acceptodds
Under review as a conference paper at ICLR 2027

Evidence-Guided Action World Model for Time-Series Forecasting

Abstract

Time-series forecasting aims to predict future system states from historical observations and supports decision-making in applications such as energy management and retail planning. While modern forecasting methods effectively model endogenous temporal dynamics, real-world systems are also influenced by external factors that can alter their future evolution. Existing approaches typically incorporate such information as additional predictive inputs, directly learning mappings from external evidence to future observations. However, this treatment leaves the underlying dynamic response implicit. Importantly, the same external evidence may induce different responses depending on the current system state, making it difficult for existing methods to fully exploit external evidence. We therefore argue that forecasting should explicitly model the dynamic response implied by external evidence and characterize how this response changes system dynamics. Motivated by this perspective, we formulate forecasting under external evidence as an action-conditioned world-model problem, where latent actions explicitly represent dynamic responses to external influences and modulate future state transitions. We propose the Evidence-Guided Action World Model (EGA-WM), which learns a latent action space through a primitive bank, infers state-dependent actions from external evidence, and incorporates them into latent dynamics while modeling their persistent effects. Experiments on electricity-market and retail datasets demonstrate the effectiveness of EGA-WM across diverse forecasting scenarios.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series forecasting, Exogenous variables, World models, Action-conditioned dynamics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_99pAOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes an interesting framework called EGA-WM for time-series forecasting with external evidence. The framework includes three steps to solve how to model actions, how to activate the actions, and how to leverage the actions to help time-series forecasting. The results on five electricity markets and two retail datasets support the usefulness of this approach.

Critical strengths
  1. The motivation is interesting and easy to follow, i.e., the same external event can lead to different responses under different system conditions. This gives the proposed formulation a clear purpose.
  2. The technical contribution is good. It separates action activation from the persistence of its effects, which is a useful design choice.
  3. The experiments are comprehensive, covering electricity and retail forecasting datasets, with different types of external evidence.
Critical weaknesses
  1. The authors used ‘action’ as the core contribution. I think it is necessary to clearly define the used ‘action’, and how the ‘action’ is related to the external evidence.
  2. I notice the model used two latent trajectories to compute the forecast adjustment. I suggest the authors to provide the clearer rationale of why using the two trajectories.
  3. The temporal activation section introduces onset, duration, strength, and a stepwise mask together. I suggest the authors to provide a short example that would help explain how these terms combine into one activation profile.
  4. Figure 4 is helpful, but it is unclear how temporal activation is set when each primitive is activated separately. I think the authors could provide a brief explanation and that would make the curves easier to interpret.
  5. The authors mentioned these datasets can be divided into forecast-based evidence and event-based evidence. However, which dataset belongs to the former and which belongs to the later? and what are their differences?
Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 99pA

Thank you for the positive assessment and for recognizing our motivation, the separation between action activation and the persistence of its effects, and the comprehensive experiments. We apologize for the delayed response, as we conducted a substantial number of additional experiments to address the reviewers' concerns and further strengthen the paper. We have addressed your comments point by point below, and the revised manuscript incorporating these clarifications will be provided shortly.

W1: Definition of action and its relationship to external evidence.

In EGA-WM, an action represents a learned modification of the latent state transition, inferred from the historical state and external evidence. Formally, it is represented as a=(π,α,β,s,m)a=(\pi,\alpha,\beta,s,m), where π\pi determines the combination of action primitives, while the remaining parameters control when and how strongly the action affects the state transition.

Importantly, external evidence serves as the input, whereas the action represents the response inferred from that evidence under the current system state. Thus, the same external evidence may lead to different inferred actions depending on the historical state. These actions are learned predictive representations rather than observed event labels or executable controls. We will clarify this definition and its relationship to external evidence in Secs. 3.3 and 3.4.

W2: Rationale for using two latent trajectories.

The two latent trajectories are designed to isolate the effect of action modulation while ensuring that the correction vanishes when no action is applied. Specifically, as defined in Eq. (10),

Y^=b+D(za)−D(z0),\hat{Y}=b+D(z^a)-D(z^0),

both trajectories share the same initialization and baseline inputs, and are processed by the same decoder. Their only difference lies in whether action modulation is applied. The correction D(za)−D(z0)D(z^a)-D(z^0) therefore captures the change in the decoded prediction associated with action modulation.

Importantly, when no action is applied, za=z0z^a=z^0, making the correction exactly zero and recovering the original direct forecast bb. In contrast, directly adding D(za)D(z^a) would alter bb even without action modulation, since D(z0)D(z^0) is generally non-zero. We will clarify this rationale in Sec. 3.5 and refer to the corresponding proof in Appendix C.

W3: How temporal parameters form the activation profile.

The temporal activation combines four components: onset determines when the action begins, duration controls how long it remains active, strength determines its overall magnitude, and the mask adjusts the activation at individual forecast steps.

For example, consider a forecast horizon of H=6H=6, where onset is concentrated at step 3, duration is concentrated at 3 steps, strength is s=2s=2, and the mask is mh=0.5m_h=0.5 for all steps. Under this simplified setting, Eq. (6) yields

g=(0,0,1,1,1,0).g=(0,0,1,1,1,0).

The action is therefore activated during steps 3–5, with an amplitude of 2×0.5=12\times0.5=1. This example uses concentrated onset and duration distributions for clarity, while the actual model learns soft distributions that allow more flexible activation profiles. Moreover, the action's effect may persist after activation through the latent state retention mechanism. We will include this illustrative example in Sec. 3.4.

W4: Temporal activation of individual primitives in Fig. 4.

In Fig. 4, all primitives are evaluated using the same inferred temporal activation profile. To visualize each primitive separately, we replace the primitive mixture weights with a one-hot vector while keeping the historical state, onset, duration, strength, mask, and frozen direct forecast unchanged. Each curve represents the resulting prediction adjustment relative to the direct forecast. We will clarify these visualization settings in the caption of Fig. 4.

W5: Distinction between the two types of external evidence.

Our experiments consider two types of external evidence: forecast-based evidence and event-based evidence. Their dataset assignments and examples are summarized below.

DatasetEvidence typeExamples
EPFForecast-basedDay-ahead forecasts of load, generation or wind
RossmannEvent-basedPromotions, opening status and holidays
M5Event-based, including numerical covariatesCalendar events, SNAP and weekly prices

The key distinction lies in how the future evidence is obtained. Forecast-based evidence consists of predictions of external variables, such as future electricity load or wind generation. In contrast, event-based evidence provides records or known schedules of external events, including numerical variables such as prices.

Both types are provided to EGA-WM as numerical inputs under the benchmark assumption that the corresponding future covariates are available at the forecast cutoff. We will explicitly clarify this distinction and the dataset assignments in Sec. 4.1.

Dear Reviewer 99pA,

Thank you again for your detailed and constructive comments. We have uploaded the revised manuscript, which incorporates the changes made in response to your suggestions and those of the other reviewers to improve the clarity and rigor of our work. We would be grateful if you could let us know whether our responses and revisions have addressed your concerns, and we will be glad to address any further feedback.

The authors have addressed my concerns, so I have raised my score.

Reviewer_bwC7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies time-series forecasting with external evidence and proposes the Evidence-Guided Action World Model (EGA-WM). The central motivation is that the same external evidence may induce different future responses depending on the current system state. EGA-WM introduces a learnable bank of latent action primitives, infers a state-conditioned mixture of these primitives together with temporal activation parameters, and uses the inferred action to modulate both the input and retention terms of a latent recurrent transition model. The resulting action-dependent forecast adjustment is added to the output of a pretrained and frozen evidence-aware forecaster.

The method is evaluated on five electricity markets from EPF and two retail datasets, Rossmann and M5. EGA-WM obtains the best aggregate results in Table 1, and the component ablations generally support the use of the primitive bank, state/evidence conditioning, temporal activation, state propagation, and retention modulation.

Critical strengths
  1. The paper targets an important and practically relevant setting. Modeling how known external conditions alter future dynamics is well motivated for electricity-price and retail-demand forecasting.

  2. The proposed architecture is coherent. The primitive bank, temporal activation, input modulation, retention modulation, and recurrent propagation have clearly defined roles.

  3. The empirical results are promising. On EPF, EGA-WM achieves the lowest MSE in four of five individual markets, suggesting that the macro-average improvement is not driven by a single market. The improvement on Rossmann is also substantial.

Critical weaknesses
  1. The direct forecasting baseline is not consistently reported, making the source of the improvement unclear.

    Appendix B states that BdirB_{\mathrm{dir}} is a pretrained evidence-aware Sonnet model, while Table 2 defines “Baseline” as the evidence-informed direct forecast without the action branch. Nevertheless, Table 1 and Table 2 report substantially different Sonnet/direct-baseline results. For example, the Rossmann MSE is 0.0194 for Sonnet in Table 1 and 0.0105 for the Baseline in Table 2; on M5, the corresponding values are 1.5246 and 1.6513.

    This matters because the improvement on Rossmann is 56.7% relative to the Table 1 Sonnet result but 20.0% relative to the direct baseline actually used by EGA-WM. On M5, the full model improves by 7.8% over the Table 2 baseline but only 0.17% over the Table 1 Sonnet result. Without explaining this discrepancy, it is difficult to determine how much of the reported gain is attributable to the proposed action module.

  2. The experiments do not isolate the value of the proposed action formulation from that of an additional conditional residual model.

    The primitive mixture and activation mechanism ultimately produce a state- and evidence-dependent modification of a latent recurrence. A generic residual MLP, a conditionally parameterized recurrent model, or a matched mixture-of-experts correction could potentially provide similar gains. The existing “w/o primitive bank” and component ablations are useful, but they do not constitute a capacity-matched comparison against simpler alternative correction architectures using the same encoders, inputs, and frozen direct forecaster.

    Consequently, the current experiments show that the complete architecture is effective, but do not yet establish that framing the correction as an action-conditioned world model provides a distinct advantage.

  3. The empirical evaluation does not directly test the central state-dependent-response claim, and the external-evidence settings are relatively limited.

    The motivating claim is that identical or similar evidence should lead to different responses under different system states. The evaluation reports aggregate forecasting errors but does not stratify examples by state, compare matched evidence under different states, or otherwise show that EGA-WM is particularly effective in the setting highlighted by the motivation.

    In addition, EPF and Rossmann have only 2 and 8 evidence channels. M5 has 39 encoded channels, but the paper does not demonstrate performance under many independent external drivers, redundant evidence, or multiple distinct events within the same prediction horizon. This is relevant because EGA-WM predicts one primitive mixture for the entire horizon from a fixed initial-state context. Its ability to handle several qualitatively different events or state changes within one horizon remains unclear.

  4. The statistical evidence is insufficient for some of the reported improvements.

    Table 1 reports averages over three seeds but provides neither standard deviations nor confidence intervals. This is particularly important for M5: relative to the best reported competing method, EGA-WM improves MSE by approximately 0.17%, MAE by 0.14%, and MASE by 0.18%. These differences may be meaningful, but the reported means alone do not establish that they exceed training or evaluation variability.

    The results are also not uniformly better at the market level. For example, on Belgium, the MSE increases from 0.4207 for DAG to 0.4469 for EGA-WM. A more calibrated conclusion should acknowledge this heterogeneity.

Questions for authors
  1. Why do the Sonnet results in Table 1 differ from the direct forecasting Baseline in Table 2 if BdirB_{\mathrm{dir}} is an evidence-aware Sonnet model? Were the architecture, preprocessing, training budget, hyperparameter selection, and seed sets identical? Reporting the performance of the exact frozen BdirB_{\mathrm{dir}} used by every EGA-WM run would resolve this concern.

  2. How does EGA-WM compare with a parameter-matched generic residual network or conditional recurrent correction model that receives the same state and evidence encodings and uses the same frozen direct forecaster? A favorable result under this controlled comparison would materially strengthen the claimed contribution.

  3. Can the authors report per-seed results or suitable uncertainty estimates, particularly for M5? Is the sub-0.2% improvement over the strongest Table 1 baselines statistically distinguishable from the relevant training and test-period variability?

  4. Can the authors directly evaluate the main state-dependent-response hypothesis, for example by grouping or matching windows with similar external evidence but different history states? It would be useful to determine whether EGA-WM’s relative gain is larger precisely when the response to evidence varies across states.

  5. How does the model behave when multiple qualitatively different external events occur within the same forecast horizon? Since one primitive mixture is inferred for the whole horizon from a fixed initial context, a focused analysis of such cases would clarify the scope of the proposed formulation.

7 comments
AuthorsOpenReview ↗

Response to Reviewer bwC7 (1/4)

Thank you for the careful review and for recognizing the practical relevance of our setting, the coherence of the proposed architecture, and the promising empirical results. We apologize for the delayed response, as we conducted a substantial number of additional experiments to address the concerns and further strengthen the paper. We have addressed your concerns point by point below, and the revised manuscript incorporating these changes will be provided shortly. Results are means ± SD over three seeds, with EPF averaged over five markets.

W1 & Q1: Clarification of the Sonnet backbone configurations and baseline differences.

The baseline in Table 2 uses Sonnet as its backbone but differs in configuration from the standalone Sonnet reported in Table 1. Specifically, the backbone configuration in Table 2 is selected based on the validation performance of the complete EGA-WM framework and subsequently kept fixed when disabling the action branch for ablation. In contrast, Table 1 uses a configuration independently selected based on the validation performance of standalone Sonnet forecasting. Both settings use the same Sonnet architecture family, data splits, preprocessing, and random seeds (2026/2027/2028), but their hyperparameters are selected separately, resulting in different pretrained checkpoints. The detailed differences are summarized below (Table 1 Sonnet → EGA-WM backbone):

Datasetd_modelSonnet αDownsampling scalesLearning rateParameters
EPF64 → 640.50 → 0.25[1,2,4] → [1,2,3]3e-4 → 1e-3195,539 → 195,491
Rossmann32 → 440.50 → 0.75[1,2] → [1,2]1e-3 → 5e-4140,523 → 154,290
M564 → 560.25 → 0.75[1,2,4] → [1,2,4]3e-4 → 1e-3182,375 → 163,193

Here, α denotes a Sonnet-specific hyperparameter, distinct from the onset parameter in EGA-WM. Parameter counts refer only to the direct forecasting backbone (per market for EPF), excluding the history forecaster and correction module. These differences in configurations and pretrained checkpoints explain the discrepancy between the two baselines. Since validation-based selection does not guarantee the best test performance, neither configuration consistently outperforms the other on the test set: the configuration in Table 2 achieves lower MSE on Rossmann (0.0105 vs. 0.0194), whereas the standalone configuration achieves lower MSE on M5 (1.5246 vs. 1.6513). We will clarify this distinction in the main manuscript and provide the detailed configurations in Appendix B.

Importantly, the Baseline in Table 2 is evaluated using the exact frozen direct forecaster employed in each corresponding EGA-WM run, rather than a separately trained baseline. Thus, Table 2 isolates the contribution of the action branch under identical backbone settings and checkpoints. Adding the action branch reduces MSE by 13.0%, 20.0%, and 7.8% on EPF, Rossmann, and M5, respectively. In particular, the 20.0% reduction on Rossmann is achieved over an already stronger configured baseline, indicating that the improvement cannot be attributed solely to backbone configuration.

To further disentangle the contribution of EGA-WM from that of backbone configuration, we conducted an additional experiment using the exact Sonnet model reported in Table 1. We kept the original Sonnet frozen and added our action-based correction module without modifying or retuning its backbone configuration.

ModelEPF MSERossmann MSEM5 MSE
Original Sonnet0.2416 ± 0.00350.0194 ± 0.00961.5246 ± 0.0045
Same Sonnet + EGA-WM0.2173 ± 0.00360.0128 ± 0.00521.5250 ± 0.0038

The correction improves all three metrics on EPF and Rossmann. On M5, although MSE remains nearly unchanged, MAE and MASE improve from 0.9008/1.6350 to 0.8878/1.6140, respectively. These results provide further evidence that the gains are not merely due to backbone configuration but also arise from the proposed action-based correction mechanism. We will include this controlled comparison in the revised manuscript to clarify the contribution of EGA-WM.

AuthorsOpenReview ↗

Response to Reviewer bwC7 (2/4)

W2 & Q2: Contribution of the action formulation beyond generic residual correction.

To examine whether the action formulation provides a distinct advantage over a generic learned correction, we conducted additional experiments with parameter-matched MLP and GRU residual correctors using the same frozen backbones. All methods receive the same state and evidence encodings, use the same encoder architectures, and follow the same additional training stage with an identical training budget.

CorrectionEPF MSERossmann MSEM5 MSE
Residual MLP0.2352 ± 0.00310.0094 ± 0.00481.5721 ± 0.0039
Residual GRU0.2333 ± 0.00330.0090 ± 0.00471.5824 ± 0.0042
EGA-WM0.2087 ± 0.00320.0084 ± 0.00291.5220 ± 0.0027

The compared correction modules are matched in parameter count (83,977/33,076/28,069 on EPF/Rossmann/M5, respectively), with the shared encoders also included and trained under the same settings.

EGA-WM achieves the lowest MSE across all three datasets, reducing MSE by 10.5%, 6.7%, and 3.2% relative to the stronger residual corrector on EPF, Rossmann, and M5, respectively. It also achieves lower MAE and MASE than both alternatives across all datasets (full results will be provided in the revised manuscript).

Since the backbone, input information, parameter budget, and training protocol are controlled, these results support the benefit of our action-conditioned formulation beyond simply introducing a generic learned correction module. We will include these comparisons in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer bwC7 (3/4)

W3 & Q4: State-dependent actions and performance under matched external evidence.

We agree that the original evaluation does not directly establish whether EGA-WM infers different actions under similar evidence but different historical states, or whether its improvements are more pronounced in such settings. To address these concerns, we conducted two additional analyses: (1) a PJM case study examining state-dependent action inference, and (2) a comparison of forecasting gains under similar external evidence but different historical states.

  1. State-dependent action inference. We extended the PJM case study to compare two 168-hour price histories with substantially different characteristics. Their mean/SD/maximum prices are 25.08/9.83/58.09 (State A) and 35.23/25.85/184.48 (State B) USD/MWh, respectively, while the external evidence, i.e., the 24-hour ComEd load forecasts, have similar means (10.208 vs. 10.241 GW). Under these conditions, the inferred weight of Primitive 1 is 93.18% for State A while changing to 44.18% for State B. The lower-volatility state predominantly activates Primitive 1, which captures a rise-and-fall response, whereas the higher-volatility state assigns greater weight to other primitives. This provides direct evidence that action inference depends on the historical state rather than external evidence alone. We will incorporate this comparison into the expanded Fig. 4.

  2. Forecasting gains under similar evidence and different historical states. Following the reviewer's suggestion, we first identify pairs of windows with similar external evidence and then group them according to whether their historical states are similar or different. We compare the forecasting gains between these groups to examine whether EGA-WM achieves larger improvements when similar external evidence is associated with different historical states. Specifically, we follow the procedure below:

Evidence matching: We first identify pairs of windows with similar evidence, using standardized distances with equal weight assigned to each. Only pairs with an evidence distance at or below the 25th percentile determined from training data are retained.

State grouping: We characterize each historical state using three statistics of the target history: mean, standard deviation, and linear trend, each normalized using statistics from training data. Among the selected pairs, those with state distances at or below the training data's 25th percentile are classified as similar states, while those at or above the 75th percentile are classified as different states.

Controlled evaluation: EPF pairs are drawn from the same market at sufficiently separated times; Rossmann pairs from the same store with non-overlapping forecast windows; and M5 pairs from different stores sharing the same item, U.S. state, and forecast origin. We further balance the two groups in sample size and evidence distance distribution, with each window used only once. All thresholds are determined from training data (separately for each EPF market), and neither future targets nor prediction errors are used for grouping.

The resulting MSE reductions relative to the corresponding frozen direct baseline are:

DatasetMSE reduction, similar statesMSE reduction, different states
EPF7.9%14.9%
Rossmann11.9%13.0%
M52.7%4.0%

Under matched external evidence, EGA-WM achieves larger relative MSE reductions in the different-state group across all three datasets. Together with the PJM case study, these results provide further support for the proposed state-dependent action modeling, beyond the original aggregate forecasting results. We will include this analysis and the detailed matching protocol in the revised manuscript.

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

Existing time-series forecasting methods typically use historical observations and external information to directly predict future values, without explicitly modeling how external events change system dynamics. The same promotion or change in electricity load may have effects of different strengths and durations depending on the current system state. This paper adds a recurrent correction module to an existing forecaster. It combines learnable response primitives based on historical observations and external information to determine the form, timing, and strength of an adjustment. These adjustments propagate through latent state updates, allowing their effects to persist. The difference between decoded trajectories with and without action modulation is added to the baseline forecast. The correction module is trained using forecasting error, without action annotations.

Critical strengths
  1. The motivation is reasonable. The same external event can produce different responses under different system states, and its effects may be delayed or persistent.
  2. The model separates the introduction of new effects from the propagation of existing ones through additional inputs and state retention. This allows an effect to persist after direct modulation ends and provides explicit temporal dependence beyond independent adjustments at each forecast step.
  3. The correction module requires only forecasting supervision, without annotations of response types, onset times, or durations. The experiments cover electricity and retail datasets with both forecast-based and event-based evidence. The ablations examine the primitive bank, state and evidence conditioning, temporal activation, and state propagation.
Critical weaknesses
  1. There is insufficient evidence that the model learns the actual dynamic responses induced by external events. The baseline forecaster already uses external information, and the correction module is trained solely through the final forecasting error. It may therefore correct various baseline errors, including trend and seasonal errors, rather than specifically isolate responses to external events. The results support the architecture’s predictive usefulness but do not establish this stronger interpretation.
  2. On M5, the MSE decreases from 1.5246 for Sonnet to 1.5220, a relative improvement of approximately 0.17%. Although the authors report averages over three random seeds, they provide no standard deviations or confidence intervals. It is therefore unclear whether this small improvement is consistent across runs.
  3. The temporal parameters appear interpretable, but their interpretation is not sufficiently validated. The model predicts onset, duration, overall strength, and a per-step mask, then combines them into an activation profile. Different parameter combinations may produce identical or very similar profiles, making their individual meanings ambiguous. Since the effective quantity is the activation profile ghg_h, the paper should justify this parameterization against directly predicting ghg_h.
  4. The action input takes the form uhact=gh∑jπjfj,h(c)u_h^{\mathrm{act}}=g_h\sum_j\pi_j f_{j,h}(c). The primitive weights πj\pi_j remain fixed throughout the forecast window, and all primitives share the same activation profile ghg_h. Although the time-varying parameters within each primitive may accommodate different temporal patterns, the timing of individual responses becomes entangled with the primitive shapes and shared activation profile. This weakens the claimed separation between response form and timing.
  5. Actions are inferred once from a fixed state representation before forecasting and are not updated based on the evolving latent state. The transition itself consists of elementwise scaling and additive inputs, without nonlinear interactions between state dimensions. The nonlinear encoders and decoder may partly compensate for these restrictions, but the paper does not sufficiently establish whether this simple recurrence supports its broader interpretation of system dynamics.
  6. The ablations show that individual components help within the proposed architecture, but do not establish the necessity of the overall complexity. Adding parameters and an additional training stage to an existing forecaster may itself improve performance. It remains unclear how much of the gain comes from the specific proposed structure rather than from a more general correction module.
Questions for authors
  1. Why do the baseline results differ between Tables 1 and 2? Appendix B states that the base forecaster BdirB_{\mathrm{dir}} is Sonnet, while Section 4.3 defines the Baseline in Table 2 as the direct forecast without the action branch. Please clarify any differences in configuration, training, or checkpoints between this baseline and Sonnet in Table 1.
  2. For M5, are future prices taken from price schedules known at the forecast cutoff or from subsequently recorded actual prices? What is the source of the future opening-status information used for Rossmann, and was it available at prediction time?
7 comments
AuthorsOpenReview ↗

Response to Reviewer UVMt (1/4)

Thank you for the detailed review and for recognizing our motivation of state-dependent and persistent responses, as well as the separation between introducing new effects and propagating existing ones. We apologize for the delayed response, as we conducted a substantial number of additional experiments to address the concerns and further strengthen the paper. We have addressed your concerns point by point below, and the revised manuscript incorporating these changes will be provided shortly. Results below are reported as mean ± SD over three random seeds, with EPF results averaged over five markets.

W1: Whether the correction captures responses to external evidence.

To further investigate whether the learned corrections capture responses to external events rather than merely correcting general forecasting errors, we conducted two complementary analyses: (1) evidence shuffling, to examine whether the correction gains depend on correctly aligned external evidence rather than merely correcting baseline errors; and (2) evidence-period analysis, to assess whether the corrections are particularly beneficial during evidence periods.

  1. Sensitivity to external evidence. We shuffled only the external evidence provided to the correction module, while keeping the direct forecaster's inputs and predictions unchanged.
ConditionEPF MSERossmann MSEM5 MSE
Direct baseline0.2398 ± 0.00360.0105 ± 0.00361.6513 ± 0.0035
EGA-WM0.2087 ± 0.00320.0084 ± 0.00291.5220 ± 0.0027
EGA-WM, shuffled evidence0.3330 ± 0.04580.0178 ± 0.00841.6962 ± 0.0093

Shuffling the evidence reverses the improvements on these datasets. These results show that the correction module relies on the correct external evidence; otherwise, incorrect evidence may even bring negative impacts.

  1. Performance during evidence and ordinary periods. We further examined whether the correction provides greater benefits during periods with the external evidence. For Rossmann, evidence periods include promotions, holidays, and store closures; for M5, they include calendar events and price changes. Since EPF has no explicit event labels, we identify high-change windows using the 75th percentile of standardized evidence changes relative to the preceding week, with the threshold determined from the training set.
DatasetMSE reduction in evidence periodsMSE reduction in ordinary periods
EPF18.8%15.2%
Rossmann24.8%10.7%
M59.0%7.6%

EGA-WM achieves larger MSE reductions during evidence periods across all three datasets, suggesting that its corrections are particularly beneficial when external conditions change.

Together, these analyses provide additional evidence that the correction module utilizes external evidence and is particularly effective during external evidence-related changes. We will clarify this distinction and incorporate the additional analyses into the revised manuscript.

W2: Consistency of the improvements on M5.

We agree that the small MSE improvement on M5 requires further evidence of its consistency across random seeds. To address this concern, we report the mean ± standard deviation (SD) over three random seeds for all three datasets, with EPF results averaged over five markets.

MethodEPF MSERossmann MSEM5 MSE
XLinear0.3176 ± 0.04140.0586 ± 0.00701.5344 ± 0.0022
Sonnet0.2416 ± 0.00350.0194 ± 0.00961.5246 ± 0.0045
TimeMosaic0.2473 ± 0.00290.2089 ± 0.00841.5429 ± 0.0031
DAG0.2188 ± 0.00270.0727 ± 0.02361.5719 ± 0.0232
EGA-WM0.2087 ± 0.00320.0084 ± 0.00291.5220 ± 0.0027

To further examine whether the small improvement on M5 is consistent across runs, we additionally report the MSE for each random seed:

Methodresult 1result 2result 3
XLinear1.53281.53351.5369
Sonnet1.52031.52421.5293
TimeMosaic1.53961.54351.5458
DAG1.54741.57461.5936
EGA-WM1.51971.52101.5250

EGA-WM achieves the consistent improved performance compared to SOTAs in these runs. It also exhibits a lower SD than Sonnet (0.0027 vs. 0.0045). Although the average improvement is modest (0.17%), the per-seed results show that the advantage is observed across all three runs rather than being driven by a single favorable seed. We will include the SDs and clarify the consistency of the M5 results in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer UVMt (2/4)

W3: Why parameterize temporal activation instead of predicting it directly?

The temporal parameters are designed as a structured parameterization of the temporal activation profile ghg_h, rather than as individually identifiable quantities. Specifically, onset and duration define its soft temporal support, strength controls its overall magnitude, and the per-step mask allows flexible temporal adjustments. Although different parameter combinations may produce similar profiles, this parameterization introduces a structured inductive bias for modeling when and how strongly a response is activated, rather than predicting each time step independently.

To further validate the benefit of this parameterization, we conducted an additional experiment that directly predicts ghg_h, while keeping all other components unchanged.

Temporal activationEPF MSERossmann MSEM5 MSE
Direct prediction0.2139 ± 0.00600.0091 ± 0.00281.5414 ± 0.0057
Our parameterization0.2087 ± 0.00320.0084 ± 0.00291.5220 ± 0.0027

Our parameterization achieves lower MSE across all three datasets, with relative reductions of 2.4%, 7.7%, and 1.3% on EPF, Rossmann, and M5, respectively. It also achieves lower MAE and MASE across all datasets. These results support the benefit of structured temporal activation over unconstrained direct prediction.

We will include this comparison in our manuscript and clarify that the temporal parameters provide an inductive bias for learning ghg_h.

W4: Shared temporal activation and the separation of response form and timing.

The separation described in Sec. 3.4 refers to the factorization of the action input, rather than a complete disentanglement of the realized response. Specifically, the primitive weights π\pi determine which transition modifications are combined, while the shared activation profile ghg_h controls when and how strongly these modifications affect the action input and retention. When ghg_h approaches zero, the action's influence on both components is suppressed, regardless of the primitive combination. As the primitive components Pj,hP_{j,h} also vary over time, the realized temporal response is jointly shaped by the primitives and ghg_h. We will clarify this distinction by replacing separates with factorizes in Sec. 3.4.

Regarding the shared activation design, we treat primitives as components of a single window-level action rather than independent event-specific actions. A shared ghg_h provides a common temporal envelope for the combined action while avoiding separate temporal parameters for each primitive.

To further evaluate this choice, we conducted an additional experiment assigning each primitive its own onset, duration, strength, and mask, resulting in independent activation profiles gj,hg_{j,h}, while keeping the remaining components unchanged.

Temporal activationEPF MSERossmann MSEM5 MSE
Shared (ours)0.2087 ± 0.00320.0084 ± 0.00291.5220 ± 0.0027
Independent0.2078 ± 0.00590.0086 ± 0.00151.5293 ± 0.0029

The independent design yields a marginal MSE improvement on EPF but performs worse on Rossmann and M5, while increasing the action-head and dynamics parameters from approximately 84k/33k/28k to 126k/63k/51k across the three datasets. Thus, allowing primitive-specific timing introduces substantially more parameters without consistent performance gains. We therefore retain the shared activation design and will clarify this factorization and the corresponding trade-off in the manuscript.

AuthorsOpenReview ↗

Response to Reviewer UVMt (3/4)

W5: Horizon-level action inference and transition expressiveness.

  1. Horizon-level action inference. We infer the action once at the forecast cutoff because it represents a horizon-level response conditioned on the historical state and the full future evidence available at that time. Since no new observations become available within the forecast horizon, the latent states along the horizon are fully determined by these inputs and the inferred action, so re-inferring the action from them would not introduce new information. Although the action is not re-inferred from evolving latent states, its effects vary across forecast steps through the temporal activation ghg_h and step-dependent primitive components Pj,hP_{j,h}. This provides a structured way to model time-varying responses without repeatedly updating the action.

  2. Transition expressiveness. The elementwise linear transition is a deliberate design choice that enables efficient parallel computation while maintaining stable latent-state propagation through bounded retention coefficients. Although the transition itself does not explicitly model nonlinear cross-dimensional interactions, the nonlinear encoders and decoder allow such dependencies to be represented in the overall forecasting model.

To further evaluate whether a more expressive transition is necessary, we introduced an additional nonlinear interaction term Atanh⁡(z)A\tanh(z) into both the action-conditioned and no-action transitions, while keeping the action inference, primitive bank, temporal activation, encoders, and decoder unchanged.

TransitionEPF MSERossmann MSEM5 MSE
Original0.2087 ± 0.00320.0084 ± 0.00291.5220 ± 0.0027
With nonlinear interaction0.2096 ± 0.00550.0083 ± 0.00341.5214 ± 0.0032

The nonlinear transition achieves comparable performance, with marginal improvements on Rossmann and M5 but slightly worse results on EPF. However, it introduces state-dependent interactions that require sequential latent updates rather than the parallel scan supported by our original transition. These results suggest that the simpler transition provides sufficient expressiveness for the evaluated forecasting tasks, without consistent benefits from additional nonlinear interactions.

We will clarify the motivation and scope of our transition design in Sec. 3.5 and discuss more expressive nonlinear transitions as potential extensions for more complex dynamical systems.

W6: Contribution of the proposed structure beyond additional parameters and training.

To examine how much of the gain comes from the proposed structure rather than from additional parameters and training, we conducted additional experiments with parameter-matched MLP and GRU residual correctors on the same frozen backbones. All methods use the same state and evidence inputs, encoder architectures, and additional training stage with an identical training budget.

CorrectionEPF MSERossmann MSEM5 MSE
Residual MLP0.2352 ± 0.00310.0094 ± 0.00481.5721 ± 0.0039
Residual GRU0.2333 ± 0.00330.0090 ± 0.00471.5824 ± 0.0042
EGA-WM0.2087 ± 0.00320.0084 ± 0.00291.5220 ± 0.0027

The compared correction components have 83,977/33,076/28,069 parameters on EPF/Rossmann/M5, respectively. The shared encoder architectures are also included and trained under the same settings.

EGA-WM achieves the lowest MSE across all three datasets, reducing MSE by 10.5%, 6.7%, and 3.2% relative to the stronger residual corrector on EPF, Rossmann, and M5, respectively. It also achieves lower MAE and MASE than both alternatives across all datasets (full results will be provided in the revised manuscript).

Since the additional parameter budget and training stage are controlled, these results support the contribution of our proposed structure beyond simply adding model capacity or an additional correction stage. We will include these comparisons in the revised manuscript.

Reviewer_qo37OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper looks at time-series forecasting with external evidence, where the same evidence can affect the future differently depending on the current state of the system. EGA-WM infers a latent action from the evidence and the current state and uses the action to correct a base forecast through a small recurrent latent model. On electricity-price and retail sales forecasting, EGA-WM has lower average errors than all compared forecasters, including ones that also use external evidence.

Critical strengths

The paper treats external evidence as something whose effect depends on the current state of the system, and this is a useful way to think about forecasting with external information. Table 2 supports this view, since removing state conditioning from action inference raises the error on every dataset and metric. The experiments pair hourly electricity prices with day-ahead load, generation or wind forecasts, and daily retail sales with promotion, holiday, calendar and price information. The same components handle both, so the approach does not seem tied to one kind of evidence.

Critical weaknesses
  • Most of my doubts are about the base forecaster, which the main text never names beyond "a pretrained, frozen forecaster". Appendix B says it is Sonnet, which is also a baseline in Table 1, so Table 2's "Baseline" (the base forecast without the action branch) should match the Sonnet row in Table 1, but the two are close only on EPF. On Rossmann the ablation baseline already has about half the MSE of the Sonnet row, so much of the MSE gain over Sonnet comes from the base forecaster itself, and on M5 the gap goes the other way.

  • There is also no comparison with a plain residual corrector of similar size on the same frozen Sonnet, and every Table 2 variant keeps the action-inference structure, so the experiments don't fully separate the action formulation from a generic learned correction. Table 4 gives only layer widths, and the paper reports no parameter counts or training cost, even though EGA-WM runs two separately pretrained Sonnet models plus the action model.

  • The paper's central claim that the same evidence leads to different responses depending on the state is shown only indirectly. The ablation shows that state conditioning helps, and Fig. 4 covers one PJM window, but there is no analysis of the same evidence leading to different inferred actions under different states.

  • No standard deviations are reported, so it is not clear that EGA-WM actually wins on M5, where it leads the strongest baseline by only a fraction of a percent on each metric.

Questions for authors

Q1. Table 2's Baseline is the direct forecast btb_t, and Table 1's Sonnet receives the same evidence. Why do the two numbers differ?

Q2. Could you compare EGA-WM with a plain residual corrector of similar size on the same frozen Sonnet (see above)?

Q3. Could you compare inferred actions for windows with similar evidence but different states, either as examples or as a statistic over the test set?

Q4. Could you report standard deviations over seeds, at least on M5?

7 comments
AuthorsOpenReview ↗

Response to Reviewer qo37 (1/4)

Thank you for the careful review and for recognizing the motivation behind state-dependent responses. Your constructive comments have helped us improve the clarity and rigor of our work. We apologize for the delayed response, as we conducted a substantial number of additional experiments to address the concerns and further strengthen the paper. We have addressed your concerns point by point below, and the revised manuscript incorporating these changes will be provided shortly. The results below are means ± SD over three seeds, and EPF is the five-market average.

W1 & Q1: Clarification of the backbone and baseline differences.

The baseline in Table 2 is based on Sonnet, but uses different configurations from the standalone Sonnet reported in Table 1. Specifically, the backbone configuration in Table 2 is selected based on the validation performance of the complete EGA-WM framework and subsequently kept fixed when disabling the action branch for ablation. In contrast, Table 1 uses a configuration independently selected based on the validation performance of standalone Sonnet forecasting. The two Sonnet models are trained separately with their respective configurations and therefore use different pretrained checkpoints. The detailed differences are summarized below (Table 1 Sonnet → EGA-WM backbone):

Datasetd_modelSonnet αDownsampling scalesLearning rateParameters
EPF64 → 640.50 → 0.25[1,2,4] → [1,2,3]3e-4 → 1e-3195,539 → 195,491
Rossmann32 → 440.50 → 0.75[1,2] → [1,2]1e-3 → 5e-4140,523 → 154,290
M564 → 560.25 → 0.75[1,2,4] → [1,2,4]3e-4 → 1e-3182,375 → 163,193

Here, α denotes a Sonnet-specific hyperparameter, distinct from the onset parameter in EGA-WM. Parameter counts refer only to the direct forecasting backbone (per market for EPF), excluding the history forecaster and correction module. These differences in configurations and pretrained checkpoints explain the discrepancy between the two baselines. Since validation-based selection does not guarantee the best test performance, neither configuration consistently outperforms the other on the test set: the configuration in Table 2 achieves lower MSE on Rossmann (0.0105 vs. 0.0194), whereas the standalone configuration achieves lower MSE on M5 (1.5246 vs. 1.6513). We will clarify this distinction in the main manuscript and provide the detailed configurations in Appendix B.

Importantly, the Baseline in Table 2 is evaluated using the exact frozen direct forecaster employed in each corresponding EGA-WM run, rather than a separately trained baseline. Thus, Table 2 isolates the contribution of the action branch under identical backbone settings and checkpoints. Adding the action branch reduces MSE by 13.0%, 20.0%, and 7.8% on EPF, Rossmann, and M5, respectively. On Rossmann, this 20.0% improvement is obtained on top of the stronger configured baseline, so the gain is not solely due to the backbone settings.

To further disentangle the contribution of EGA-WM from that of backbone configuration, we conducted an additional experiment using the exact Sonnet model reported in Table 1. We kept the original Sonnet frozen and added our action-based correction module without modifying or retuning its backbone configuration.

ModelEPF MSERossmann MSEM5 MSE
Original Sonnet0.2416 ± 0.00350.0194 ± 0.00961.5246 ± 0.0045
Same Sonnet + EGA-WM0.2173 ± 0.00360.0128 ± 0.00521.5250 ± 0.0038

The correction improves all three metrics on EPF and Rossmann. On M5, although MSE remains nearly unchanged, MAE and MASE improve from 0.9008/1.6350 to 0.8878/1.6140, respectively. These results provide further evidence that the gains are not merely due to backbone configuration but also arise from the proposed action-based correction mechanism. We will include this controlled comparison in the revised manuscript to clarify the contribution of EGA-WM.

AuthorsOpenReview ↗

Response to Reviewer qo37 (2/4)

W2 & Q2: Action formulation versus generic residual correction and model complexity.

We agree that the original ablations do not fully distinguish the contribution of our action formulation from that of a generic learned correction. To directly address this concern, we conducted additional experiments with parameter-matched MLP and GRU residual correctors and expanded the model complexity analysis to account for all components.

  1. Comparison with generic residual correctors. We replaced the proposed action-based correction module with standard MLP and GRU residual correctors while keeping the same frozen backbones, state and evidence inputs, encoder architectures, and training protocol. All methods use the same correction-stage training budget, with correction modules matched in parameter count (83,977/33,076/28,069 on EPF/Rossmann/M5, respectively).
CorrectionEPF MSERossmann MSEM5 MSE
Residual MLP0.2352 ± 0.00310.0094 ± 0.00481.5721 ± 0.0039
Residual GRU0.2333 ± 0.00330.0090 ± 0.00471.5824 ± 0.0042
EGA-WM0.2087 ± 0.00320.0084 ± 0.00291.5220 ± 0.0027

EGA-WM achieves the lowest MSE across all three datasets, reducing MSE by 10.5%, 6.7%, and 3.2% relative to the stronger residual corrector on EPF, Rossmann, and M5, respectively. It also achieves lower MAE and MASE than both alternatives across all datasets (Full results will be provided in the main manuscript). Since the backbone, input information, parameter count, and training budget are controlled, these results provide evidence that the improvement stems from the proposed action formulation rather than simply introducing an additional learned correction module. We will include this comparison in the main manuscript.

  1. Model complexity and training procedure. We acknowledge that the original Table 4 does not fully reflect the model complexity. EGA-WM consists of two separately pretrained Sonnet models and trainable encoders and action model. We have expanded Table 4 to report the parameter counts of all components:
DatasetFrozen history forecasterFrozen direct forecasterTrainable parameters, including encodersTotal
EPF, per market195,443195,491335,754726,688
Rossmann153,795154,290153,205461,290
M5159,959163,193388,262711,414

Both Sonnet models are pretrained separately and subsequently frozen. During correction training, only the encoders and action model are updated, while the frozen backbone outputs can be precomputed and cached. We will clarify the complete model composition, parameter breakdown, and separate training stages in Table 4 and Appendix B.

Importantly, while EGA-WM introduces additional model capacity, the parameter-matched comparisons above show that its improvements cannot be explained by increased correction-module capacity alone, further supporting the effectiveness of the proposed action formulation.

AuthorsOpenReview ↗

Response to Reviewer qo37 (3/4)

W3 & Q3: Direct evidence of state-dependent action inference.

To directly address this concern, we conducted two additional analyses: (1) an extended PJM case study to visually examine whether different historical states lead to different inferred actions under similar evidence, and (2) a quantitative evaluation to investigate whether EGA-WM achieves larger forecasting gains under similar evidence but different historical states.

To visually show this, we first selected two 168-hour price histories with substantially different characteristics. Their mean/SD/maximum prices are 25.08/9.83/58.09 (State A) and 35.23/25.85/184.48 (State B) USD/MWh, respectively, while their evidence, i.e., the 24-hour ComEd load forecasts, have very similar means (10.208 vs. 10.241 GW). Under these conditions, the inferred weight of Primitive 1 is 93.18% for State A while changing to 44.18% for State B. The lower-volatility state predominantly activates Primitive 1, which captures a rise-and-fall response, whereas the higher-volatility state assigns greater weight to other primitives. This analysis will be incorporated into the expanded Fig. 4 in the main manuscript.

Beyond the PJM case study, we further quantitatively examine whether EGA-WM achieves larger forecasting gains under similar external evidence but different historical states. Specifically, we first identify pairs of windows with similar external evidence and then group them according to whether their historical states are similar or different. We then compare the forecasting gains between these groups. The detailed procedure is as follows:

Evidence matching: We first identify pairs of windows with similar evidence, using standardized distances with equal weight assigned to each. Only pairs with an evidence distance at or below the 25th percentile determined from training data are retained.

State grouping: We characterize each historical state using three statistics of the target history: mean, standard deviation, and linear trend, each normalized using statistics from training data. Among the selected pairs, those with state distances at or below the training data's 25th percentile are classified as similar states, while those at or above the 75th percentile are classified as different states.

Controlled evaluation: EPF pairs are drawn from the same market at sufficiently separated times; Rossmann pairs from the same store with non-overlapping forecast windows; and M5 pairs from different stores sharing the same item, U.S. state, and forecast origin. We further balance the two groups in sample size and evidence distance distribution, with each window used only once. All thresholds are determined from training data (separately for each EPF market), and neither future targets nor prediction errors are used for grouping.

The resulting MSE reductions relative to the corresponding frozen direct baseline are:

DatasetMSE reduction, similar statesMSE reduction, different states
EPF7.9%14.9%
Rossmann11.9%13.0%
M52.7%4.0%

Under matched external evidence, EGA-WM achieves larger relative MSE reductions in the different-state group across all three datasets. Together with the PJM case study, these results provide further support for the proposed state-dependent action modeling. We will include this analysis in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.