acceptodds
Under review as a conference paper at ICLR 2027

Morpheus-2.0: Multiscale Probabilistic Forecasting with Dual-Axis Attention

Abstract

A universal forecaster is trained once and applied everywhere, so it must serve tasks whose schemas, forecast horizons, and temporal resolutions all differ. The number of targets and covariates changes from dataset to dataset, horizons vary by orders of magnitude, and sampling rates range from seconds to years. Existing time-series foundation models address these only in part, since they isolate each series or visit the two grid relations in separate blocks, and they fix the horizon and resolution before training begins. We introduce Morpheus-2.0, a compact encoder-only foundation model that represents targets and covariates on a single variable–time grid. Its dual-axis attention runs both relations side by side in every layer, one across variables at a timestep and one across time within a variable; a learned gate weights the two per layer, and an axis-aware rotary encoding preserves both grid coordinates. The forecast length is sampled during training, so one set of weights spans short to long horizons at no additional inference cost. Morpheus-Multiscale extends the model to several temporal resolutions at once, weighting them according to the series being forecast. Across GIFT-Eval, which spans several domains and frequencies, the two releases give complementary operating points: the Morpheus-2.0 attains relative MASE with 81M parameters, while the larger Morpheus-Multiscale attains , the lowest relative MASE of any published system on the benchmark.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, foundation models, probabilistic forecasting, attention mechanisms, multiscale representation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YzabOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper propose MORPHEUS-2.0, an encoder-only time-series foundation model that use single variable-time grid to represent targets and covariates. The MORPHEUS-2.0 achieve 0.689 relative MASE with 81M in GIFT-EVAL and its multiscale counterpart attains the lowest relative MASE compared to other system on benchmark.

Critical strengths
  1. MORPHEUS-MULTISCALE achieves the lowest relative MASE on GIFT-Evaluation, while its 81M-parameter variant also delivers competitive performance.
  2. The paper is well written, with clear explanations and intuitive illustrations that make the proposed method easy to understand.
  3. The prediction-length randomization and four-stage synthetic-to-real curriculum are thoughtfully designed, demonstrating strong consideration of practical forecasting challenges.
Critical weaknesses
  1. Although the paper emphasizes the computational benefits of dual-axis attention and reduced activation memory, it does not systematically evaluate inference efficiency. MORPHEUS-MULTISCALE’s three encoders with 247M parameters may incur substantial overhead; reporting latency, peak GPU memory, throughput, and FLOPs would better characterize the efficiency–accuracy trade-off.

  2. The performance improvements of MORPHEUS-MULTISCALE may stem from ensembling rather than multiscale modeling. As shown in Appendix D4, combining three encoders with the same patch width already reduces relative MASE from 0.689 to 0.669, while using different patch widths further improves it to 0.653. This suggests that much of the improvement comes from standard ensembling, with the distinct contribution of multiscale fusion being relatively modest.

  3. Table 2 demonstrates the benefits of attention factorization and learnable gating, but it does not establish whether simultaneous processing of both axes outperforms the conventional alternating axial attention discussed in Section 4.3. A comparison with an alternating axial attention baseline under matched parameter counts, model depth, and training budgets would therefore strengthen the paper.

Questions for authors
  1. Could the authors provide additional zero-shot evaluation results on independent benchmarks beyond GIFT-Eval?

  2. Could mixing-weight visualizations clarify whether the context-conditioned weights align meaningfully with signal characteristics, forecast horizons, and sampling frequencies?

7 comments
AuthorsOpenReview ↗

Summary of revisions

Thank you for the review. We added everything you requested: measured inference cost, an alternating axial baseline at matched budget, equal-weight and per-horizon controls that separate ensembling from the multiscale gain, and an analysis of the learned mixing weights on GIFT-Eval. Results on an additional benchmark will follow in this thread. With the revisions, Morpheus-Multiscale reaches 0.642 relative MASE and 0.454 relative CRPS, the lowest of any published system on both (Table 1).

AuthorsOpenReview ↗

W1 - Inference efficiency

Appendix E7 now reports this. Both models run on the same A100 at fp32 on identical GIFT-Eval requests, timing every encoder, the mixing head and pooling. Every timed forecast matched the evaluated output, so cost and accuracy refer to the same predictions.

On short-horizon configurations, Morpheus-Multiscale needs about 3.5× the FLOPs of Morpheus-2.0, 4.5× the latency and 2.2× the peak memory, with about 4× lower throughput. This overhead is the price of accuracy: it buys 7.8% lower MASE and 12.2% lower CRPS, with both lower on almost every configuration. The cost comes from running several encoders, the narrow-patch ones on longer sequences, and is fixed by request shape rather than data (FLOPs 3.4–3.7×, memory 2.2–2.3×), so it is known in advance. In high-stakes forecasting, a forecast error costs far more than the extra seconds of GPU time needed to avoid it, and that is the trade the multiscale model makes. Dual-axis attention keeps each encoder compact, with TN2+NT2TN^2+NT^2 query–key pairs instead of (TN)2(TN)^2 and under half the activation memory (Section 7.1).

We therefore release two operating points. Morpheus-2.0 suits high-volume forecasting and is more accurate than the larger Chronos-2. Morpheus-Multiscale targets decisions where errors cost more than compute, such as energy dispatch, inventory and capacity planning, which usually run at planning intervals rather than under latency limits.

Morpheus-Multiscale relative to Morpheus-2.0 (Appendix E7)Change
FLOPs per request3.5×
Latency, one request per call4.5×
Throughput, four requests per call4× lower
Peak GPU memory, four requests per call2.2×
Relative MASE, short horizons−7.8%
Relative CRPS, short horizons−12.2%
AuthorsOpenReview ↗

W2 - Ensembling versus multiscale modeling

Appendix D4 addresses this question, and we have expanded it with equal-weight and best-per-horizon controls (Table 13, below). They show that ordinary ensembling saturates, and that the remaining gain needs both multiple resolutions and context-dependent weighting.

Ensembling at one resolution reaches its limit quickly. Three patch-4 encoders improve MASE from 0.689 to 0.669, and that is all ensembling at one resolution can contribute; spending the same budget on one larger patch-4 encoder is far worse. Adding resolutions goes further. Replacing two of the patch-4 copies with patch-2 and patch-8 encoders improves the combination on both metrics, even though each new encoder is individually weaker than the copy it replaces, because each width sees structure the others cannot resolve: in Table 3, patch 2 wins on oscillations and wide schemas, patch 4 on impulses and noisy series, and patch 8 on persistent levels. No single width, and no ensemble of one width, covers all of these, which is why multiple resolutions are required.

Equal averaging cannot fully exploit this, because where one width is right the others pull the forecast away from it. The mixing head reads the context and weights the right width, lowering MASE to 0.653, whereas choosing the best encoder per horizon, even with hindsight, does worse than equal averaging.

Together, resolution diversity and context weighting account for 2.4% of the 5.2% gain over the standalone model, nearly half. At the top of GIFT-Eval that is substantial: it exceeds the spread across the five systems from Falcon-2.0 to Toto-2.0-2.5B-FT in Table 1.

Combination of encoders (Table 13)ParamsMASECRPS
Patch 2, single encoder84.3M0.6960.507
Patch 4, single encoder81.4M0.6890.522
Patch 8, single encoder80.1M0.7080.526
Patch 4, single encoder scaled to multiscale budget247M0.8380.602
Three patch-4 encoders, equal weights245M0.6690.492
Widths 2, 4, 8, best encoder per horizon246M0.6670.494
Widths 2, 4, 8, equal weights246M0.6610.490
Widths 2, 4, 8, mixing head247M0.6530.485
Reviewer_wAGLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Morpheus-2.0, a time series foundation model that represents targets and covariates on a variable-time grid. Each layer computes attention across variables and across time in parallel, then combines the two outputs with a learned scalar gate. The model is trained with randomly sampled prediction lengths and outputs marginal quantiles. The multiscale version combines three separately trained models with patch widths 2, 4, and 8, using context-dependent weights. The paper evaluates these models on GIFT-Eval and reports improvements in point forecasting accuracy.

Critical strengths
  1. The framework handles targets, past-only covariates, and known-future covariates in one representation. The observation masks and attention restrictions make the available information at the forecast origin explicit.

  2. The experimental analysis is comprehensive. The authors separately examine attention, positional encoding, prediction lengths, patch widths, synthetic data, and the training curriculum.

  3. The multiscale experiments include useful controls. In particular, comparing three models with the same patch width against three different widths helps separate ordinary ensembling from the benefit of multiple resolutions. The reported point forecasting results are competitive at a relatively small parameter count.

Critical weaknesses
  1. The novelty relative to previous forecasting models needs to be clarified. Axial attention is already established (Ho et al., 2019), and Moirai already uses a masked-encoder formulation for universal forecasting with varying numbers of variables and multiple patch sizes. The paper also acknowledges existing methods that vary prediction spans during training. Therefore, the variable-time grid, horizon randomization, and use of multiple resolutions are not individually new.

  2. The most direct comparison for the main architectural claim is missing. The paper emphasizes the advantage of computing both axes in parallel rather than alternating them. However, Table 2 only compares full attention, parallel dual-axis attention with a fixed gate, and parallel dual-axis attention with a learned gate. There is no alternating-axis model trained under the same setting.

  3. The role of the scalar gate needs a clearer explanation. In Equation (1), both output matrices W^I and W^T are trainable. The factors 1-g_l and g_l can therefore be absorbed into these matrices. Even with a fixed gate of 0.5, the model can learn different relative contributions by changing the output projections. This may still be a useful optimization choice, but it does not add expressive power.

  4. The probabilistic forecasts are under-calibrated. Table 16 reports that the nominal 80% interval covers only 47.9% of observations for Morpheus-2.0 and 57.7% for the multiscale model. These are large deviations for a paper on probabilistic forecasting. The multiscale model achieves the best reported MASE in Table 1, but not the best CRPS. The authors discuss median feedback during rollout as a limitation, but do not separate calibration within the native prediction block from calibration after rollout.

  5. The additional value of context-dependent scale weighting could be isolated more directly. I would also like to see the same three scale models combined using equal weights, learned global weights, and horizon-only weights on GIFT-Eval. Otherwise, it is difficult to tell how much the context-dependent mixing head contributes beyond combining complementary models.

Questions for authors
  1. Could the authors add an alternating-axis attention baseline with matched parameter and training budgets? This is the closest comparison for the main architectural claim.

  2. Since the scalar gate can be absorbed into the output projections, what explains its large empirical benefit? Does the improvement remain when the fixed-gate model is given comparable branch-output scaling or initialization?

  3. Could the authors report interval coverage and CRPS separately for forecasts within the native block and forecasts requiring rollout? Is the severe undercoverage already present before median feedback is used?

  4. On the same GIFT-Eval configurations, how does the mixing head compare with equal weighting, learned global weighting, and weighting based only on the requested horizon?

8 comments
AuthorsOpenReview ↗

Summary of revisions

Thank you for the detailed review. We added the alternating-axis baseline, the native-versus-rollout calibration split, and equal- and horizon-weighted controls for the mixing head. Alternating axes is significantly worse at a matched budget; rollout turned out to cause the undercoverage, which uncertainty propagation now fixes (79.4% coverage); and both weighting controls trail the context-conditioned head. We also clarify what is new relative to Moirai and axial attention, and why the scalar gate matters. Each point is answered below.

AuthorsOpenReview ↗

W1 - Novelty relative to Moirai, axial attention and prior span or resolution methods

We agree that the grid, the masked-encoder objective, varied spans and multiple patch sizes each have precedents. They define the setting, not our contribution. What is new is one design in which the operator, its positional encoding, the horizon objective and the resolution weighting are built around the asymmetry of the variable–time grid, and depend on one another.

Moirai's masked encoder. Moirai attends over all pairs of the flattened sequence and marks variables only with a same/different bias, so each token must learn which of TNTN keys are its own history and which are contemporaneous covariates. Our operator builds that distinction in: separate neighborhoods, projections and normalizations, fused by a learned per-layer gate, with relative geometry on both axes at cost TN2+NT2TN^2+NT^2 rather than (TN)2(TN)^2.

Axial attention (Ho et al., 2019). Axial attention updates symmetric axes in sequence, which suits images. Forecasting axes are asymmetric, so our pathways run in parallel on the same input, with per-axis normalization and a balance learned at each depth. A cross-axis dependence, such as a covariate's current value shaping the target's temporal pattern, forms within one layer rather than through an intermediate representation.

Positional encoding. A two-dimensional rotation aliases time with variable offsets over unrestricted pairs, which is why prior models use flattened or time-only positions. Within our neighborhoods each pathway sees one offset, so the rotation is exact and independent of schema width. The encoding is valid only because of the operator, and the operator has relative geometry on both axes only because of the encoding.

Multiple patch sizes. Moirai picks one patch size per series by a fixed rule on frequency, regardless of content. Morpheus-Multiscale learns per-forecast weights over resolution-specialized encoders from the observed context, so resolution follows the signal, not the metadata. A fixed rule is measurably worse: the best width per horizon, chosen with hindsight, trails even equal averaging (Table 13).

Horizon randomization. Related schemes vary context–prediction splits or infill interior spans. Ours samples a terminal prediction budget, so every task is a true forecast, and because the budget is shared across an interleaved canvas, one draw trains a range of horizons and schema widths at once.

To our knowledge, no prior forecasting model combines these four elements. Each is tested against its closest precedent at matched capacity, the operator against full attention, a fixed gate and alternating axes (Section 7.1), the encoding against flattened and time-only RoPE (Section 7.2), horizon sampling against fixed and reweighted objectives (Section 7.3), and the mixing head against same-width, equal and per-horizon combinations (Section 7.4, Appendix D4). Each is the best arm in its comparison, and together they reach the lowest MASE and CRPS of any published system on GIFT-Eval.

AuthorsOpenReview ↗

W2 / Q1 - Alternating-axis baseline

We agree this comparison was missing and have added it. We trained an alternating-axis encoder from scratch: each layer runs only one pathway, intra-time and inter-time in alternation. It is matched to a parallel dual-axis encoder in architecture, total parameters, data, curriculum, optimizer, batch and training steps.

Attention operatorMASECRPSMultivariate MASEMultivariate CRPS
Parallel dual-axis, learned gate0.6850.5190.7030.511
Alternating axes across layers0.7280.5340.7680.543
Change from alternating+6.2% [3.8, 8.6]+2.8% [0.01, 5.8]+9.2% [5.1, 13.5]+6.3% [0.8, 12.0]

Relative metrics; brackets are 90% paired bootstrap intervals.

Alternating is significantly worse: +6.2% MASE and +2.8% CRPS overall, and +9.2% MASE and +6.3% CRPS on multivariate configurations. All paired bootstrap intervals exclude zero. The gap is largest where cross-variable and temporal information must be combined, which supports the claim that computing both axes in parallel within each layer is what matters. This is why dual-axis attention is required in our design: it is superior to alternation under the matched comparison and helps explain why Morpheus-2.0 outperforms Chronos-2, which alternates its mixing operations across axes. We report this in Section 7.1 and Appendix D.

Reviewer_MWz3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This research work introduces Morpheus-2.0, a time-series foundation model that uses dual-axis attention to capture relationships across variables and temporal patterns within each variable, which is a timely extension. It combines a flexible variable-time representation with axis-aware positional encoding and prediction-length randomization. The proposed Morpheus-Multiscale combines three models with different patch widths using a context-conditioned head. Experiments on 97 GIFT-Eval configurations show competitive forecasting performance, with the multiscale model achieving the best reported relative MASE among the compared systems.

Detailed Review: This work addresses an important problem in time-series foundation models. The dual axis attention mechanism is a reasonable design choice, and the experimental section includes several controlled ablations to investigate the contribution of individual components. The multiscale extension is particularly interesting because different patch widths can capture different temporal patterns. The authors also provide useful experiments on synthetic data and a real-world forecasting benchmark to support their design choices. The analysis across domains and forecast horizons helps explain where the proposed models perform well and where they remain limited. However, the improvements are not consistent across all evaluation metrics. Although Morpheus-Multiscale achieves the best relative MASE, its relative CRPS remains worse than that of TimesFM-3. The inference cost of combining three separate encoders also deserves further examination. Overall, the paper presents a technically sound and carefully evaluated approach, but the claims about probabilistic forecasting quality and efficiency need stronger support. The main contribution appears to be a useful combination of architectural and training improvements rather than a fundamentally new forecasting principle.

Critical strengths

Strengths: S1. The variable-time representation supports different schema widths, covariate types, and forecast horizons within a shared architecture. This is useful for developing general-purpose time-series foundation models. S2. The authors evaluate dual-axis attention, positional encoding, horizon randomization, multiscale forecasting, and the synthetic-data curriculum. These experiments help explain the contribution of individual design choices. S3. The experiments cover 97 configurations across several domains and sampling frequencies. The paper also reports results by horizon and domain rather than relying solely on aggregate metrics.

Critical weaknesses

Weaknesses: W1. Morpheus-Multiscale achieves a relative MASE of 0.653 but has a relative CRPS of 0.476, compared with 0.456 for TimesFM-3. Its nominal 80% prediction interval covers only 57.7% of observations. This weak calibration limits the strength of the claims about probabilistic forecasting. W2. The MASE improvement over TimesFM-3 is very low on the aggregate benchmark. The multiscale model also performs worse at short horizons and does not lead on CRPS. The paper should be more careful when presenting the results as evidence of broadly superior forecasting performance. W3. The multiscale model combines three trained encoders trained at different patches and has approximately 247M parameters, compared with 81.4M for Morpheus-2.0. Although the paper discusses attention complexity and memory savings, actual inference latency and throughput are not reported in sufficient detail to establish the practical efficiency of the proposed approach. W4. The proposed approach combines several established ideas, including axial attention and rotary positional encoding. While their combination is useful, the paper could better explain what fundamental advance the method offers beyond adapting and integrating these existing techniques.

Questions for authors

Questions: Q1. Can the authors provide paired confidence intervals for the MASE and CRPS differences between Morpheus-Multiscale and TimesFM-3? Would the main result remain consistent under separate domain-level and horizon-level aggregation? Q2. Can the authors report inference latency, throughput, and peak memory for Morpheus-2.0 and Morpheus-Multiscale on the same hardware? This would help clarify whether the accuracy gains of the multiscale model justify its additional computational cost. Q4. The paper argues that the multiscale combination benefits from context-conditioned weights and uses Jensen’s inequality to motivate quantile averaging. Can the authors clarify the precise scope of this guarantee and provide evidence that the learned weights adapt meaningfully across different forecasting contexts?

Suggestions: R1. Provide additional calibration results, preferably across different forecast horizons and domains. If possible, evaluate a simple post-hoc calibration method using only validation data to show whether the undercoverage can be reduced without substantially degrading CRPS. R2. Can the authors provide confidence intervals for the comparisons with TimesFM-3 using paired bootstrap analysis per configuration results are available? Summarizing the main trade-offs in the main text would make the overall claim clearer. R3. Include practical inference measurements (latency, throughput, and peak memory under a consistent evaluation) for both models. I believe this setup would make the efficiency claims more convincing. R4. Clarify the main methodological contribution relative to existing axial-attention and multiscale forecasting methods. A concise explanation of which aspects are genuinely new and which are adaptations of established techniques would strengthen the paper’s positioning.

8 comments
AuthorsOpenReview ↗

Summary of revisions

Thank you for the careful review. We ran the analyses you asked for and revised the paper. Uncertainty propagation brings 80% interval coverage to 79.4% and gives Morpheus-Multiscale the lowest CRPS in Table 1. Paired bootstrap intervals show that no baseline is significantly better overall, while Morpheus-Multiscale is significantly better than all of them at long horizons. Inference cost is now measured on matched hardware, and we clarify what is new in the attention design and how the mixing weights adapt. Each point is answered below.

AuthorsOpenReview ↗

W1 / R1 - Calibration

You were right that the submitted intervals were too narrow and that this cost probabilistic accuracy. Morpheus-2.0 covers 47.9% overall and 32.7% on rollout configurations; uncertainty propagation raises these to 79.2% and 78.9%. The submitted Multiscale model covers 57.7% overall and 41.3% on rollout configurations, while the final propagated model reaches 79.4% and 81.6%. Both final models are therefore essentially at the nominal 80% target (Appendix E5, Table 20).

Following Reviewer wAGL's request, we traced the shortfall to median feedback, which carries no uncertainty between passes. We now propagate quantile paths: each path feeds its own quantile back through successive passes, and the resulting predictive distributions are pooled before the final quantiles are read. This preserves uncertainty accumulated during rollout; Appendix E5 gives the full procedure.

Central 80% interval coverage (%)All configurationsRollout configurations
Morpheus-2.0, median feedback47.932.7
Morpheus-2.0, uncertainty propagation79.278.9
Submitted Multiscale, median feedback57.741.3
Final Multiscale, uncertainty propagation79.481.6

CRPS improves from 0.476 to 0.454, the best CRPS of any system in Table 1, ahead of TimesFM-3, Chronos-2 and Falcon-2.0. Propagating uncertainty through rollout therefore turns probabilistic accuracy from a weakness into a leading strength of the model.

As you asked (R1), Appendix E6 now reports coverage for every horizon group and domain. It also isolates the one post-hoc step in the final model, a spread factor fitted on the GIFT-Eval training region only. Without it, the propagated model already covers 80.0% overall (79.4% with it), no horizon group or domain moves by more than 2.1 points, and CRPS changes only from 0.4542 to 0.4538. Once uncertainty is propagated, the model is therefore calibrated without any post-hoc adjustment.

AuthorsOpenReview ↗

W2 / Q1 / R2 - Paired confidence intervals by horizon and domain

This analysis strengthens the paper's central claim. The table reports paired bootstrap 90% intervals over the 97 GIFT-Eval configurations (100k resamples) for the final Morpheus-Multiscale model. Values are the relative gain (%) over each baseline; positive means lower error, and bold marks intervals that exclude zero.

TimesFM-3 MASETimesFM-3 CRPSChronos-2 MASEChronos-2 CRPSFalcon-2.0 MASEFalcon-2.0 CRPS
Overall (97)+3.8 [−0.5, 8.1]+0.4 [−4.0, 4.8]+8.0 [3.9, 12.1]+6.5 [2.2, 10.8]+3.6 [−0.9, 8.2]+6.7 [1.4, 11.7]
Short (55)−5.4 [−11.2, 0.9]−9.2 [−15.1, −3.1]−2.4 [−7.5, 3.1]−4.6 [−10.0, 0.9]−4.7 [−10.7, 1.7]−5.5 [−12.7, 1.4]
Medium (21)+11.8 [3.9, 19.3]+8.0 [−0.3, 15.6]+16.7 [9.1, 23.9]+15.5 [7.1, 23.2]+9.3 [−0.5, 18.1]+16.1 [4.9, 25.5]
Long (21)+17.2 [11.1, 23.3]+15.4 [8.6, 21.7]+23.4 [17.2, 29.3]+22.9 [15.8, 29.4]+17.6 [10.5, 24.1]+24.7 [17.6, 31.0]
Horizon-balanced+8.4 [4.5, 12.3]+5.2 [1.1, 9.3]+13.2 [9.4, 16.9]+12.0 [7.8, 16.0]+7.9 [3.3, 12.2]+12.6 [7.6, 17.2]

Morpheus-Multiscale has the best MASE point estimate among the compared systems. Against TimesFM-3, the interval rules out a loss larger than 0.5% while the point estimate favours our model by 3.8%, and no baseline can be shown to outperform it on either metric. It is significantly better than Chronos-2 on both metrics and than Falcon-2.0 on CRPS. Aggregate scores over 97 configurations rarely separate leading systems: Falcon-2.0 and TimesFM-3 are likewise level on MASE (+0.1% [−2.0, 2.2]).

At medium and long horizons the advantage is decisive. Morpheus-Multiscale is significantly better than all three baselines at long horizons on both metrics, by 15–25%, and no baseline is significantly better at medium or long horizons on either metric. With horizons weighted equally, it is significantly better than every baseline on both metrics. TimesFM-3 retains an edge only at short horizons, where the single significant loss in the table occurs. No comparison against Chronos-2 or Falcon-2.0 is a significant loss.

On Energy, the largest domain (32 configurations), Morpheus-Multiscale is significantly better than all three baselines on both metrics, by 12.7–19.4%. It trails on Healthcare and Nature. Per-domain intervals are in Appendix E1. Domain labels are not available at inference, so exploiting domain-specific behaviour requires identifying the regime from the raw context, which is precisely what the mixing head does when it selects the resolution suited to each series (Appendix D3).

Section 6 states this profile, and Appendix E1 now reports the full intervals overall, by horizon and by domain.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to reviewers and Area Chair

We thank the three reviewers for their careful reading, and the Area Chair for overseeing the discussion. We took every weakness and question seriously. Each one is answered with new experiments in the revised paper or with arguments and justification in the individual threads. Since the threads have grown long, we summarize everything below, grouped by theme.

Calibration and significance. Reviewer wAGL suggested separating calibration within the native block from calibration after rollout. That traced the undercoverage to median feedback, which discards uncertainty during rollout. Propagating quantile paths fixes this at inference time without retraining, raising 80% coverage from 57.7% to 79.4% and improving CRPS from 0.476 to 0.454. Reviewer MWz3 asked for paired bootstrap intervals, and they show that no baseline is significantly better overall. Our model is significantly better than TimesFM-3, Chronos-2 and Falcon-2.0 at long horizons, by 15 to 25%, and on Energy, the largest domain.

Novelty, alternating attention and the gate. Moirai attends over all pairs of the flattened grid, and axial attention updates one axis after the other. Our operator is built for the asymmetry of forecasting data. Time and variables each get their own neighborhood, projections and softmax, so a few variables never compete with hundreds of timesteps. Both pathways read the same input, balanced by a learned gate at each depth. Because each pathway sees only one offset, our two-dimensional rotary encoding is exact at any schema width. This operator is the backbone of a model that ranks among the strongest on multiple time series benchmarks. Alternating the axes at the same budget is significantly worse on every measure, by 6.2% MASE and 2.8% CRPS overall and by 9.2% and 6.3% on multivariate data. The fixed and learned gates start from the same function and differ only in nine scalars, yet learning them lowers MASE by 5.5%.

Resolution diversity and context weighting drive the multiscale gain. Morpheus-Multiscale combines encoders with patch sizes 2, 4 and 8. A single encoder scaled to the same budget is 21.6% worse in MASE. Ensembling three patch-4 encoders helps by only 2.9% and then plateaus. Replacing two of them with patch sizes 2 and 8 and weighting by context gains a further 2.4%. In total, MASE drops 5.2% and CRPS 7.1% over a single encoder, and this is the only combination best on both metrics. Even picking the best encoder for each horizon does worse than equal weights. On GIFT-Eval, the mixing head favors small patches for high-frequency and wide multivariate series, intermediate patches for spiky series and large patches for long horizons, as our synthetic study predicts. In every case it picks the more accurate patch size without seeing the targets, and the others trail it by up to 22% in MASE and 32% in CRPS.

Cost and generalization. On the same A100, the multiscale model uses 3.5 times the FLOPs of Morpheus-2.0 in exchange for 7.8% lower MASE and 12.2% lower CRPS at short horizons, so practitioners can choose either. We also evaluated zero-shot on fev-bench, 35 datasets we never developed on. There it is within 1.2% of TimesFM-3 in MASE and ahead of Chronos-2, TimesFM-2.5, TiRex-2 and Toto-2.0, and it is the best model on more datasets than any other system.

We are grateful for these reviews, several of which led to a better model as well as a clearer paper. Morpheus-Multiscale now achieves the lowest MASE and CRPS of any published system on GIFT-Eval with calibrated intervals, and we believe the revised paper makes a clear and well-supported contribution. We would be glad to discuss anything further before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.