acceptodds
Under review as a conference paper at ICLR 2027

SCEU: STEP-CONDITIONED EXOGENOUS UTILITY LEARNING FOR TIME SERIES FORECASTING

Abstract

Exogenous variables provide valuable forecasting context, but their usefulness can vary across variables and future prediction steps. Existing methods organize cross-variable information through channel-independent processing, channel-dependent interaction, or selective clustering, but these interaction strategies alone do not specify how exogenous information should be weighted at each future prediction step. To address this limitation, we propose SCEU, a Step-Conditioned Exogenous Utility learning framework. SCEU constructs an endogenous-only base forecast, assigns relative Channel Utility scores to exogenous variable-scale representations, and derives Temporal Utility representations from multi-scale endogenous historical patches. A gated bounded residual pathway selectively applies the correction obtained by fusing the two branches while controlling its magnitude at each forecast step. The learned utility weights serve as model-internal, step-conditioned scores for weighting endogenous and exogenous representations. Across diverse benchmarks, SCEU achieves strong forecasting performance. Ablation studies show that Channel Utility, Temporal Utility, and the adaptive response range generally improve forecasting accuracy in the evaluated settings. Overall, SCEU provides a step-conditioned mechanism for selectively utilizing exogenous information in time series forecasting. The code is available at: https://anonymous.4open.science/r/sceu-EBCE.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Exogenous Variables, Step-Conditioned Utility, Multi-Scale Modeling

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_po8kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SCEU refines an endogenous-only linear forecast using multi-scale Temporal and Channel Utility representations conditioned on the current target state and forecast position. A gated, bounded residual controls the correction magnitude. The model is trained end to end with MAE loss, without utility annotations. Experiments on 13 datasets report competitive accuracy, component ablations, parameter sensitivity and evaluation under different future-covariate availability levels.

Critical strengths
  1. The motivation is clear. Beyond the CI, CD and CCM perspectives on whether and how channels interact, the paper considers how exogenous information differs in usefulness across future prediction steps. This provides a well-defined motivation for Step-Conditioned Exogenous Utility learning.
  2. The empirical evaluation is broad. The paper evaluates 13 real-world datasets, covering multivariate data and covariate-aware forecasting in electricity-market, energy and other settings.
Critical weaknesses
  1. Fine-grained ablations of the core design are insufficient. Table 2 mainly removes the entire Channel Utility or Temporal Utility branch, or the Adaptive Response Range. However, key choices include step-conditioned queries, state-conditioned utility, Softmax importance scores ah,c,kexa_{h,c,k}^{ex}, tanh signed modulation dh,c,kexd_{h,c,k}^{ex}, and their product u=a⋅du=a\cdot d. Their individual contributions relative to ordinary conditional attention or horizon-invariant weighting remain unclear. Useful ablations would remove step embeddings or state conditioning, replace signed utilities with nonnegative weights, or substitute standard cross-attention for utility scoring.
  2. The impact of bounded residuals on extreme-value forecasting is insufficiently evaluated. The final prediction is y^hnorm=y^hbase+ghrhδh\hat y_h^{norm}=\hat y_h^{base}+g_hr_h\delta_h, with gh∈[0,1]g_h\in[0,1], rh∈[rmin⁡,rmax⁡]r_h\in[r_{\min},r_{\max}] and δh∈[−1,1]\delta_h\in[-1,1]. Proposition 3 bounds the correction by rmax⁡r_{\max}. This may improve stability but could suppress responses to spikes, extreme weather or abrupt electricity-price changes. Aggregate MSE/MAE sensitivity analyses should be complemented by extreme-value subsets, peak errors, quantile losses or high-volatility periods.
Questions for authors
  1. Could utility learning and bounded correction become reusable components for forecasting backbones such as TimeXer, CrossLinear and PatchTST? What interfaces and additional assumptions would this require?
  2. Could the authors investigate patch lengths PP and QQ more thoroughly, including different multi-scale combinations, shared versus distinct P/Q sets, the relationship between scale count KK and horizon HH, and effects on short- and long-horizon forecasting?
1 comment

W1. Fine-grained ablations

We conducted four controlled ablations over 14 dataset–horizon settings: Shared Utility (removing step conditioning), No State (removing state conditioning), Softmax Only (removing signed modulation), and Cross-Attn (replacing utility scoring with standard cross-attention).

VariantAvg. MSE/MAE increaseFull wins (MSE/MAE)
Shared+3.28%/+2.42%12/14, 13/14
No State+1.08%/+0.97%8/14, 10/14
Softmax Only+2.40%/+1.07%10/14, 9/14
Cross-Attn+2.96%/+1.46%11/14, 13/14

Positive values indicate higher errors relative to Full SCEU. Step conditioning provides the clearest benefit. Cross-attention replacement and removal of signed modulation also increase average errors, while state conditioning provides smaller, dataset-dependent gains. These results directly address the individual design choices and will be included in the revision.

W2. Bounded residuals and extreme events

Thank you for pointing out that aggregate errors can obscure performance on extreme values. We added a quantile-sliced evaluation on BE (H=24H=24) and Weather (H=96H=96). At each forecast step, we grouped test predictions by their true target values into the lowest 5%, middle 90%, and highest 5%. We then compared full SCEU, with the residual bound in Proposition 3, against the endogenous-only base forecast from the same model. The table reports relative error reductions.

Dataset (HH)True target valuesMSE reductionMAE reduction
BE (24)Lowest 5%75.05%56.39%
BE (24)Middle 90%50.34%35.87%
BE (24)Highest 5%14.34%17.63%
Weather (96)Lowest 5%51.21%45.92%
Weather (96)Middle 90%41.78%26.10%
Weather (96)Highest 5%24.18%14.68%

SCEU improves forecasts of both unusually low and unusually high target values while keeping its corrections within the bound in Proposition 3. The largest gains occur for unusually low values, where SCEU reduces MAE by 56.39% on BE and 45.92% on Weather. For unusually high values, SCEU also reduces both MSE and MAE, with MAE reductions of 17.63% on BE and 14.68% on Weather relative to the endogenous-only base forecast.

Q1. Transferability to other backbones

We thank the reviewer for this suggestion. We integrated SCEU's adaptive bounded correction into TimeXer and evaluated it on BE and PJM at a forecast horizon of H=24H=24, using a correction range of [0.1,1.0][0.1, 1.0].

DatasetTimeXer + AdaptiveOriginal TimeXer
BE0.414 / 0.2660.392 / 0.253
PJM0.071 / 0.1600.075 / 0.166

Values are MSE / MAE. Relative to the original TimeXer, the correction reduces MSE and MAE on PJM by 5.3% and 3.6%, respectively, and increases them on BE by 5.6% and 5.1%, respectively.

This experiment demonstrates that adaptive bounded correction can be integrated into TimeXer and can improve its forecasts, as observed on PJM. The different outcomes under the same correction range suggest that the correction range and gate initialization should be tuned for the target backbone and dataset.

Q2. Patch lengths and multi-scale configurations

We appreciate the suggestion. Default patch lengths follow DAG. We additionally evaluate alternative P/Q sets and scale configurations on BE and Energy at H=24,168H=24,168, covering short- and longer-horizon forecasting.

P/Q combinations: Let A=[4,8,24]A=[4,8,24] and B=[8,12,24]B=[8,12,24]. AA shares P/Q sets, whereas AB and BA use distinct sets.

Dataset (H)AA MSE/MAEAB MSE/MAEBA MSE/MAE
BE (24)0.342/0.2260.349/0.2290.341/0.225
BE (168)0.435/0.2740.453/0.2810.456/0.281
Energy (24)0.079/0.2180.082/0.2220.081/0.221
Energy (168)0.112/0.2620.111/0.2600.114/0.265

BB was also evaluated but did not achieve the best result. Neither shared nor distinct P/Q sets consistently dominate.

Scale combinations: K1–K4 contain one to four scales, respectively.

Dataset (H)K1 MSEK2 MSEK3 MSEK4 MSE
BE (24)0.3490.3390.3510.358
BE (168)0.4420.4410.4520.447
Energy (24)0.0800.0790.0770.076
Energy (168)0.1160.1170.1110.112

The results reveal horizon-dependent preferences: BE favors K2 at both horizons, while Energy favors K4 at H=24 and K3 at H=168. Thus, longer horizons do not necessarily require more scales, and increasing K does not guarantee better accuracy. Patch lengths determine temporal granularity, whereas K controls representation diversity and H determines prediction length. Their optimal combinations depend on dataset characteristics and horizon. Since scale count and patch composition vary together, their effects cannot be fully isolated. These experiments complement Section 5.3.2 and Figure 4 and will be clarified in the revision.

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

In this paper, the authors propose a time series forecasting framework called SCEU, which aims to adaptively leverage exogenous information at different forecasting horizons. The method first constructs a linear baseline forecaster that relies solely on historical information from the target series, and then introduces two complementary branches. Temporal Utility extracts historical information from the target series through multi-scale temporal modelling, while Channel Utility assigns relative weights to scale representations of exogenous variables for different forecasting horizons. The outputs of the two branches are fused and used to refine the baseline forecasts through a gating mechanism and a bounded residual correction mechanism. The authors evaluate the proposed method on 13 datasets and compare its performance with that of several existing forecasting models.

Critical strengths

S1: This paper investigates the possibility that exogenous information may have different predictive value at different forecasting steps, which is a meaningful research direction for multi-step time series forecasting. S2: By combining a baseline forecaster that relies solely on historical target-series information with a bounded residual correction mechanism, the model can incorporate additional information without fully replacing the baseline forecasts. The overall design is relatively clear. S3: The paper evaluates the proposed method on 13 datasets and compares it with several existing forecasting methods. The ablation studies and hyperparameter sensitivity analyses also provide some empirical evidence to help understand the contributions of the individual model components.

Critical weaknesses

W1: The paper positions step-wise utility learning as its core contribution, but it does not sufficiently clarify how the proposed method fundamentally differs from existing attention mechanisms or variable selection methods. A more direct comparison with the closest existing approaches is needed, particularly methods that employ multi-scale modelling for time series data. W2: The experiments primarily demonstrate improvements in forecasting performance, but do not sufficiently establish whether the utility scores can reliably identify exogenous variables that provide genuine additional predictive information at specific forecasting steps. W3: The method assumes that future exogenous variable trajectories are available at the time of forecasting. However, the actual future values of variables such as traffic occupancy, electricity consumption, and weather observations are generally unavailable at the forecast origin. Since these are commonly used datasets in time series forecasting, the authors should explicitly clarify which future covariates are genuinely known in each dataset and distinguish forecasting based on known future covariates from evaluation using realised future observations.

Questions for authors

Q1: What is the fundamental difference between the proposed step-wise utility weights and the attention weights or variable selection weights used in existing methods for forecasting with exogenous variables? Can the authors validate the practical benefits of the proposed mechanism in isolation through controlled comparative experiments? Q2: Given that learned utility scores are not guaranteed to reflect incremental predictive value, what evidence do the authors provide to demonstrate that exogenous representations assigned higher weights actually yield greater forecasting improvements at the corresponding prediction steps? Q3: For each dataset, which future exogenous variables are genuinely available at the forecast origin? How does the model's performance change when the future exogenous inputs are replaced with forecasts that contain realistic prediction errors? Q4: The proposed method employs multi-scale processing. How are the portions of the historical sequence that cannot form complete patches handled? Q5: The paper uses datasets commonly adopted in time series forecasting, and the performance differences on some datasets appear relatively small. For the datasets in Table 1 where the performance gaps are marginal, what would be a more appropriate way to demonstrate the advantages of the proposed method?

1 comment

W1 & Q1. Novelty and utility versus attention

We thank the reviewer. We agree that multi-scale modeling itself is not novel. We therefore compare SCEU with TimeKAN, a multi-scale forecasting model, under the same protocol (MSE/MAE):

DatasetSCEUTimeKAN
Weather0.001/0.0210.007/0.067
Electricity0.193/0.3210.301/0.429
Traffic0.090/0.1660.211/0.285
ETTh10.104/0.2520.172/0.331
ETTh20.327/0.4420.301/0.432

Across 13 datasets, SCEU wins on both metrics in 12/13. Its contribution lies in step-conditioned utility learning and selective correction of an endogenous-only forecast, rather than multi-scale representations alone.

For Q1, we replace utility scoring with standard cross-attention while retaining the remaining framework:

VariantAvg. MSE/MAE increaseFull wins (MSE/MAE)
Cross-Attn+2.96%/+1.46%11/14, 13/14

Across 14 settings, utility scoring outperforms cross-attention in most cases. Unlike standard attention, it combines relative importance with signed modulation to regulate bounded corrections. We will clarify this distinction without claiming that attention components themselves are novel.

W2 & Q2. Step-dependent utility and incremental value

Thank you for raising this question. We evaluated SCEU on BE (H = 24) and Weather (H = 96) by masking one exogenous variable at a time during testing. Using the same trained model, we measured the change in absolute prediction error between the original and masked inputs. We then grouped the masking results into four groups based on the variable’s Channel Utility weight at the corresponding forecast step, computed from the original, unmasked input.

Average increase in absolute prediction error after masking. Larger values indicate a greater loss in forecast accuracy.

DatasetForecast horizonLowest 25% of weights25–50%50–75%Highest 25% of weights
BE240.0470.0670.0790.087
Weather960.0000340.0001110.0002090.000296

On both datasets, the average error increase rose from one weight group to the next. Masking variables in the highest-weight group increased absolute error 1.86 times as much on BE and 8.50 times as much on Weather as masking variables in the lowest-weight group. The pooled Spearman correlations were 0.059 on BE and 0.075 on Weather, indicating a weak association across individual masking results. Under this masking protocol, SCEU was therefore more sensitive, on average, to masking variables that it assigned higher weights.

W3 & Q3. Future covariates and noise

We thank the reviewer. Benchmark-provided future covariates are not always known at forecast origin. Calendar variables may be available in advance, whereas future weather, traffic, and electricity measurements may require external forecasts. Our original masking experiments address availability, not forecast errors.

We therefore perturb only future exogenous inputs using Gaussian noise, ϵ∼N(0,σ2I)\epsilon\sim\mathcal N(0,\sigma^2I), at σ∈{0,0.1,0.2,0.5}\sigma\in\{0,0.1,0.2,0.5\}, with three realizations per nonzero level. Historical inputs remain unchanged.

DE results (MSE/MAE):

σ\sigmaSCEUDAG
00.271/0.3220.294/0.329
0.10.281/0.3300.312/0.344
0.20.302/0.3450.349/0.371
0.50.389/0.4020.454/0.436

Across seven datasets, the comparison is:

σ\sigmaMSE/MAE winsAvg. MSE/MAE reduction
06/7, 5/77.98%/3.25%
0.16/7, 4/78.81%/4.12%
0.27/7, 6/710.18%/5.68%
0.56/7, 6/710.79%/7.40%

The reductions are equal-weight averages of per-dataset relative improvements against DAG. At σ=0.5\sigma=0.5, SCEU also shows smaller average relative degradation from clean inputs: 30.25% versus 34.89% MSE, and 22.46% versus 28.14% MAE.

These results support average robustness, not universal superiority. Gaussian noise remains a proxy for forecast errors.

Q4. Incomplete patches

We thank the reviewer. Temporal Utility retains ⌊T/Pr⌋\lfloor T/P_r\rfloor complete non-overlapping patches and discards trailing incomplete observations. Channel Utility instead uses last-value padding at the right boundary. We will clarify both implementations in the revision.

Q5. Marginal improvements

We appreciate the suggestion. SCEU does not outperform every baseline on every dataset, and small differences should not be overstated. Its main contribution is a step-conditioned perspective for selective exogenous integration, supported by competitive results on most datasets. We will discuss small gains cautiously.

Reviewer_2j2GOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SCEU, a forecasting model that estimates how useful exogenous variables are at different forecast steps, conditioned on the target’s recent state. It combines this step-conditioned utility with multiscale temporal and cross-variable representations, then applies a bounded gated residual correction to an endogenous-only forecast. The authors evaluate it on 13 time-series datasets, compare it with forecasting baselines, and report ablations and robustness analyses. Results show gains on several datasets, though the experiments do not fully isolate the benefit of step-conditioned exogenous utility from the other components.

Critical strengths

1,The paper studies a practically relevant forecasting problem: the usefulness of exogenous variables can vary with the target’s recent state and across forecast steps. SCEU offers a coherent way to incorporate this variation while retaining an endogenous-only forecast as a stable base, making the overall modeling approach meaningful beyond a single fixed variable-weighting scheme.

2,The evaluation spans 13 datasets and includes comparisons with recent forecasting methods, ablations, and robustness analyses. The reported gains on several datasets suggest that the approach can be effective across more than one setting, which supports the value of the proposed overall architecture.

Critical weaknesses

The experiments do not clearly isolate the paper’s central claim: that exogenous-variable utility should be conditioned on the target state and vary across forecast steps. The ablations remove whole utility branches, changing the model’s information paths and capacity, but do not compare against a matched model that retains the same inputs and architecture while sharing utility weights across forecast steps. As a result, the reported gains support the effectiveness of the overall architecture more strongly than the claimed importance of step-conditioned utility, leaving the paper’s main contribution insufficiently established.

The method assumes access to future exogenous trajectories, but this assumption may not hold for several benchmark variables at the forecast origin. The paper’s random-masking experiments test reduced input availability, but do not establish performance when future exogenous inputs are themselves forecasts or are noisy. This limits how confidently the results support the paper’s practical forecasting claims.

Questions for authors

1,Clarification: Can the authors specify which parts of SCEU are intended to constitute the main novelty relative to existing horizon-dependent variable selection or attention methods—target-state conditioning, joint variable–scale weighting, signed modulation, or the bounded residual design? What forecasting difficulty does this address that those methods do not?

2,Clarification: Can the authors provide a matched comparison that keeps the same inputs, encoders, and prediction head but shares the exogenous utility weights across forecast steps? This would help determine whether step-conditioned utility itself accounts for the gains, rather than the additional branches or model capacity.

3,Clarification: For the benchmarks, which future exogenous variables are assumed to be known at the forecast origin, and which are observed or forecast proxies? Does the reported advantage persist when future exogenous inputs are supplied as noisy forecasts rather than ground-truth trajectories?

1 comment

Q1. Clarification of the main novelty

We thank the reviewer for this important question.

The central contribution of SCEU lies in step-conditioned exogenous predictive utility learning, rather than any individual component in isolation.

Specifically, SCEU jointly considers the current target state and forecast position to estimate the relative predictive utility of exogenous representations. This is realized through three complementary mechanisms:

  • Joint variable–scale weighting: Evaluates exogenous representations across variables and temporal scales, allowing their contributions to vary across forecast steps.
  • Signed modulation: Allows positive or negative modulation rather than restricting utility to nonnegative importance weights.
  • Bounded residual correction: Retains an endogenous-only base forecast and selectively constrains exogenous corrections.

Unlike TFT's variable selection, TiDE's covariate integration, or DAG's correlation transfer, SCEU explicitly learns step-dependent relative utility and uses it to regulate bounded forecast corrections.

This addresses the difficulty that exogenous information may contribute differently across forecast positions under different target states. We will clarify these distinctions without attributing novelty to dynamic weighting alone.

W1&Q2. Matched comparison with Shared Utility

We appreciate the reviewer for identifying this experimental gap.

To isolate the contribution of step conditioning, we introduce Shared Utility, which shares utility scores across forecast steps while retaining the remaining architecture.

The following table compares Full SCEU with Shared Utility (MSE/MAE):

Dataset (H)FullShared Utility
BE (24)0.351/0.2280.357/0.235
DE (24)0.271/0.3220.304/0.339
Energy (24)0.077/0.2130.077/0.217
FR (24)0.342/0.1810.347/0.188
NP (24)0.187/0.2160.206/0.231
PJM (24)0.058/0.1460.063/0.152
Weather (96, x10)0.007/0.1760.007/0.177

Across all 14 dataset–horizon settings, Full achieves lower MSE in 12/14 and MAE in 13/14 cases. Sharing utility scores increases average MSE by 3.28% and MAE by 2.42%.

These results support the predictive benefit of forecast-step-specific utility over shared weighting. We will include this controlled ablation in the revision.

W2&Q3. Future exogenous availability and noise robustness

We thank the reviewer for raising this practical concern.

Future exogenous inputs. Our experiments use benchmark-provided future covariates, which should not be interpreted as universally known at the forecast origin. Calendar features may be available in advance, whereas future weather or traffic measurements may require external forecasts.

Our original missing-exogenous experiments primarily evaluate whether SCEU maintains predictive capability when future covariates are partially or entirely unavailable.

Additional Gaussian noise experiments. To investigate imperfect future covariates, we conducted new experiments comparing SCEU with DAG on seven datasets. Gaussian noise is added to future exogenous inputs:

\quad \boldsymbol{\epsilon}\sim\mathcal N(0,\sigma^2\mathbf I).$$ We evaluate $\sigma\in\{0,0.1,0.2,0.5\}$ with three noise realizations per nonzero level. Representative results at $\sigma=0$ and $\sigma=0.5$ are shown below. | Dataset | $\sigma$ | SCEU MSE | DAG MSE | SCEU MAE | DAG MAE | |---|---:|---:|---:|---:|---:| | NP | 0 | **0.187** | 0.203 | **0.216** | 0.236 | | NP | 0.5 | **0.253** | 0.264 | **0.277** | 0.304 | | DE | 0 | **0.271** | 0.294 | **0.322** | 0.330 | | DE | 0.5 | **0.390** | 0.454 | **0.402** | 0.436 | | FR | 0 | **0.342** | 0.355 | 0.182 | **0.172** | | FR | 0.5 | **0.377** | 0.415 | **0.224** | 0.245 | Across all seven datasets, SCEU achieves lower average errors than DAG at every evaluated noise level. At $\sigma=0.5$, SCEU reduces MSE and MAE by an average of **10.79%** and **7.40%**, respectively. These results complement our missing-exogenous experiments and provide further evidence of SCEU's robustness under imperfect future exogenous inputs. We will clarify the benchmark assumptions and incorporate these evaluations in the revision.
Reviewer_An4yOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This article proposes a prediction model SCEU that can adjust the use of exogenous information based on the prediction steps. This model combines only endogenous linear basis prediction with two branches, one branch is the multi-scale objective historical representation of time utility aggregation, and the other branch is the channel utility scale representation of exogenous variable allocation with signed, step based weights. Their outputs are fused through residual correction. This article evaluated SCEU on 13 real-world datasets and validated its effectiveness.

Critical strengths

S1. Exogenous variables can have different predictive relevance across forecast steps. Conditioning their use on both the target’s current state and the forecast position is a sensible modeling choice.

S2. Retaining an endogenous-only base forecast and controlling the correction magnitude offers a clear structural design.

S3. The empirical evaluation covers a broad range of benchmarks.

Critical weaknesses

W1. The problem setting largely follows DAG, while the distinction from existing methods such as TFT and TiDE is insufficiently clear. The contribution appears primarily incremental, adapting established time-series modeling components to a previously studied task.

W2. Some borrowed baseline results differ from those reported in DAG, while the changed ETT protocol complicates comparison with previously published results. The omission of GCGNet further weakens the evaluation. Taken together, the evidence for a substantial and consistent improvement over DAG is not yet convincing.

W3. The missing-data experiments use random masking and retrain the model at each missingness level. This setting does not adequately reflect noisy forecasts or anomalous exogenous inputs encountered in practice, leaving it unclear whether SCEU maintains its advantage under realistic conditions.

Questions for authors

See weaknesses.

3 comments

W1. What is new, and how it differs from DAG, TFT and TiDE

A new question. Existing exogenous forecasters model how strongly a covariate relates to the target. SCEU asks: at one forecast origin and one system state, which exogenous information should correct the forecast at each individual future step, and in which direction? SCEU is the first method to make this per-step decision the object of learning: for every step hh it learns a signed utility over all exogenous variable–scale representations (Eq. 15–17) and uses it only to correct an endogenous-only forecast (Eq. 24).

Why it is hard. (1) Utility is never observed; it must be learned from the forecasting loss alone. (2) Dependence is not utility: a covariate correlated with the target can add nothing beyond the target's history, and correlation or attention similarity cannot separate the two. (3) Step-wise weighting needs H×C×KH\times C\times K weights per forecast, HH times more than horizon-invariant weighting; unconstrained, irrelevant covariates corrupt the forecast.

How SCEU solves it. (1) Residual formulation: covariates only correct an endogenous-only base forecast, so utility is learned relative to what the target's history already predicts. (2) Signed utility: Softmax importance × tanh direction lets one representation raise the forecast at one step and lower it at another; convex attention weights cannot. (3) Guarantees for every input: utility mass ≤1\le 1 (Prop. 1), fusion never amplifies the context norm (Prop. 2), and the forecast stays within rmax⁡r_{\max} of the endogenous base (Prop. 3).

Core objectCovariates enterBound
TFTnon-negative selection weights per input time step, from that step's covariate valuesmain pathwaynone
TiDEshared covariate projection, no selectionmain pathwaynone
DAGtemporal/channel correlations from auxiliary prediction taskstransferred into the forecastnone
SCEUsigned utility per target step × variable × scalecorrection to an endogenous-only forecast≤rmax⁡\le r_{\max}

None of the three ties a signed weight to the target step or bounds the effect of covariates. SCEU does both.

Evidence. With everything else fixed, sharing one utility across forecast steps raises MSE on 10 of 12 settings, and replacing utility scoring by standard cross-attention raises it on 9 of 12 (BE/DE/Energy/FR/NP/PJM, H=24H=24 and 168168). Masking a covariate at test time raises the error 1.86× (BE) and 8.50× (Weather) more when its utility weight at that step is in the highest quartile than in the lowest.

W2. Baseline numbers, ETT protocol, GCGNet

Baseline numbers. We reran every baseline, including DAG, on one server with the hyperparameters DAG reports for each baseline. Reproduced scores differ from DAG's published values by about 0.0010.001, the size of numerical differences between computing environments. The caption of Table 5 says the numbers are taken from the DAG paper; this is corrected in the revision.

ETT. We use the full-length ETT series (17,420 hourly points; the common truncation keeps 14,400), which lengthens the test span by 21%. All models are trained and evaluated by us on the same chronological split, so Table 4 is like-for-like.

GCGNet (official code, our protocol; average MSE/MAE):

DatasetSCEUGCGNetDatasetSCEUGCGNet
Weather0.001/0.0210.002/0.030BE0.420/0.2740.431/0.294
Electricity0.192/0.3210.376/0.467DE0.355/0.3700.402/0.389
Traffic0.090/0.1660.384/0.469Energy0.122/0.2660.132/0.278
ETTh10.104/0.2520.133/0.284FR0.422/0.2320.415/0.234
ETTh20.327/0.4420.368/0.466NP0.322/0.3190.370/0.348
ETTm10.080/0.2080.091/0.221PJM0.091/0.1820.095/0.187
ETTm20.225/0.3400.236/0.352

SCEU has lower MSE on 12/13 datasets (all but FR) and lower MAE on all 13.

W3. Noisy future exogenous inputs

Each model is trained once on clean inputs; noise ϵ∼N(0,σ2I)\epsilon\sim\mathcal N(0,\sigma^2 I) (normalized units) is added to future exogenous inputs only at test time, without retraining, three draws per level. SCEU vs. DAG on seven settings (BE/DE/Energy/FR/NP/PJM at H=24H=24, Weather at H=96H=96):

σ\sigmaSCEU lower (MSE, MAE)Avg. reduction vs. DAG (MSE / MAE)Avg. MSE rise, SCEU / DAG
06/7, 5/77.98% / 3.25%–
0.16/7, 4/78.81% / 4.12%2.3% / 3.4%
0.27/7, 6/710.18% / 5.68%7.1% / 10.2%
0.56/7, 6/710.79% / 7.40%30.3% / 34.9%

SCEU's advantage over DAG holds at every noise level and widens as the covariates degrade; SCEU also degrades less than DAG at every level. On BE/DE/FR/NP/PJM the exogenous inputs are themselves day-ahead load and generation forecasts (Lago et al., 2021), so the σ=0\sigma=0 row already uses real forecasts with real errors.

Reviewer_An4yOpenReview ↗

Response

Thank you for clarifying the contribution. I see value in learning how exogenous information should adjust a forecast at each future step. My main reservation concerns the novelty of both the research question and the method. Several of the ideas the authors highlight, particularly residual correction, are familiar from the broader time-series forecasting literature. My impression is that the main contribution lies in adapting these established ideas to a more specialized forecasting task.

I therefore maintain my score of 2. My confidence is also 2, as my limited familiarity with this specific task makes it difficult for me to judge how novel the formulation is within that literature.

Thank you for the follow-up. Residual correction is not our core innovation. SCEU’s core methodological contribution is how it constructs the correction: it learns a signed weight for each forecast step × covariate × temporal scale combination, conditioned on the recent target state, and uses the resulting exogenous representation in a bounded forecast correction.

1. Residual correction is an established design that provides the reference forecast

TiDE already adds a linear projection of the target history to a nonlinear forecast that uses future covariates. SCEU adopts this established structure to provide a target-only reference (Eq. 4). The gate and adaptive range then limit how far the fused correction can move the prediction from that reference at each step. In normalized space, ∣y^hnorm−y^hbase∣≤ghrh≤rmax⁡|\hat y_h^{\mathrm{norm}}-\hat y_h^{\mathrm{base}}|\le g_h r_h\le r_{\max} (Eqs. 22–24 and Proposition 3).

2. SCEU’s core innovation is explicit signed allocation at each forecast step

Within the established exogenous forecasting task, we ask a specific modeling question: at the same forecast origin and under the same target state, how should the model weight each covariate at each temporal scale for each future step?

MethodHow it incorporates exogenous information
TiDEFeeds each step’s projected covariates and decoded representation into a shared temporal decoder.
TFTSelects among known future variables at each input time using non-negative weights before temporal processing.
DAGLearns temporal and channel correlations through auxiliary prediction tasks and transfers the learned attention parameters into forecasting.
SCEUComputes an explicit, state-conditioned signed coefficient for each output step, covariate, and scale to construct the exogenous context used in the bounded correction.

SCEU computes uh,c,kex=ah,c,kexdh,c,kexu_{h,c,k}^{\mathrm{ex}}=a_{h,c,k}^{\mathrm{ex}}d_{h,c,k}^{\mathrm{ex}} (Eqs. 15–18). The softmax term assigns relative importance across covariates and scales, and the separate signed term modulates how each latent representation enters the aggregate. Even when two forecast steps share the same future covariate patch, their distinct step embeddings allow the scorer to assign different coefficients. The model thus has separate mechanisms for allocating exogenous information and controlling the magnitude of the final correction.

3. The ablations test the contribution of this mechanism within the residual framework

The variants below retain the target-only base pathway and change the allocation or correction mechanism.

ModificationObserved change in MSE
Share one allocation across forecast stepsHigher in 10 of 12 settings across BE, DE, Energy, FR, NP, and PJM at two horizons each
Replace the proposed scoring with standard cross-attentionHigher in 9 of the same 12 settings
Remove Channel Utility (Table 2)Electricity (H=96H=96): 0.161 → 0.187. NP (H=24H=24): 0.187 → 0.274.
Remove Adaptive Response Range (Table 2)Electricity (H=96H=96): 0.161 → 0.220. NP (H=24H=24): 0.187 → 0.221.

The first two ablations test the allocation mechanism with the residual pathway retained. Sharing one allocation across steps increases MSE in 10 of 12 settings, supporting the use of step-specific weights. Replacing the proposed scorer with standard cross-attention increases MSE in 9 of 12 settings, supporting the proposed scoring mechanism relative to this alternative. We will clarify these architectural differences in Sections 1 and 4 and document what each ablation changes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.