the history of the coarse forecasts to the 96 steps of the fine model (same history) or to the last 48 of them (half history). Parallax still reduces MSE on all 7 datasets and keeps 84% of its gain with the same history and 82% with half of it (Table R4 below).

Table R4: Parallax on LightGTS with the history of the coarse forecasts controlled. ΔMSE shows improvement compared to LightGTS only, averaged over 4 horizons.

DatasetΔMSE, standard setting (coarse sees 96k steps)ΔMSE, same history (all resolutions ≤ 96 steps)ΔMSE, half history (coarse sees 48 steps)
ETTh1−10.82%−9.87%−9.41%
ETTh2−6.94%−2.51%−1.91%
ETTm1−54.66%−49.74%−49.19%
ETTm2−19.39%−11.59%−10.60%
Weather−25.79%−14.55%−14.32%
Solar-Energy−79.83%−73.19%−72.62%
Electricity−49.46%−44.99%−43.75%
Average−35.27%−29.49%−28.83%
Improvement retained84%82%

The gain remains because, even when the coarse forecasts see the same or a shorter history than the fine forecast, the frozen model makes different errors at the coarse resolution. So, the disagreement between the coarse forecast and the aggregated fine forecast carries information about the fine-resolution forecast errors, and Parallax uses this information to correct the fine forecast (Section 3.1).

Q2 (Part 1): Do the reported running times include preparing the inputs, making all fine and coarse predictions, and applying the correction?

Please refer to W1 (Part 1) above.

Q2 (Part 2): How long does it take to update the correction module? This would help assess whether the method can be used in streaming settings and how often it can be updated.

Updating the corrector costs 8.5 ms to 1,389.5 ms per test window with LightGTS on the 7 datasets when refitting every H windows (Table R3 above), since it only re-estimates closed-form moments from existing forecasts, with no forward pass or gradient training. This is shorter than the interval between two observations (10 minutes to 1 hour), so even refitting every H windows is affordable in streaming. Such frequent refits are rarely needed on these benchmarks. Refitting changes the MSE reduction little on every dataset, and refitting every 16H windows, at 0.5 ms to 86.8 ms per test window, stays close to refitting every H windows. Deciding when to refit under stronger shifts, particularly in streaming settings, is beyond the paper's focus and remains open. We revised Appendix E to clarify these limitations and scope of the paper.

Reviewer_C45ZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Parallax, a model-agnostic method for improving time-series forecasts using disagreement between predictions made at different temporal resolutions. The key idea is that the difference between a coarse-resolution forecast and the aggregated fine-resolution forecast contains information about the fine-resolution prediction error. Parallax estimates the fine-resolution residual from cross-resolution disagreement using an affine corrector. To estimate the required moments, it constructs a structured covariance based on orthogonal temporal bands and resolution-specific error variances, and combines it with empirical moments estimated from historical forecast–target pairs. Experiments with 7 task-specific forecasting models and 5 pretrained foundation models report consistent MSE improvements.

Critical strengths

Consistent empirical improvements across diverse forecasting models and datasets.

Parallax achieves substantial and consistent performance improvements across a broad range of forecasting tasks. The results suggest that cross-resolution information can provide useful complementary signals beyond those captured by individual forecasting models. The improvements across different model architectures, including models that already incorporate multiscale processing, strengthen the practical significance of the approach.

A simple and interpretable statistical framework for forecast correction.

Parallax formulates forecast correction as residual estimation from cross-resolution disagreement, providing a clear statistical interpretation of the correction process. The resulting approach is potentially useful when multiple resolution-specific forecasts are available.

Critical weaknesses

Insufficient coverage of recent and competitive forecasting models.

7 task-specific backbones and 5 foundation models are evaluated, but 5 of the backbones were introduced in 2023-2024. More recent works of task-specific backbones should be included. Some representative FMs should also be included, such as TimesFM 2.5 and Moirai 2.0.

Insufficient experimental evidence to isolate and substantiate the claimed contributions.

The ablation study shows that structured moments alone achieve nearly the same performance as the full Parallax, leaving the practical contributions of empirical moments and moment mixing insufficiently established.

Current experiments do not adequately disentangle improvements attributable to cross-resolution disagreement and the proposed covariance estimator from those arising from supervised calibration, additional forecasting models, or model ensembling. Fine-forecast-only calibration and appropriately regularized forecast fusion baselines are needed to establish these distinctions. In particular, although Appendix D.3 includes several fusion baselines, the learned variants includes high amounts of parameters without sufficient regularization. For example, Linear fusion has H(H+m+1) trainable parameters and uses no weight decay. Parallax uses a similar amount of parameters, but the ablation shows the structured component dominates the improvements and the parameters of it are only 2J+1. The design itself provides a structured regularization. If low-dimensional statistical parameterization is intended to be a central contribution, it should be explicitly articulated and supported by corresponding comparisons.

Despite the emphasis on model-agnostic and post-hoc correction, applying Parallax to task-specific backbones requires independently training additional forecasters at each coarse resolution, rather than simply reusing an existing fine-resolution model. The additional training, inference, and model-storage costs should be explicitly accounted for when assessing the practical advantages of the method.

Questions for authors
  • Experiments on more backbones and FMs, as mentioned in Critical Weaknesses.
  • More experiments of fine-grained-only calibration and appropriately regularized forecast fusion baselines, as mentioned in Critical Weaknesses. If intended as contributions, provide experiments and explanation on low-dimensional statistical parameterization.
  • Informantion about additional training and inference costs introduced by Parallax.
5 comments
AuthorsOpenReview ↗

Response to the Reviewer C45Z (Q1)

We sincerely thank you for the constructive comments, which have helped us to improve our paper. We provide new experiments for each point, and all results have been added to the revised paper (in blue-colored text).

Q1: Experiments on more backbones and FMs.

We thank the reviewer for the suggestion. The original evaluation already included recent models, OLinear (2025), SiGMA (2026) and all five foundation models (2025 and 2026), alongside widely used baselines from 2023 and 2024. Following the suggestion, we added three recent task-specific backbones from different architecture families, FACT [1] (ICLR 2026, convolutional), ReNF [2] (ICML 2026, MLP) and SimpleTM [3] (ICLR 2025, attention-based), and two foundation models, TimesFM 2.5 and Moirai 2.0. The results are summarized in Table R1 below (full results in Tables 10 and 11, Appendix D.1.1, Pages 28-30).

Table R1: Parallax performance on five recent backbones. The number of improvement cases and Avg. ΔMSE (changes relative to the backbone alone) are reported. Negative ΔMSE means improvement compared to backbone only. For TimesFM 2.5 and Moirai 2.0, we excluded some datasets included in their pretraining lists, resulting in fewer cases compared to others.

TypeBackboneDatasetsParallax improvement (cases)Avg. ΔMSE
Task-specificFACT1352/52−10.53%
Task-specificReNF1352/52−15.39%
Task-specificSimpleTM1352/52−10.55%
Foundation modelTimesFM 2.5728/28−35.79%
Foundation modelMoirai 2.0312/12−37.64%

Parallax improves every added model, including strong recent backbones. It reduces MSE in 196 over studied 196 dataset-horizon cases by 10.53% to 15.39% on average for the three task-specific backbones and by 35.79% and 37.64% for the two foundation models (see Tables 10 and 11, Appendix D.1.1, Pages 28-30). This supports the model-agnostic contribution of Parallax. With these additions of the most recent task-specific and foundation models, Parallax still consistently improves all 17 backbones, lowering MSE in 695 over 696 studied backbone-dataset-horizon cases while acting only on their forecasts.

References:

[1] Huiqiang Wang, Jieming Shi, and Li Qing. FACT: Fine-grained across-variable convolution for multivariate time series forecasting. ICLR, 2026.

[2] Yihang Lu, Xianwei Meng, and Enhong Chen. ReNF: Rethinking the design of neural long-term time series forecasters. ICML, 2026.

[3] Hui Chen, Viet Luong, Lopamudra Mukherjee, and Vikas Singh. SimpleTM: A simple baseline for multivariate time series forecasting. ICLR, 2025.

AuthorsOpenReview ↗

Response to the Reviewer C45Z (Q2)

Q2: More experiments of fine-grained-only calibration and appropriately regularized forecast fusion baselines, as mentioned in Critical Weaknesses. If intended as contributions, provide experiments and explanation on low-dimensional statistical parameterization.

  • More experiments of fine-grained-only calibration and appropriately regularized forecast fusion baselines.

Thank you for these detailed suggestions. We added the requested baselines, each covering one possible source of gain: calibration of the fine forecast, additional coarse models, ensembling, and regularized fusion. We summarize the results in Table R2 below.

Table R2: Ablation studies on other calibration and fusion schemes. ΔMSE shows changes in MSE with respect to the backbone, including OLinear and LightGTS, over seven datasets (20 dataset-backbone pairs, 80 dataset-horizon cases) with the same evaluation setting. Negative ΔMSE means improvement. Each dataset is averaged over 4 horizons. Bold indicates the best value.

GroupMethodparameters per channel (given H=96)Avg. ΔMSE, OLinearAvg. ΔMSE, LightGTSDataset-backbone pairs with lower MSEDataset-horizon cases with lower MSE
CalibrationBias96+0.66%−1.13%10/2037/80
CalibrationRidge (fine)9.3k−1.09%−22.78%14/2062/80
Additional modelsCoarse only0−1.66%−14.70%13/2048/80
EnsemblingEqual average0−0.66%+1.67%15/2057/80
EnsemblingWeighted average5 to 6−4.26%+19.12%14/2056/80
Regularized fusionRidge (fine+coarse)15k to 18k−6.99%−28.50%13/2062/80
Regularized fusionLinear fusion, tuned weight decay15k to 18k−5.07%−30.06%13/2062/80
Regularized fusionMLP fusion, tuned4.1k to 205k−6.65%−32.00%15/2065/80
OursParallax3 to 9k−13.92%−35.27%20/2080/80

Parallax consistently improves MSE on all dataset-backbone pairs and dataset-horizon cases, while calibration, ensemble, and regularized fusion schemes only improve the backbone forecasting at most 15/20 and 65/80 cases, respectively. Our Parallax also outperforms other schemes in Avg. ΔMSE on both OLinear and LightGTS backbones.

  • Appropriate regularization.

We perform regularization on every learned fusion and hyperparameter tuning to select the best model for each fusion scheme. Linear fusion now tunes its weight decay and learning rate. Ridge fits the same linear model with a penalty searching on 49 values. MLP fusion tunes weight decay, width, depth, dropout, and learning rate searching on 567 combinations (see Table 13, Appendix D.3.1, Page 33). The best fusion model only improves MSE on 65/80 cases, while Parallax improves all 80/80 cases.

  • Low-dimensional statistical parameterization.

Low-dimensional parameterization is one of our central contributions (Secs 1 and 3.2). Parallax constructs its covariance from coarse/fine forecasts with one residual variance per temporal band and one error variance per coarse resolution with at most 2⌊log₂H⌋ + 1 parameters in total. With an additional weight λ to mix with empirical moments, Parallax fits only 3 to 9k parameters per channel, against 4.1k to 205k for the fusions (see Table 14, Appendix D.3.2, Page 33-35). By this scheme, Parallax exploits the relationship among different levels to fix the fine forecast while avoiding the overfitting, thus improving the forecasting performance.

AuthorsOpenReview ↗

Response to the Reviewer C45Z (Q3, W1, W2, W3)

Q3: Informantion about additional training and inference costs introduced by Parallax

We agree that for task-specific backbones, the coarse models are part of the cost of Parallax. The reported cost on LightGTS (see Table 18, Appendix D.6, Page 43) did not include coarse-resolution running time since its roles are both fine and coarse forecasts. We have measured training, inference, and storage costs of Parallax for three foundation models and for OLinear. We summarize the result in Table R3 below.

Table R3: Costs (additional training time, storage, time per window, and correction time) added by Parallax over ETTm1, Weather, and Electricity at H=96. Per-window time covers coarse-input preparation, all fine and coarse forecasts, and the correction time. Models are trained on GPU, and the corrector is fitted on CPU. Foundation models have no additional training stage for coarse levels. The range of MSE change is reported. Negative MSE change means improvement.

BackboneAdditional models to trainAdditional training timeAdditional storage (MB): coarse models + correctorTime per window (ms): backbone → ParallaxCorrection time (ms)MSE change
LightGTSnonenone0 + 0.23 to 10.979.70 to 12.12 → 10.45 to 21.580.19 to 0.20−32.7% to −64.0%
Toto-2.0-4mnonenone0 + 0.25 to 10.4814.18 to 14.53 → 14.39 to 21.360.18 to 0.20−37.8% to −64.8%
Toto-2.0-22mnonenone0 + 0.24 to 9.4920.20 to 22.23 → 20.46 to 68.230.19 to 0.22−37.1% to −63.9%
OLinearone per coarse resolution+26% to +112%, once64.36 to 74.82 + 0.25 to 10.971.42 to 1.93 → 8.23 to 10.150.18 to 0.20−6.9% to −11.2%

With these costs included, Parallax still keeps its practical advantages. For the foundation models, it needs no additional training model and stores only the corrector because the frozen backbone can produce both the fine and coarse forecasts. For task-specific backbones such as OLinear, the coarse models are the main extra cost, but they are trained once before deployment. Parallax only requires the CPU to fit its corrector, without gradients.

W1. Insufficient coverage of recent and competitive forecasting models. 7 task-specific backbones and 5 foundation models are evaluated, but 5 of the backbones were introduced in 2023-2024. More recent works of task-specific backbones should be included. Some representative FMs should also be included, such as TimesFM 2.5 and Moirai 2.0.

We provided the experiments on the most recent backbones in Q1.

W2. Current experiments do not adequately disentangle improvements attributable to cross-resolution disagreement and the proposed … If low-dimensional statistical parameterization is intended to be a central contribution, it should be explicitly articulated and supported by corresponding comparisons.

Please refer to our answer in Q2 and the additional ablation below.

Additional ablation study. We have expanded the ablation with LightGTS and added the variant without structured moments (λ = 1) in Table R4 below (full results in Tables 16-17, Appendix D.4, Pages 39-40).

Table R4: Ablation study on Parallax components with OLinear and LightGTS backbones on four datasets each. MSE averaged over 4 horizons.

Variant (average MSE)OLinear: ETTh1 / ETTm2 / Weather / PEMS04LightGTS: ETTh1 / ETTm2 / Weather / Solar-Energy
Backbone without Parallax0.4268 / 0.2707 / 0.2385 / 0.09080.486 / 0.315 / 0.295 / 1.060
Without empirical moments (λ = 0)0.4068 / 0.2483 / 0.2168 / 0.07180.466 / 0.272 / 0.245 / 0.402
Without structured moments (λ = 1)1.1575 / 0.3346 / 1.3107 / 0.07510.597 / 0.292 / 0.271 / 0.223
Parallax (λ selected on validation)0.4058 / 0.2474 / 0.2144 / 0.07160.434 / 0.254 / 0.219 / 0.214

These results show that empirical moments bring a large part of the gain with LightGTS. Removing them raises MSE on all 4 datasets (0.214 → 0.402 MSE on Solar-Energy). Mixing empirical and structured moments is needed because the empirical moments alone can give a higher MSE than the backbone. With λ = 1 (without structured moments), MSE exceeds the backbone’s on ETTh1 for LightGTS and on 3 of the 4 datasets for OLinear. Mixing both kinds of moments, with λ selected on validation, gives the lowest MSE in every case of Table R4.

W3. Despite the emphasis on model-agnostic and post-hoc correction, applying Parallax to task-specific backbones requires independently training additional forecasters at each coarse resolution, rather than simply reusing an existing fine-resolution model. The additional training, inference, and model-storage costs should be explicitly accounted for when assessing the practical advantages of the method.

Please refer to our answer in Q3.

Reviewer_6fqkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Parallax, a post-hoc method that corrects fine-resolution forecasts using disagreements with coarse-resolution forecasts. It estimates forecast residuals through an affine correction combining structured covariance estimates with empirical moments. Experiments across 13 datasets, seven task-specific models, and five foundation models demonstrate substantial MSE improvements.

Critical strengths
  1. A simple, interpretable, model-agnostic correction requiring no backbone updates.

  2. Extensive empirical evaluation showing consistent improvements across diverse datasets and forecasting architectures, supported by ablations.

Critical weaknesses
  1. Comparisons against classical temporal forecast reconciliation methods (e.g., MinT) and regularized linear residual correction are missing, making it difficult to isolate the benefit of the proposed covariance structure.

  2. The validation protocol needs clarification, particularly for short datasets such as ILI. Furthermore, experiments use a single fixed seed, limiting evidence of robustness.

Questions for authors
  1. How does Parallax compare against covariance-shrinkage temporal reconciliation or ridge-regularized residual correction using the same forecasts?

  2. How are validation windows constructed for ILI at horizon 60, given that each validation half appears shorter than the forecasting horizon?

  3. Does the reported computational overhead include generating coarse-resolution forecasts and training additional coarse models?

4 comments
AuthorsOpenReview ↗

Response to the Reviewer 6fqk (Q1, Q2)

We sincerely thank you for the constructive comments, which have helped us improve the paper. We ran new experiments for each point, and all results have been added to the revised paper (in blue-colored text).

Q1: How does Parallax compare against covariance-shrinkage temporal reconciliation or ridge-regularized residual correction using the same forecasts?

We compared Parallax with OLS [1], WLS [1], MinT-shrink reconciliation [2], and ridge-regularized residual correction (Ridge). All methods are applied to the same frozen forecasts of OLinear and LightGTS, with the same forecasts. The results on 13 datasets with OLinear and 7 datasets with LightGTS (20 dataset-backbone settings and 80 dataset-horizon settings in total) are summarized in Table R1 below (full results in Table 15, Appendix D.3.2, Page 36-38).

Table R1: Parallax vs OLS, WLS, MinT, and Ridge. Avg. ΔMSE represents changes in MSE relative to the backbone (OLinear and LightGTS) only, averaged over 4 horizons and all datasets (negative = better). Last column: number of dataset–horizon cases where each method achieves the lowest MSE (ties included). Best methods are in bold.

MethodAvg. ΔMSE, OLinearAvg. ΔMSE, LightGTSDataset–backbone pairs with worse MSE (over 20 cases)Dataset–horizon cases with the best MSE (over 80 cases)
OLS−7.76%−16.03%412
WLS−10.56%−12.50%010
MinT-0.71%+6.41%130
Ridge (fine+coarse)-6.99%−28.50%718
Parallax−13.92%−35.27%055
  • Reconciliation methods: Reconciliation aims at a different goal, adjusting every resolution so that the fine forecasts aggregate exactly to the coarse ones. It enforces coherence but may not reliably correct the fine forecast. In contrast, Parallax keeps the coarse forecasts unchanged and treats the cross-resolution disagreement as a noisy measurement of the fine-forecast residual, which it estimates via structured and empirical moments to correct the fine forecast (Secs. 3.1-3.2). In Table R1, MinT decreases performance on 13 of 20 dataset-backbone pairs, and WLS gains much less than Parallax on average MSE (12.50% vs 35.27% with LightGTS).

  • Ridge-regularized residual correction method: First, Parallax exploits a structured covariance from the nested aggregation chain, which captures the relationship among different levels. On the other hand, Ridge regularizes all of its coefficients with a single penalty, neglecting the connection among coarse/fine forecasts. Second, Parallax has fewer parameters compared to Ridge, which helps our method avoid overfitting. Ridge treats the correction as a generic regression from the fine/coarse forecasts to the residual, which requires around 15k coefficients (see Table 14, Appendix D.3.2, Page 35). In contrast, Parallax constructs its covariance from how the disagreement arises in Eq. (1) with temporal bands and one error variance per resolution. Given the same setting with Ridge, Parallax’s structured moments need at most 2⌊log₂H⌋ + 1 parameters. In Table R1, with its design, Parallax improves all 20 dataset-backbone pairs, while Ridge increases MSE on 7 of 20 and gives a smaller average MSE reduction with both backbones.

References:

[1] George Athanasopoulos, Rob J Hyndman, Nikolaos Kourentzes, and Fotios Petropoulos. Forecasting with temporal hierarchies. European Journal of Operational Research, 2017.

[2] Shanika L Wickramasuriya, George Athanasopoulos, and Rob J Hyndman. Optimal forecast recon- ciliation for hierarchical and grouped time series through trace minimization. Journal of the American Statistical Association, 2019.

Q2: How are validation windows constructed for ILI at horizon 60, given that each validation half appears shorter than the forecasting horizon?

Thank you for carefully checking this. The two validation halves split the validation windows, not the validation time steps, so no validation windows are cut short. The ILI validation set has 97 steps, so half of it would be shorter than H=60. But following the standard benchmark loader, the first validation windows take their 36-step lookback from the end of the training set. So, all validation windows are drawn from 133 consecutive steps. With a one-step stride, this gives 98 - H windows, each with the full lookback and the full H-step target: we get 74/62/50/38 validation windows for H=24/36/48/60, i.e., 37/31/25/19 windows per half.

We have revised the experiment setting description in Appendix C.4 (Fitting the Parallax corrector) to make this more explicit.

AuthorsOpenReview ↗

Response to the Reviewer 6fqk (Q3)

Q3: Does the reported computational overhead include generating coarse-resolution forecasts and training additional coarse models?

The original submission did not include these two costs, and we added both in the revised version. The initial efficiency analysis timed only the corrector fit for LightGTS on ETTm1. We have re-measured every stage for three post-hoc correctors, including Parallax, δ-Adapter (ICLR 2026), and RGMR (ICML 2026). We study 3 foundation models and OLinear on ETTm1, Weather, and Electricity (7, 21, 321 channels, respectively; H=96). With every stage counted, Parallax still gives the largest MSE reduction of the 3 correctors on the foundation models, with a correction step that is almost free at inference, strengthening its practical contribution. We summarize these findings in Table R2 below (full results in Table 18 and Figure 9, Appendix D.6, Page 42-43).

Table R2: Computational overheads. Ranges span ETTm1, Weather and Electricity (7-321 channels) at H = 96, plus LightGTS and Toto-2.0-4m/22m for δ-Adapter and RGMR. Per-window time covers coarse-input preparation, all fine/coarse forecasts, and the correction (0.18-0.22 ms for Parallax). OLinear’s coarse models add 64.36–74.82 MB. Negative ΔMSE means improvement compared to backbone only.

Method (backbone)Coarse models to trainFitting time, forecasts included (s)Time per window (ms)Corrector storage (MB)ΔMSE
Parallax (LightGTS)none54.1 to 730.510.45 to 21.580.23 to 10.97−32.7% to −64.0%
Parallax (Toto-2.0-4m)none64.9 to 919.214.39 to 21.360.25 to 10.48−37.8% to −64.8%
Parallax (Toto-2.0-22m)none128.7 to 1,803.120.46 to 68.230.24 to 9.49−37.1% to −63.9%
Parallax (OLinear)+26% to +112% training time, once16.4 to 171.88.23 to 10.150.25 to 10.97−6.9% to −11.2%
δ-Adapter (ICLR 2026), 3 foundation modelsnone93.1 to 2,681.6, GPU training9.94 to 22.993.81 to 127.40−0.6% to −13.4%
RGMR (ICML 2026), 3 foundation modelsnone8 h to 10 days560 to 41,5460.60 to 19.75−11.9% to +14.5%

Parallax has some advantages in computation. First, the Parallax corrector is almost free. On the foundation models, the correction itself takes about 0.2 ms, and a full test window takes 10.45-68.23 ms, against 560-41,546 ms with RGMR. Fitting the corrector only requires a CPU without gradient updates for its parameters. Moreover, it can be used to boost the performance of foundation models without training coarse models. Second, the fitted Parallax corrector takes the least storage of the 3 correctors. It needs 0.23-10.97 MB, compared with 0.60-19.75 MB for RGMR and 3.81-127.40 MB for δ-Adapter.

AuthorsOpenReview ↗

Response to the Reviewer 6fqk (W1, W2)

W1: Comparisons against classical temporal forecast reconciliation methods (e.g., MinT) and regularized linear residual correction are missing…

Please refer to our answer in Q1.

W2: The validation protocol needs clarification, particularly for short datasets such as ILI. Furthermore, experiments use a single fixed seed, limiting evidence of robustness.

  • The validation protocol needs clarification, particularly for short datasets such as ILI.

Please refer to our answer in Q2.

  • Furthermore, experiments use a single fixed seed, limiting evidence of robustness.

Given the backbone's forecasts and the historical forecast-target pairs, Parallax is deterministic, so a random seed can only affect the results through the backbone. We ran 5 seeds with OLinear, one of the strongest task-specific backbones, and with Sundial, the only foundation model in this work that is non-deterministic at inference time as it samples forecast trajectories through its flow-matching head. We also added the recent post-hoc corrector δ-Adapter (ICLR 2026) for comparisons. Results are in Table R3 below (full results in Table 20, Appendix D.7, Page 45-46).

Table R3: Averaged MSE over 4 horizons with 5 random seeds (Mean ± std (ΔMSE)). ΔMSE is the change relative to Backbone only, and negative values indicate improvement.

BackboneDatasetBackbone onlyδ-AdapterParallax
OLinearETTh10.4254 ± 0.00110.4259 ± 0.0010 (+0.12%)0.4062 ± 0.0008 (−4.51%)
OLinearETTh20.3697 ± 0.00210.3700 ± 0.0020 (+0.11%)0.3497 ± 0.0009 (−5.39%)
OLinearETTm10.3746 ± 0.00030.3739 ± 0.0003 (−0.19%)0.3343 ± 0.0004 (−10.76%)
OLinearETTm20.2704 ± 0.00020.2700 ± 0.0002 (−0.15%)0.2464 ± 0.0008 (−8.90%)
OLinearWeather0.2393 ± 0.00140.2356 ± 0.0020 (−1.57%)0.2142 ± 0.0010 (−10.48%)
SundialETTh10.4755 ± 0.00010.4753 ± 0.0001 (−0.04%)0.4196 ± 0.0001 (−11.76%)
SundialETTh20.4014 ± 0.00020.4016 ± 0.0002 (+0.05%)0.3638 ± 0.0001 (−9.38%)
SundialETTm10.5432 ± 0.00020.5407 ± 0.0002 (−0.45%)0.3460 ± 0.0001 (−36.30%)
SundialETTm20.3075 ± 0.00010.3068 ± 0.0001 (−0.22%)0.2625 ± 0.0001 (−14.65%)
SundialWeather0.2846 ± 0.00010.2730 ± 0.0005 (−4.06%)0.2144 ± 0.0001 (−24.65%)

The gain of Parallax is stable across seeds. Parallax lowers MSE on all 10 dataset-backbone pairs, by 4.51% to 36.30%, and the standard deviation of its gain is at most 0.83 percentage points. Its standard deviation is at most 0.0010, comparable to that of the backbone (up to 0.0021), so the seed variation comes mainly from the backbone (see Table 20, Appendix D.7, Page 45-46). δ-Adapter changes MSE by less than 0.5% on 8 of 10 pairs and increases it on 3, with the largest reduction of 4.06%. With the same seeds, Parallax reduces MSE by at least 4.51% on every pair.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer 6fqk,

Thank you again for your detailed and constructive review. We have posted our rebuttal and would hope to follow up on your key suggestions.

In our response, we compared Parallax with temporal reconciliation methods and ridge-regularized residual correction on the same forecasts (W1, Q1). We clarified how the validation windows are constructed for short datasets such as ILI, and revised Appendix C.4 accordingly (W2 Part 1, Q2). Moreover, we added experiments with five random seeds on OLinear and Sundial, with δ-Adapter for comparison (W2 Part 2). Also, we provided a detailed analysis of costs, including coarse forecasts and coarse models, against δ-Adapter and RGMR correctors (Q3).

We are hoping that these new experiments and clarifications help to address your concerns. Thank you again for your precious time and guidance.

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers and the AC

We thank all reviewers and the AC for their careful reading and constructive feedback. In addition to the individual response to each reviewer, we have uploaded a revised paper with all changes marked in blue.

Paper Strengths. All reviewers (6fqk, C45Z, and b1iz) acknowledge these following strengths. Parallax is a simple, interpretable, and model-agnostic correction that needs no backbone update. By exploiting the disagreement between coarse and fine forecasts, it gives consistent and substantial improvements across diverse datasets, foundation models, and task-specific models, which is a practical contribution, as Reviewers C45Z and b1iz also say. The paper has a clear presentation of the correction process and its design. Reviewers 6fqk and b1iz also note that the ablations support this design.

New experiments in this rebuttal. Following the reviewers' suggestions, we ran the experiments and added the clarifications below. The details and result tables are in the individual responses.

ConcernWhat we addedReviewers
Comparison with temporal reconciliation and ridge correctionOLS, WLS, MinT-shrink and Ridge (fine+coarse) on the same frozen forecasts of OLinear and LightGTS6fqk
Validation protocol on short datasets such as ILIHow the validation windows are built and split6fqk
Robustness to the random seedFive seeds with OLinear and Sundial, with δ-Adapter as a reference6fqk
More recent backbonesFACT, ReNF and SimpleTM, and the foundation models TimesFM 2.5 and Moirai 2.0C45Z
Fine-only calibration and regularized fusion baselinesBias, Ridge (fine), Coarse only, Equal and Weighted average, Ridge (fine+coarse), and Linear and MLP fusion tuned on validationC45Z
Contribution of the empirical moments and of their mixing with the structured momentsComponent ablation with LightGTS, and λ from 0 to 1 on all five foundation modelsC45Z
Full cost, including coarse forecasts and coarse modelsEvery stage timed, with storage, for Parallax, δ-Adapter and RGMR on three foundation models and OLinear (RGMR only on the foundation models)6fqk, C45Z, b1iz
Does agreement learned in pretraining make the correction unnecessary?Why agreement between resolutions does not imply a small error, and reconciliation methods that enforce this agreement but do not reliably correct the fine forecastb1iz
Updating the correction in streaming settingsRefits every H, 4H and 16H test windows against δ-Adapter with online update, update time per window, and ΔMSE per test quarterb1iz
Benefit of the longer coarse historyCoarse forecasts limited to the same or half the history of the fine forecastb1iz
How much past data is needed?Fitting on the most recent 2% to 100% of the training windowsb1iz
Tuning of λ across foundation modelsλ sensitivity on all five foundation modelsb1iz

Main findings. Parallax improves every added backbone and lowers MSE in nearly every backbone-dataset-horizon case of the paper. Reconciliation methods enforce agreement between resolutions but can increase the fine forecast error, while Parallax reduces it in every dataset-backbone pair. The tuned calibration and fusion baselines also increase the error in some cases, while Parallax lowers it in every case. Removing either the structured or the empirical moments raises the MSE of LightGTS on all four ablation datasets, so both parts of the mixture are needed. On cost, with every stage counted, the Parallax correction itself is almost free, and the fitted corrector needs less storage than two recent post-hoc correctors. Parallax also gives a larger MSE reduction than both correctors on the foundation models. On robustness, the gain of Parallax is stable across seeds. As for updates, a correction fitted once stays useful over the whole test period, and a refit takes less time than the interval between two observations of the data. Regarding coarse history, Parallax keeps most of its gain when the coarse forecasts see no more history than the fine forecast. Regarding the amount of fitting data, a small fitting set already gives most of the gain of Parallax. Finally, with 5 foundation models, a wide range of λ consistently improves every dataset.

Limitations and future works. First, the corrector fitted once stays useful on our benchmarks and is cheap to refit. Detecting when a refit is needed under stronger shifts, particularly in streaming settings, is beyond the focus of this work and remains open. Moreover, dealing with streaming data is currently out of scope of our paper, though it would be an interesting future work. Second, task-specific backbones need one trained model per coarse resolution, a cost that foundation models avoid. A single task-specific model serving all resolutions is left for future work (Appendix E).

AuthorsOpenReview ↗

Final summary of the discussion

We thank all reviewers and the ACs again for their time and for the constructive discussion. We would like to summarize our rebuttals as follows:

All three Reviewers (6fqk, C45Z, and b1iz) see value in the methodology and practical significance of our paper. All Reviewers highlight that Parallax is a simple, interpretable, and model-agnostic correction without backbone update. Parallax gives consistent and substantial improvements across diverse datasets, foundation models, and task-specific models, which is a practical contribution (acknowledged by Reviewers C45Z and b1iz). The paper has a clear presentation of the correction process and its design, which is supported by the ablations (noted by Reviewers 6fqk and b1iz).

We are grateful that Reviewer C45Z noted that their concerns have been addressed and raised the score accordingly. The concerns include (1) adding more recent backbones, (2) additional calibration and regularized fusion baselines, (3) ablation of moments, and (4) running cost analysis. We provided new experiments with the 5 most recent backbones, 8 calibration and fusion schemes, ablation of moments, and detailed running cost analysis. Based on their feedback, we summarize our key findings. Parallax improves every added backbone and lowers MSE in nearly every backbone-dataset-horizon case of the paper. Tuned calibration and fusion baselines increase the error in some cases, while Parallax lowers the MSE in every case. The ablation shows that removing either the structured or the empirical moments raises the MSE and that empirical moments alone are unstable, which our design avoids. On cost, the Parallax correction itself is almost free, and the fitted corrector needs less storage than two recent post-hoc correctors.

We sincerely thank Reviewers 6fqk and b1iz for their constructive feedback. Although we did not have chances to engage with Reviewers 6fqk and b1iz during the rebuttal, except the initial reviews, their constructive comments helped us improve our paper. We believe our responses have addressed their concerns. Reviewer 6fqk raised (1) temporal reconciliation and ridge correction baselines, (2) the validation protocol on short datasets, and (3) random seed analysis. Reviewer b1iz raised (4) forecast correction in streaming settings, (5) whether pretraining could make the correction unnecessary, and (6) the effect of the coarse history, the amount of fitting data, and the parameter λ. Both reviewers asked about (7) running cost analysis. We added new reconciliation and correction baselines, a description of the validation setting, results on five seeds, experiments in a streaming-like setting, ablations of the coarse history, the fitting-set size, and λ, and a detailed running cost analysis. We summarize our key findings. For Reviewer 6fqk, reconciliation methods enforce agreement between resolutions but can increase the fine forecast error, while Parallax reduces it in every dataset-backbone pair. Every validation window keeps its full lookback and horizon, and the gain of Parallax is stable across seeds. For Reviewer b1iz, the reconciliation results also show that agreement between resolutions, even if learned during pretraining, does not guarantee an accurate fine forecast. A correction fitted once stays useful over the whole test period, and a refit takes less time than the interval between two observations. Parallax keeps most of its gain when the coarse forecasts see no more history than the fine forecast, and a small fitting set already gives most of the gain. With 5 foundation models, every λ in a wide range improves every dataset, so little dataset-specific tuning is needed. For both reviewers, all reported times include the input preparation, every fine and coarse forecast, and the training of coarse models. The correction itself is almost free, and Parallax gives a larger MSE reduction than both post-hoc correctors on the foundation models.

We thank all Reviewers and ACs once more for their insightful feedback.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.