acceptodds
Under review as a conference paper at ICLR 2027

STEP: Long-Term Time Series Forecasting Through the Lens of the Forecast Start

Abstract

Long-term time-series forecasting typically focuses on modeling the distant future, yet prediction becomes increasingly uncertain as the forecast horizon grows. We revisit this problem from the forecast start and argue that it may play a larger role in long-horizon accuracy than its position suggests: improving the forecast start can benefit subsequent forecasting. We propose STEP, which improves forecasting by providing a more accurate initial forecast and further strengthening the forecast start, while also serving as an enhancement for other forecasting backbones. First, STEP encodes adjacent temporal evolution through historical differences, then enriches these representations with learned feature shifts, cross-variable interaction, and subspace interaction to improve the initial forecast. Second, STEP treats the initial forecast as a natural continuation of history and jointly models the two, reinforced by a transition-aware training objective that explicitly emphasizes the forecast start. Experiments on multiple real-world datasets show that STEP achieves competitive forecasting performance and improves diverse forecasting backbones. With training-based integration, STEP improves seven different backbones in 96.4% of the evaluated settings. At inference time, STEP can provide an improved forecast start to a frozen backbone, reducing subsequent forecasting error. Code is available at https://anonymous.4open.science/r/STEP-DDD5.

Primary area:
learning on time series and dynamical systems
Keywords:
Long-Term Time Series Forecasting, Multivariate Time Series Forecasting, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_We15OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that the forecast start, i.e., the transition from history to the first predicted step, plays a larger role in long-term forecasting accuracy than its position suggests. The motivation comes from an experiment in which the ground-truth first future value is appended to the input of a frozen model and the subsequent horizon is re-forecast (Fig. 1, App. C). The authors propose STEP, which builds an initial forecast from multi-order historical differences, learned circular shifts on feature segments, cross-variable attention, and cross-head ("subspace") attention, and then refines it with a shared linear map over the concatenated history and initial forecast. Training uses a transition-aware loss with exponentially decaying horizon weights and value and adjacent-change terms. STEP is evaluated as a standalone forecaster, as a module integrated into seven backbones, and as an inference-time provider of an improved forecast start for frozen backbones.

STEP achieves the best MSE at all horizons on Electricity, Traffic, and Weather in Table 1, and the 3-seed results show low variance. The appendix is unusually transparent: it includes a capacity-matched FFN control, paired block-bootstrap confidence intervals, a dose-response analysis of the start correction (Fig. 8), and a comparison against the untouched one-shot forecast that also reports negative results (Tables 16–17).

However, the forecast-start narrative that defines the paper is only weakly supported by the paper's own controlled results, and it does not explain where the accuracy gains come from. The inference-time enhancement has a very small effect against an honest baseline, and the backbone-integration results mix architecture, capacity, and loss changes. The method is largely a combination of known components. Several reported numbers also need clarification.

Critical strengths
  • The history-to-future boundary as an object of study, with a clean intervention protocol: Instead of asking only how to encode the history, the paper asks how much a frozen forecaster depends on where its forecast begins, and tests this by supplying a corrected start a(q) = a + q(Y1 − a) and varying q, including the opposite direction (Fig. 8, App. C). This turns a vague intuition into a measurable dose-response question that can be asked of any direct forecaster. The protocol is useful beyond STEP, and it is also what lets a reader check how large the effect really is.
  • Treating the initial forecast as a continuation of the history: STEP places the initial forecast next to the observed window and refines the joint sequence [X; Ỹ] (Eq. 12), so each forecast position can see the recent history levels and changes rather than only a latent summary. This is a simple, architecture-agnostic principle, and it is what allows the same block to be attached to seven different backbones (Table 2) and to supply a start to frozen models (Sec. 4.3).
Critical weaknesses
  • The central claim is contradicted by the paper's own full-test evidence: The paper is framed around the claim that "long-range forecasting performance is closely related to how prediction begins" (Sec. 1), but its controlled measurements put this effect at a small fraction of the reported gains.
    • The 44%–98% reductions in Fig. 1 are selected examples from frozen STEP variants ("in these examples," p. 1–2). On the full test sets, even a perfect ground-truth start reduces later MSE by only 1.30% (Electricity), 0.88% (Traffic), 1.57% (Weather), and 3.70% (Solar) (Fig. 8, App. C).
    • At the same H = 96, STEP's margin over the best baseline is larger than this upper bound (Table 1: Electricity .127 vs. .134, Weather .148 vs. .156, about 5% each). So most of the improvement must come from something other than a better forecast start.
  • What the method does differs from the mechanism the narrative describes: The narrative is that first-step error propagates over the horizon, but neither the diagnostic nor the model works through such propagation.
    • The intervention appends a value to the input window (X⁺(a) = [x2, …, xL, a], Eq. 13). For a direct multi-output model, this shows that one more recent observation helps; it does not show that the model's own first-step output error drives later errors, since later outputs do not depend on that output.
    • Inside STEP, the start is singled out only by the loss weights w_h = exp[−λ(h−1)] (Eq. 7); the refinement treats all of Ỹ alike. The evidence that the transition design specifically fixes the start (16.9% step-1 gain, Fig. 6(b)) comes from a control that also switches the loss to MSE (App. G.2), so refinement and objective are not separated. The components themselves (residual linear refinement, horizon-weighted MSE+MAE+change loss similar to TDAlign, difference embeddings, cross-variable attention) are largely known, so the forecast-start framing is what would make the contribution distinct.
  • The two settings meant to show the idea's generality do not isolate the forecast-start effect:
    • Inference-time enhancement: the main "Control" is worse than the untouched one-shot forecast (Table 15 vs. Table 17), and the Q4 subgroup is selected with the ground truth (s_n = MSE^Control − MSE^GT, Eq. 14; acknowledged in App. K). Against the one-shot forecast on the full test set, the gains range from −0.36% to +0.99%, with 4/28 regressions (Table 17).
    • Backbone integration ("96.4% of settings"): integrated models are trained with the Eq. 7 objective (App. E.1), with no "backbone + Eq. 7 loss only" or capacity-matched control, although Table 4 suggests the loss effect is large (Solar MAE .199 → .271 without the STEP objective). For iTransformer and Client, the native encoder is replaced by STEP (Table 12).
Questions for authors
  • Contribution of the forecast start: Given that a perfect ground-truth start reduces later MSE by only 1–4% (Fig. 8), what share of STEP's margin in Table 1 can be attributed to forecast-start improvement? Can you show, e.g., the correlation (controlling for window difficulty) between step-1 error and later-step error in trained models?
  • Integration controls: Can you report (a) each backbone trained with the Eq. 7 loss only and (b) each backbone with a parameter-matched FFN block, plus seed variance and ETT results for +STEP?
  • Inference-time enhancement without ground truth: Does the effect remain when windows are selected without ground truth, and is it better than simply using STEP's own forecast or a simple average of STEP and the backbone?
  • Table 1 numbers and baselines: Which run is reported in Tables 1/10 (best seed, a separate run, or a mean)? Were baselines re-run under the same pipeline and test windows, and were they tuned comparably?
  • Refinement vs. loss in Fig. 6(b): What is the step-1 reduction when the control keeps the Eq. 7 loss and only replaces refinement with an FFN? How do per-horizon error curves at H = 720 compare between the chosen λ and λ = 0?
3 comments
AuthorsOpenReview ↗

Response to Reviewer We15 (Part 1)

We thank the reviewer for the detailed and constructive feedback, particularly for recognizing the value of studying the history-to-future boundary through controlled interventions.

Q1. How much of STEP's gain comes from the forecast start? Are first-step and later improvements related?

The 1%–4% gains in Figure 8 come from changing only the supplied start of a frozen predictor. In contrast, Table 1 evaluates the full STEP model, which is designed around the forecast start and trained end to end. Training changes the model parameters and improves its overall forecasting capability. Therefore, the input-intervention gain is not a ceiling on STEP's gain, and subtracting the two cannot identify the causal contribution of the first step.

We also performed a paired window–variable analysis on Electricity (L = H = 96), comparing STEP with w/o all modules in Table 4. Among samples whose step-1 MSE improved, we formed ten groups by the magnitude of the step-1 improvement. To reduce differences in sample scale and improvement magnitude, we aggregated the logarithms of per-sample relative RMSE improvement factors. The reduction in RMSE at steps 2–96 increased from 2.84% to 7.92% across the groups. This conditional analysis shows that samples with larger first-step improvements generally also have larger later-error reductions.

Electricity: Conditional groups by step-1 improvement

GroupPaired samplesMedian step-1 MSE improvementRMSE reduction, steps 2–96
186,5540.0001002.84%
286,5550.0004823.62%
386,5550.0011063.69%
486,5550.0020343.89%
586,5550.0034244.03%
686,5550.0055384.38%
786,5550.0089164.65%
886,5550.0149295.19%
986,5550.0278825.96%
1086,5550.0780667.92%

Q2. What do loss-only, parameter-matched FFN, multi-seed, and ETT integration controls show?

We added Eq. (7) loss-only and parameter-matched FFN controls. The following are two completed, strictly matched comparisons on Electricity (L = H = 96):

Electricity: MSE with loss-only, matched FFN, and +STEP

BackboneLoss-onlyMatched FFN+STEP
Client0.14220.15430.1345
GTR0.13400.13380.1328

For Client and GTR, +STEP reduces MSE relative to loss-only by 5.43% and 0.95%, respectively, and relative to the matched FFN by 12.85% and 0.79% (calculated from unrounded values). These matched comparisons show that neither the loss change nor extra parameters alone explain the gains.

We further evaluated +STEP integration with seven backbones on ETTh1 and ETTh2 using three paired random seeds for each backbone (L = H = 96). The table reports mean MSE ± SD, from the original backbone to +STEP.

ETT integration: Three-seed MSE (original → +STEP; mean ± SD)

BackboneETTh1ETTh2
DLinear0.390 ± 0.012 → 0.375 ± 0.0060.315 ± 0.007 → 0.294 ± 0.002
PatchTST0.379 ± 0.005 → 0.373 ± 0.0030.290 ± 0.004 → 0.287 ± 0.004
iTransformer0.390 ± 0.003 → 0.370 ± 0.0030.314 ± 0.007 → 0.295 ± 0.002
TimeMixer0.388 ± 0.006 → 0.377 ± 0.0030.302 ± 0.004 → 0.294 ± 0.001
Client0.392 ± 0.004 → 0.374 ± 0.0010.315 ± 0.009 → 0.296 ± 0.001
TQNet0.378 ± 0.003 → 0.372 ± 0.0010.297 ± 0.010 → 0.289 ± 0.001
GTR0.373 ± 0.003 → 0.369 ± 0.0020.291 ± 0.004 → 0.287 ± 0.003

Using unrounded values, MSE and MAE improve in 20/21 and 21/21 paired ETTh1 runs, respectively, and in 21/21 paired ETTh2 runs for both metrics.

AuthorsOpenReview ↗

Response to Reviewer We15 (Part 2)

Q3. Without ground-truth-based selection, how does start replacement compare with full STEP and forecast averaging?

Inference-time enhancement does not use future ground truth to select samples. Q4 is only a post-hoc sensitivity group. Appendix Table 17 compares start replacement directly with the original one-shot forecast over the full test set: subsequent MSE improves in 24 of 28 settings (range: −0.36% to +0.99%).

We also compared four strategies over the full Electricity test set (L = H = 96), evaluating the original forecast positions 2–96 in every case:

Electricity: MSE at steps 2–96 (lower is better)

BackboneOriginalFull STEP50/50 averageStart replacement
DLinear0.19600.12880.14290.1953
iTransformer0.14850.12880.13230.1484
Client0.14280.12880.13060.1426
TQNet0.13520.12880.12770.1351
GTR0.13530.12880.12670.1351

Full STEP or forecast averaging generally gives lower error than replacing only the first step. We use start replacement to test a frozen backbone's response to a local intervention and the feasibility of a targeted correction without changing its weights, not as a claim of optimal inference accuracy.

Q4. Which runs appear in Tables 1/10? Were baselines retrained under the same pipeline?

Tables 1/10 report the single-run test results used in the submission, not three-seed means, and they were not systematically selected as the best seed. For example, Traffic (H = 192) has MSE 0.409 in Table 1, while the three-run mean in Appendix Table 11 is 0.4083 ± 0.0006. Thus, at least one run was better than the main-table result. Appendix Table 11 reports the full three-run statistics separately.

For baselines, we mainly follow published benchmark results rather than rerunning and retuning every method in a single pipeline. The comparisons use the standard datasets, input lengths, prediction horizons, and evaluation metrics.

Q5. What are the separate effects of Refinement and time-weighted Loss at H = 720?

The Transition Design comparison in Figure 6(b) changes both Refinement and the training objective. Its 16.9% step-1 and 6.2% later-step MSE reductions therefore cannot be attributed to Refinement alone. We added controlled comparisons on Electricity (L = 96, H = 720):

H = 720: Relative MSE reductions under separate controls

Controlled comparisonStep 1Steps 2–720
Refinement vs. matched FFN (same Eq. (7) Loss)2.93% [1.88%, 3.71%]0.92% [−0.22%, 1.93%]
λ = 0.01 vs. λ = 0 (same STEP architecture)9.29%−0.14%

The brackets in the first row are 95% confidence intervals; the second row gives verified point estimates from a matched-seed comparison. λ = 0.01 is the Electricity setting in Table 9. With the loss fixed, Refinement clearly improves step 1, while the later-step confidence interval includes zero. With the architecture fixed, time weighting primarily affects early predictions.

We further divided the controlled λ = 0.01 versus λ = 0 comparison into forecast ranges to examine how its effect changes across the horizon:

Horizon-wise MSE reductions: λ = 0.01 vs. λ = 0

Forecast positionsMSE reduction95% CI
Steps 1–96+3.30%[2.51%, 4.08%]
Steps 97–360−1.86%[−3.78%, 0.22%]
Steps 361–720+0.65%[−0.74%, 1.99%]

The benefit is concentrated in early predictions, while changes at later horizons remain uncertain. This is consistent with the design's emphasis on the forecast start and early forecast positions.

Reviewer_We15OpenReview ↗

Reponse to Authors

Thank you for the response and the new controls. The 3-seed ETT integration results with seven backbones, and the clarifications on Tables 1/10 and the Q4 subgroup, address my questions on reporting and generality. I also agree that the frozen-model intervention (Fig. 8) is not a strict upper bound for a model trained end to end, so my "upper bound" wording was too strong.

However, the new controlled results make the forecast-start narrative harder to support:

  • Q5 (H=720): With the loss fixed, refinement improves step 1 by 2.93% but steps 2-720 by only 0.92%, with a CI that includes zero. Time weighting improves step 1 by 9.29% and changes steps 2-720 by -0.14%. In the trained model, a better start does not carry over to later steps, which is the opposite of the claim that "long-range forecasting performance is closely related to how prediction begins".
  • Q1: The grouped analysis is correlational and only uses samples whose step-1 error improved. Even the group with almost no step-1 change (median 0.0001) has a 2.84% later reduction, so a large part of the later gain does not depend on the start.
  • Q3: Start replacement improves steps 2-96 by at most about 0.4% (e.g., DLinear 0.1960 -> 0.1953), while full STEP or a simple 50/50 average gives much larger gains. This supports reading the inference-time results as a feasibility check rather than evidence for the start mechanism.
  • Q2: On GTR, +STEP improves over loss-only by 0.95% and over the matched FFN by 0.79%. On Electricity, where strong baselines are very close (TQNet 0.1352 vs. GTR 0.1353 in your Q3 table), this is a meaningful gain, so the module itself does contribute. However, these controls cover one dataset and two backbones, the Client matched FFN (0.1543) is worse than loss-only (0.1422), which suggests this control is not well tuned, and the ETT runs have no loss-only control.
  • The current PDF does not include any of these results, and the baselines are taken from published results rather than re-run.

STEP looks like a useful plug-in module, but the evidence points to the loss and the full refinement, not the forecast start, as the source of the gains. I will keep my score and suggest reframing the paper around what these controls show.

Reviewer_FEQ3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that the forecast start (the first predicted step) plays an outsized role in long-term time-series forecasting, citing 44-98% reductions in subsequent cumulative error when the ground-truth first value is fed to a frozen model (Fig. 1, Fig. 7). It proposes STEP, which improves the initial forecast via a difference embedding, learned circular shifts of feature segments, cross-variable attention and cross-head attention, and strengthens the history-to-future transition via a shared linear refinement trained with a value-plus-change loss whose weights decay along the horizon. Experiments on eight benchmarks with L=96 report competitive results (Table 1/10), and training-based integration improves seven backbones in 108/112 settings (Table 2). An inference-time protocol replaces a frozen backbone's forecast-start value with STEP's prediction (Table 3, App. J).

Critical strengths
  1. The paper goes further than most LTSF submissions in trying to isolate mechanisms: capacity-matched FFN controls (Fig. 5, Table 14), matched reruns, a graded intervention study with block-bootstrap CIs (Fig. 8), comparisons against untouched one-shot forecasts (Table 16/17), and multi-seed results (Table 11). This lets a reader check how large the forecast-start effect actually is.
  2. STEP gives the lowest MSE at all horizons on Electricity, Traffic and Weather (Table 1), and training-based integration improves 108/112 backbone-dataset-horizon settings (Table 2). If the attribution issues below are resolved, this is a practically useful module.
Critical weaknesses

These datasets are nearing performance saturation, so testing on more datasets is recommended. Existing results show that the 44–98% gains from forecast-start come from selected variables/windows; on full test sets, the oracle reduces MSE by only 1.30%/0.88%/1.57%/3.70%, while STEP’s Table 1 gains already exceed this ceiling, with the main contribution still coming from generic cross-variable attention. The inference-time enhancement (Table 3) is also inflated because Q4 uses the ground-truth start; Control is worse than the one-shot forecast in 27/28 settings, and the improvement ranges from -0.36% to +0.98%, being negative in 4/28 settings.

Questions for authors
  1. Could you clarify why STEP’s gains exceed the full-test-set oracle gain achieved with a ground-truth start? This point deserves further analysis.

  2. Please report seed means for the main results table and evaluate the inference-time enhancement without using ground truth to define Q4, if my understanding is correct.

2 comments
AuthorsOpenReview ↗

Response to Reviewer FEQ3

We thank the reviewer for the constructive feedback and for acknowledging STEP's strong forecasting performance and our efforts to investigate its underlying mechanisms.

Q1. Why can STEP's overall gains exceed those from supplying the ground-truth forecast start?

The two experiments measure different things. Figure 8 changes only the supplied forecast-start value of a frozen predictor; its learned parameters and predictive capability remain unchanged. In contrast, STEP is designed around the forecast start and trained end to end. Its new encoder changes the learned mapping from history to future, improving the model's forecasting capability. Therefore, the gain from a local input intervention is not an upper bound on the gain from training STEP.

The start-oriented encoder also brings an additional benefit: the representations learned to improve the first prediction can help later predictions. In the parameter-matched FFN comparison in Figure 6(b) and Appendix G.2, with the other components and loss unchanged, the initial forecasting design reduced MSE by 2.5% at step 1 and 11.0% at steps 2–720. The component ablations in Table 4 further support this result. Thus, STEP improves what the model learns, rather than merely replacing one input value.

Q2. What do the three-seed results show? Does inference-time enhancement require ground truth to select Q4?

Appendix Table 11 already reports the mean and standard deviation across three runs for all 32 dataset–horizon settings. Below are the H = 96 results, compared with the best baseline MSE in Table 1:

H = 96: STEP MSE over three runs (mean ± SD; Appendix Table 11)

DatasetBest baseline MSESTEP MSE (mean ± SD)
Electricity0.1340.1287 ± 0.0009
Traffic0.3980.3806 ± 0.0010
Weather0.1560.1497 ± 0.0014
Solar0.1760.1787 ± 0.0022

By the three-run mean, STEP outperforms the best baseline on Electricity, Traffic, and Weather; it is slightly worse on Solar. These statistics supplement, rather than replace, the single-run test values in Table 1.

Inference-time enhancement does not use future ground truth for sample selection. Q4 is defined only for post-hoc analysis of forecast-start sensitivity. Appendix Table 17 compares start-replacement reruns with the original one-shot forecasts on the full test set, without Q4 filtering. Subsequent-step MSE decreases in 24 of 28 settings, with changes ranging from −0.36% to +0.99%. This tests the effect of a local intervention on a frozen backbone; it is distinct from the gains of a full STEP forecast or forecast averaging.

OK, thanks for your response, I got it. But I still believe more datasets would be helpful, otherwise it’s still hard to believe the performance gains meet my expectations. I’ll keep my score at 3, and I hope you can provide more results in the camera-ready version if the paper gets accepted.

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

The paper proposes STEP for long-term multivariate time series forecasting. STEP combines historical-difference embeddings, learned feature shifts, variable interaction, and subspace interaction to produce an intermediate forecast. A shared temporal linear map combines this forecast with history, and training emphasizes early forecast positions through value and adjacent-change losses with exponentially decaying weights. Experiments cover eight benchmarks, training-based integration into seven backbones, and inference-time enhancement through forecast-start substitution in frozen-model reruns.

Critical strengths
  • Evaluation spans eight benchmarks, and training-based integration improves MSE in 108 of 112 settings. Component ablations, parameter-matched controls, and repeated-run results provide meaningful evidence that the complete design is useful.

  • The authors distinguish untouched forecasts from matched reruns, exclude the supplied point from subsequent-error evaluation, and provide full-test results and block-bootstrap intervals. These controls strengthen the evidence that changing the supplied start value can affect subsequent rerun predictions.

Critical weaknesses
  • The central forecast-start mechanism remains insufficiently established. The intermediate forecast is supervised only through an unrestricted linear refinement. Permuting its temporal coordinates and inversely permuting the refinement columns leaves final predictions unchanged, weakening its interpretation as a temporally aligned forecast. Moreover, the shifted-window interventions demonstrate sensitivity to a modified input, rather than establishing that first-step improvements explain the original one-shot forecasting gains. General history skip connections, forecast mixing, and loss regularization remain plausible explanations.

  • The practical value of inference-time correction is unclear. It requires an independently trained STEP predictor, but lacks comparisons with using STEP’s complete forecast or combining both models’ predictions. Against untouched one-shot forecasts, Table 17 reports positive full-test improvements of approximately 0.03%–0.99%, with four regressions. These modest gains need to justify the additional predictor and computation. Larger Q4 gains concern retrospectively selected groups and do not establish a deployable selection strategy.

Questions for authors
  • How accurate is the intermediate forecast before refinement, and what establishes its temporal alignment?

  • Does STEP retain its advantage over a backbone equipped with a historical skip connection, forecast-output mixing, and the same loss?

  • How does inference-time correction compare with STEP alone and full-forecast combinations in accuracy and total cost?

2 comments
AuthorsOpenReview ↗

Response to Reviewer brvj

We thank the reviewer for the thoughtful feedback and for recognizing the breadth of our evaluation and the value of our controlled experiments.

Q1. How accurate is the intermediate forecast before Refinement, and how is temporal alignment established?

We treat the intermediate forecast in Eq. (5) as a representation of future information, not as a final forecast with separate supervision and guaranteed temporal alignment. It therefore need not match the accuracy of the refined output. We directly measured both outputs on Electricity (L = H = 96):

Electricity: Errors before and after Refinement in the full STEP model

Forecast positionsBefore Refinement (MSE / MAE)After Refinement (MSE / MAE)
Step 10.570 / 0.6260.046 / 0.133
All 96 steps0.570 / 0.6140.127 / 0.220

Table 4 reports a different experiment: after removing Refinement, we retrained the initial forecasting module with direct supervision. Its MSE was 0.129 on Electricity, 0.392 on Traffic, and 0.153 on Weather. This shows that the initial module can forecast on its own, but these results should not be confused with the intermediate outputs of the fully trained STEP model in the table above.

We acknowledge the permutation equivalence of the intermediate representation and do not claim that its temporal coordinates are mathematically unique. However, inputs and future targets are always paired in a fixed chronological order, allowing the model to learn position-dependent relationships. We shifted the intermediate outputs by −8 to +8 steps relative to the targets for two Electricity checkpoints. Both achieved their lowest MSE at zero shift, providing empirical evidence of temporal correspondence.

Electricity: Temporal shift test of the intermediate forecast

CheckpointTested shifts (steps)Shift with lowest MSE (steps)
1−8 to +80
2−8 to +80

Q2. Does STEP outperform history skip connections and simple forecast mixing under the same loss?

Yes. We added simple history-skip and forecast-mixing controls trained with the same Transition-aware Loss in Eq. (7). In four verified, matched Electricity experiments (L = H = 96), +STEP achieved the following error reductions over these controls:

Electricity: +STEP gains over simple controls

BackboneMSE reductionMAE reduction
DLinear20.54%10.30%
iTransformer9.67%4.98%
Client3.19%2.39%
GTR0.50%0.41%

Thus, simple skip connections or output mixing do not fully account for STEP's gains under the same loss. The component ablations in Table 4 and the comparison of the initial forecasting and history-to-future transition designs in Figure 6(b) provide further evidence for the two design components.

Q3. How does start replacement compare with STEP alone and full-forecast averaging in accuracy and cost?

We use start replacement as a local intervention: STEP supplies the first predicted value to a frozen backbone, which then reruns to predict the later steps. This tests the backbone's response to the supplied start and explores a correction that does not change its weights. It is not intended to be the most accurate full-forecast strategy.

We compared the original backbone forecast, the full STEP forecast, their 50/50 average, and start replacement over the full Electricity test set (L = H = 96). All MSE values evaluate the original forecast positions 2–96. We show four decimals because the differences from start replacement are small.

Electricity: MSE at steps 2–96 (lower is better)

BackboneOriginalFull STEP50/50 averageStart replacement
DLinear0.19600.12880.14290.1953
iTransformer0.14850.12880.13230.1484
Client0.14280.12880.13060.1426
TQNet0.13520.12880.12770.1351
GTR0.13530.12880.12670.1351

Full STEP or full-forecast averaging generally gives lower error than replacing only the first step. Start replacement avoids retraining the backbone but requires an additional STEP forward pass and a backbone rerun. The following end-to-end per-window latencies include data transfer (RTX 4090, FP32, batch size 1):

Inference latency (ms per window)

BackboneOriginalFull STEP50/50 averageStart replacement
TQNet0.4272.3122.6792.687
GTR0.3392.3482.6122.625

Start replacement therefore provides evidence about a local mechanism and a way to preserve the backbone's learned parameters. Its practical value depends on the accuracy–cost trade-off.

Thank you to the author for supplementing the experiment. The matching comparison with the same loss on Electricity shows that the improvement of complete STEP cannot be fully explained by simple historical direct connection and output mixing. The two checkpoints have the lowest error at zero displacement, indicating that the intermediate representation retains a certain temporal order. However, what concerns me the most is: in a one-time prediction, how many of the later steps become accurate because the first step becomes accurate? Equation (6) allows for the use of all intermediate coordinates for each output position, so the first and subsequent steps may all become accurate together due to sharing better representations. Sliding window rerun is to place a value at the end of the new input window and make another prediction. It indicates that this input value will affect the subsequent results, but it does not distinguish how the original forward profit came about.

This also affects my judgment of innovation. GBT has provided a starting point for the second stage with preliminary predictions, and TDAlign has also monitored the predicted values and changes in adjacent steps together. The novelty of STEP mainly lies in integrating shift attention, linear mixing of history and preliminary prediction, and weighting of early positions into the same model. The results on multiple backbones indicate that this combination is effective, but the most recognizable proposition of the paper is that 'if the first step is done accurately, it can drive subsequent predictions', and this mechanism has not been established in current experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.