acceptodds
Under review as a conference paper at ICLR 2027

Learning Latent Actions from Joint Histories for Multivariate Time-Series Forecasting

Abstract

In multivariate time series forecasting, the same target history can warrant different futures as surrounding observations change. Forecast accuracy alone leaves the context-induced revision unlocalized. ACT defines a forecasting state–action interface: it retains a representation encoded from the target history alone, uses joint histories to select a replaceable action, and reads the updated state with a shared forecast head. A fixed-state intervention protocol replaces only that action, making its forecast effect directly measurable and exactly replayable. LAB is one realization of this interface, using structured feature and variable interactions for exact action extraction and low-rank implementation; direct scaling provides a second realization under the same interface. Across 20 models on four datasets, actions selected from the observed joint history beat zero, equal-magnitude noise, and shuffled actions; direction and target matching matter beyond magnitude. Across five Solar sites and 2,527 active site–time pairs, the observed context lowers six-hour mean squared forecast error by 28.0% relative to substitutions matched on target history and daily phase; replay reconstructs the forecast difference to numerical precision. ACT is competitive with 12 forecasters across eight benchmarks. Relative to dense LAB, low-rank LAB cuts training-and-validation time per epoch by about 42% on Electricity and 56% on Traffic.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Latent Actions

Reviews

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

ACT separates the forecast for a target variable into a state and an action. The state is encoded from the target's own history, and the action is generated from the joint history of all variables and applied as an update to that state before a shared head produces the forecast. The motivation is that existing multivariate forecasters change their forecast for a target as its context changes, even when the target's own history is fixed, yet this change is mixed into the final output and cannot be measured in isolation. ACT addresses this with a fixed-state intervention protocol. The target state is held fixed and only the action is replaced, so the forecast change caused by the context can be measured and replayed exactly. LAB is the paper's main realization of the action generator. Its multiplicative routes allow the action to be extracted in closed form, and low-rank gates keep the computational cost down. On ECL and Traffic, low-rank LAB reduces per-epoch training time by about 42% and 56% relative to dense LAB (Table 27). For validation, ACT is compared with 12 baselines on eight benchmarks and is best or tied-best on four datasets in MSE and on four in MAE (Table 2, Appendix C). In the trained models, replacing the learned state update with zero, norm-matched noise or an update shuffled across test histories raises the error in all 20 checkpoints (Table 3). The paper also tests whether context information can be read out linearly from the action (Table 6) and compares LAB with other action generators such as direct scaling (Table 7).

Critical strengths
  1. The fixed-state intervention experiments are systematic and directly test the paper's central claim that the action carries meaningful information beyond the state. In the trained models, the authors replace the update D added to the state not only with zero but also with norm-matched noise and with shuffled updates that preserve the empirical distribution. The native D performs best across all 20 models (Tables 3, 14). The shuffle condition is the most telling. It keeps the distribution of update values unchanged and still degrades performance substantially, which shows that the update carries state-specific directional information rather than mere magnitude. This supports the paper's central argument that the separation of state and action is a meaningful structure.

  2. The evaluation is broad, and the paper documents everything needed to reproduce it. Eight benchmarks of different types are used. Appendix A reports the hyperparameters, the selected configuration for every dataset and horizon cell (Table 8) and the exact run commands, and Appendix G reports the hardware and the full configuration-search cost (Table 28). The analysis of how context influences the forecast is similarly thorough, covering error alignment (Table 4), the interaction between context groups (Table 5) and linear readouts of context from the action (Table 6). This makes the results easy to verify and adds to my confidence in the claims.

Critical weaknesses
  1. The methodological novelty of ACT and LAB is limited. As the authors state in Sections 1 and 2, the core operation, multiplying the retained state by coefficients generated from the joint history, is conventional conditional modulation in the style of FiLM, and the contributions are defined as the state-action interface and the fixed-state intervention protocol. When LAB is replaced by a simpler direct-scaling generator, the MSE is similar or better on Weather, Solar and Traffic, and LAB is ahead only on ECL (Table 18). The presentation also makes the contribution harder to assess. The paper introduces many terms such as action, applied update, forecast correction, recipient and replay, while the gate gℓg_\ell in Eq. 4 is defined only in Appendix A.1 (Eq. 9) and the recurrence for the route scales StS_t, ScS_c, SrS_r only in Appendix B.2 (Eq. 11), so LAB and Prop. 1 cannot be reconstructed from the main text alone. In short, the substantive contribution is closer to recasting existing multiplicative modulation as an evaluation perspective than to a new architecture, and the paper does not show convincingly that this perspective yields insights that existing methods cannot provide. This is the main reason for my weak-reject recommendation.

  2. The practical forecasting gain from adding the action is limited. ACT is ahead on four of the eight datasets, but it is behind the dependency-aware models on the high-dimensional datasets where cross-variable information should matter most, namely ECL and Traffic (Table 2). At H=720, ACT is behind the best linear or mixer model on ETTh1, ETTh2 and ECL (Table 11). For ETTh2 at H=720, validation selects depth 0, which means no action at all (Table 8). Compared with models retrained without an action, the advantage ranges from 0.2% to 7.1% in MSE (Table 19).

Questions for authors

The following are clarification requests, and the answers would affect my assessment.

  1. The fixed-state context-swap protocol is presented as a key contribution. What does it reveal that analogous interventions on existing models cannot, for example swapping the non-target inputs to a cross-variable model like iTransformer or TimeXer, or to a channel-independent baseline? A comparison under the same protocol would make it clear whether the insight is specific to the ACT interface.

  2. Section 3.3 notes that direct scaling yields a coefficient that is explicit by construction, and Table 18 shows direct scaling matching or outperforming LAB on Weather, Solar and Traffic. What does LAB offer that direct scaling does not? The training-cost reduction in Table 27 compares low-rank LAB with dense LAB, so it does not answer this question. Is there any analysis in the paper, such as the route geometry in Section 4.5, that requires LAB and could not be carried out with direct scaling?

  3. Appendix C describes the baseline numbers as reference point estimates. Where do they come from, and were the baselines tuned with a search comparable to the 20 configurations per dataset and horizon cell that ACT receives? Appendix A.2 states that sample standard deviations are computed over the five seeds, yet Tables 2 and 11 to 13 report means only. Please report those standard deviations for ACT and, where possible, for the baselines as well.

  4. At H=720, ACT is behind the best linear or mixer model on ETTh1, ETTh2 and ECL (Table 11), and validation selects depth 0 for ETTh2 at H=720, which means no action is applied (Table 8). Does the action contribute little at long horizons? If so, please explain why, and state whether the action gives any practical forecasting gain at long horizons.

Reviewer_nFZiOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents a forecasting framework ACT that separates a target’s history representation from a latent update generated using joint histories. Its Latent Action Block combines feature, variable, and residual routes, allowing the update to be extracted and its forecast effect replayed with the target state and prediction head fixed. The evaluation combines forecasting benchmarks with action replacements, context substitutions, representation probes, and comparisons between dense and low-rank implementations.

Critical strengths
  1. Separating the retained target state from its contextual update provides a clear way to examine how surrounding observations revise a forecast. Holding the state, normalization, and prediction head fixed makes these comparisons easier to interpret.

  2. The intervention analysis goes beyond standard module removal. Norm-matched noise, shuffled updates, recipient matching, and context substitutions examine different aspects of the learned update, with a useful distinction between action coefficients and applied updates.

  3. The experiments include retrained no-action models, matched modulation controls, and detailed computational measurements. The low-rank implementation reduces the training cost of dense LAB while maintaining competitive forecasting accuracy.

Critical weaknesses
  1. The residual-route ablation does not isolate the value of LAB’s decomposition. Replacing the residual output with its input collapses the entire updated representation to the retained state, effectively repeating the zero-action comparison. This offers limited evidence for why the proposed route structure is needed.

  2. Matching donors on target-history summaries and calendar features does not ensure that the substituted multivariate histories remain jointly plausible. The resulting performance degradation may therefore partly reflect sensitivity to inconsistent inputs. Conditional resampling or a controlled synthetic experiment would help distinguish this from useful contextual dependence.

  3. ACT undergoes dataset- and horizon-specific configuration searches, whereas the competing methods are reported as reference point estimates. Comparable tuning budgets and repeated-run uncertainty are not established for these baselines, making the smaller performance differences difficult to assess.

  4. Although the paper distinguishes context readability from predictive utility, the experiments do not directly connect the decoded information to the action components responsible for forecast changes. The linear head can discard action directions, so successful readouts and useful context interventions need not reflect the same information.

Questions for authors
  1. Can the same intervention analyses be applied to the matched direct-scaling control? Which findings would distinguish the contribution of ACT/LAB from conventional conditional modulation?

  2. How sensitive are the context-substitution results to the matching variables and donor-distance threshold, and can the authors check whether the recombined histories remain within the observed joint distribution?

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

This work is motivated by the hypothesis that the future state of one variable can be influenced by surrounding variables in time series data. The authors propose ACT as a "state–action interface." It encodes each variable independently into zcz_c, and the zcz_c values of all variables are used to generate a single latent action coefficient vector, aca_c. Each zcz_c is then updated by these latent action coefficients according to the equation zc=zc+ac⊙zcz_c = z_c + a_c \odot z_c. The latent action block (LAB) is a concrete realization of ACT, utilizing three different routes that mix features, variables, and residuals. The work further claims that LAB is an axis-wise, low-rank realization of ACT, making changes in the forecast directly tractable to the latent action space. Experimental results on standard long-horizon forecasting benchmarks demonstrate that the proposed method is competitive against various baseline families. Furthermore, the results highlight the usefulness of the latent action coefficient in improving forecasting accuracy, showing that it effectively encodes contextual information.

Critical strengths
  1. The proposed methods are compared against baseline models across different families—including linear, mixer, and transformer-based models—on standard benchmarks. The results demonstrate the competitiveness of the proposed approaches, ACT and LAB.
  2. The theoretical results are mathematically correct and verified against quantitative experimental results. For example, Proposition 1 is proved in Appendix B.2, and Appendix B.4 further validates these properties with experiments. I appreciate the rigorous studies conducted by the author.
  3. Overall, the experimental setup is a key strength of the work. For example, Section 4.2 uses different latent action states, including zero-state, Gaussian noise, and randomly shuffled states, to demonstrate the usefulness of the proposed native latent action variable. The work further introduces an energy-alignment decomposition experiment (Table 4) to explain this. Additionally, the solar farm experiment shows that the learned latent action captures physically meaningful context.
Critical weaknesses
  1. On standard benchmark evaluations, the results in Table 2 show that the proposed method is competitive at best but does not consistently outperform the baseline families. For example, it trails behind the Dep.-Aware models on ECL and Traffic data and does not significantly outperform linear or mixer models on ETTh data.
  2. Despite demonstrating the correctness of the theoretical properties (see Strength Point 2), the work falls short of justifying their practical usefulness. For example, Proposition 1 shows that the latent action can be computed in closed form from LAB’s three different routes, and Eq. (7) shows that a change in the action can be mapped to a change in the forecast. What is missing are experimental results that demonstrate the practical value of these properties through comparisons with existing methods on standard benchmarks.
  3. Similar arguments apply to Table 1, which compares the proposed method against existing approaches across different properties. However, it is unclear how these properties help the proposed method achieve better forecasting accuracy or other practical benefits over existing baselines.
  4. Considering these points, the overall contribution of the work is unclear. On standard long-term forecasting benchmarks, the results are not impressive, nor do the theoretical results translate into better empirical performance compared to existing methods. While the computational efficiency mentioned in Section 3.3 could be a major strength, it is not thoroughly studied.
Questions for authors
  1. Could the authors clarify what the acronym ACT stands for?
  2. Why does the model update the state using zc=zc+ac⊙zcz_c = z_c + a_c \odot z_c? Can the identity matrix be absorbed into aca_c?
  3. I strongly encourage the authors to improve the clarity of the presentation in Sections 3.2 and 3.3. For example, how do the different equations in Eq. 3 correspond to the three routes LAB is built upon: the history-feature route, the cross-variable route, and the residual route? The subscripts in Eq.3 is never defined and could be confusion to readers. Additionally, how are StS_t, ScS_c, and SrS_r (introduced at the beginning of Section 3.3) computed?
  4. The ETTh1 and ETTm1 results in Table 5 show that using different variants of the latent action variable does not change the forecasting MSE at all. Does this mean that forecasting on these two datasets does not depend on contextual information from other variables, or that the model fails to capture this information? Could the authors provide a comparison against a baseline (another model or an ACT variant) that forecasts each variable independently without relying on such a latent action variable?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.