acceptodds
Under review as a conference paper at ICLR 2027

Pairwise-to-All: Relational Equilibrium for Simultaneous Forecasting

Abstract

We introduce Relational Equilibrium State Estimation (RESE), a forecasting framework for multiple interacting systems that models how their relative relations evolve over time. RESE estimates these relations directly from observed trajectories, and the pairwise relations are weighted by reliability and reconciled into a consistent system-level equilibrium state. Individual forecasts are then recovered by combining this state with a separately estimated aggregate. Experiments on synthetic, exchange-rate, COVID-19, and multivariate datasets show that RESE is competitive on its own and can improve existing forecasting models. Standalone RESE achieves the lowest RMSE on the synthetic dataset, while most forecasting baselines improve when combined with RESE on exchange rates. RESE also remains stable under moderate noise, missing observations, and local shocks, and benefits from stronger cross-system dependence. Its computational cost remains low for small and moderate system sizes. Source and data are available at https://anonymous.4open.science/r/RESE-6024

Primary area:
learning on time series and dynamical systems
Keywords:
Equilibrium, time series, real word data

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_Ax9hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
3
Read review
Paper summary

This paper introduces relational equilibrium state estimation (RESE), an extension of equilibrium state estimation (ESE) [1] which circumvents the original method's need for attribute data by instead estimating relations between systems based on their past behaviour, and using this to estimate an equilibrium state. Once estimated, this equilibrium state is then used for future forecasting.

The main benefit of this method over ESE is that it does not require attribute data (which may not be present for the systems one is interested in) while still retaining similar or better predictive performance. However, the computational complexity (which is quadratic in the number of systems, as opposed to ESE which is linear) means that handling larger numbers of systems becomes rapidly more expensive, a fact which is mentioned but potentially not made as clear as it should be.

[1] B. Xu, A. Song, J. Gao, F. Liu, Once-for-All: Scalable Simultaneous Forecasting via Equilibrium State Estimation

Critical strengths

As presented, the method does what the authors claim: it removes the need for attribute data that is inherent in the design of ESE, while still retaining similar or better performance. Putting aside the computational complexity of this method, this is a clear benefit when working with time series that do not have accompanying attribute data.

Critical weaknesses

The paper is difficult to follow, immediately assuming familiarity with the work on equilibrium state estimation (ESE) from [1] (in fact, the authors don't make it clear that ESE is defined in [1], as it is referenced as a paper in which "equilibrium-state forecasting has been used for simultaneous prediction across multiple systems". Given that this work is an extension, the initial contribution of [1] should be presented more clearly).

Beyond that, terms are introduced that are not then discussed in detail until later (the relation between two systems is defined in terms of a transformation T\mathcal{T} on page 2 that appears with no context, halfway down page 3 it is mentioned that the choice of T\mathcal{T} determines how differences between systems are represented, but no actual candidate transformations are discussed until the appendix). Similarly the estimators r^ij,t(k)\hat{r}^{(k)}_{ij,t} are introduced in equation (5) but the details of them are postponed until the appendix, where I feel that at least one example should be given here to aid clarity.

As it stands, I feel that the paper requires [1] to be read in parallel to understand its motivations, which is a significant weakness. More careful explanation of what is being done, how it differs from ESE and some justification as to why we would want to use equilibrium-based methods should be added.

The computational complexity of the algorithm should probably be made more prominent. Several of the tables have potentially misleading captions referring to computational costs being close to ESE but better than the baselines, when this may simply be due to the small number of systems being used (as the complexity is O(N2)O(N^2) where NN is the number of systems). The authors do note at points that it is more costly than ESE at larger numbers of systems (as well as mentioning the complexity) but I feel this should be made clearer. Tables containing larger numbers of systems (for which the runtime is noticeably worse) are left to the appendix. As it stands I would like to see efforts to reduce the complexity (I feel that comparing all possible pairs of systems is probably unnecessary, and one could be more thoughtful about selecting suitable comparisons).

Questions for authors
  1. The forecasting results in Section 4.1 are essentially the same experiments as in the original ESE paper, but I notice that in the currency exchange and Covid-19 experiments two baselines (FiLM and KVAE) which appear in that paper are not considered here. Is there a reason for this?

  2. The reliability weights wij,tw_{ij,t} seem only to depend on the pair (i,j)(i,j) and the historical data set V\mathcal{V} used for validation, which presumably is fixed, so the weights appear to me to be constant in tt. Is this the case or am I misunderstanding? Also (unless it's implicit in the estimator) the calculation of these weights seems to give every past piece of historical data equal precedence - do things change if instead we pay more attention to the most recent data?

  3. You mention in the future work section about using parallel computation to improve performance. While this would make the method run faster, it doesn't really reduce the complexity. Do you have any thoughts about doing the latter (for instance, you could potentially just restrict to the top kk most similar systems based on the historical data and then take relational estimates on those, and see if that makes any difference to the equilibrium state estimates).

4 comments
AuthorsOpenReview ↗

Response to Weaknesses 1 - 4

We sincerely thank you for the reviews. Please see our responses below.

W1. Understanding RESE requires reading the original ESE paper.

We agree that the paper should explain more about ESE and have updated the manuscript.

ESE estimates a system-level equilibrium state using system attributes. RESE retains the separation of relative-state estimation and forecasting, but replaces the attribute-based equilibrium estimation with a relation-based one, hence requires no attributes.

The Introduction states the distinction and the underlying mechanism with a new Figure 1. Section 3.5 gives the reconstruction in Eq.9. However, these passages do not sufficiently explain the ESE framework on which RESE builds.

W2. The transformation T\mathcal T and the pairwise relation estimators are introduced without enough explanation.

We appreciate your suggestion to clarify these methodological details. Both the transformation and the pairwise relation estimators are further described in the updated manuscript, with their detailed formulations provided in Appendices A and B. We clarify their roles below and will add brief examples to the main text to make the presentation easier to follow.

Transformation T\mathcal T. As explained in Section 3.1 (lines 134–144), T\mathcal T determines how pairwise differences are represented. The main formulation uses T(s)=log⁡s\mathcal T(s)=\log s for positive observations, so that pairwise relations correspond to log-ratios of system values and their relative shares (Appendix A.1, lines 711–721). Other representations, including asinh, signed-log, and additive differences, are provided in Appendix A.3 (lines 733–746).

Pairwise relation estimators. Section 3.2 (lines 146–163) describes the historical validation and model selection procedure. Appendix B.1 (lines 756–856) specifies six candidate model families covering different relation dynamics. For example, the stable relation model predicts a pairwise relation using the median of its historical values, while other models capture trends, autoregressive dynamics, nonlinear patterns, and regime changes.

W3. The motivation for equilibrium-based forecasting is not sufficiently explained.

Thank you for raising this point. We clarified the motivation behind equilibrium-based forecasting in the updated manuscript as well as below.

The central idea is that the relative states of multiple systems are shaped by their interactions. ESE estimates the resulting equilibrium using system attributes that influence these interactions. For example, in exchange-rate forecasting, macroeconomic attributes such as GDP, inflation, and trade conditions provide information about the relative states of different currencies.

RESE follows the same motivation but approaches the problem from the observed outcomes of these interactions. Changes in the underlying attributes and their combined effects may be reflected in the historical relationships between system trajectories. Rather than explicitly modeling these attributes, RESE forecasts the evolving pairwise relations and reconciles them into a consistent equilibrium state (Sections 2–3, lines 105–132 and 148–186). The estimated state is then combined with the aggregate forecast to obtain individual predictions (Eq.9).

This allows RESE to retain the equilibrium-based forecasting framework while extending its applicability to systems without available or aligned attribute data. To clarify the motivation for equilibrium-based forecasting, we have added Figure 1 in Section 2, illustrating how the historical evolution of system values and their relative relations provides information for estimating future equilibrium states.

W4. The quadratic computational cost is understated, and the runtime comparisons may give a misleading impression.

The computational cost of RESE is explicitly discussed in Section 4.3 (lines 356–372), where Table 5 shows how pairwise relation estimation scales with the number of systems. Appendix G.5 (lines 1148–1168) further provides the complexity analysis. We also identify the quadratic growth in pairwise relations as a main limitation in Section 6 (lines 479–485).

The runtime comparisons in Section 4.1 are specific to their experimental settings. In particular, the observation that RESE has costs close to ESE refers to the 16-system exchange-rate experiment (lines 262–265), rather than a general claim across system sizes. The COVID-19 discussion (lines 267–275) explicitly notes that RESE becomes slower than ESE at 79 and 320 systems.

Nevertheless, we understand the concern that these limitations could be more visible alongside the main results. We have clarified the scope of the runtime comparisons and added the larger-system results from Appendix F more prominent in the main text.

AuthorsOpenReview ↗

Responses to Questions 1 - 3

Q1. Why were FiLM and KVAE omitted from the exchange-rate and COVID-19 comparisons?

Thank you for the question. Our baseline selection was intended to cover a broad range of forecasting approaches, including statistical, recurrent, linear, graph-based, and Transformer-based models, rather than reproduce the entire baseline set of ESE. We therefore focused on a compact but diverse set of established and recent methods. FiLM and KVAE were not included in the initial comparisons as part of this selection, rather than because of any incompatibility with RESE.

Nevertheless, both models can be integrated into RESE as aggregate predictors. We are open to including their results in the revised manuscript to provide a more complete comparison with the original ESE study.

Run FiLM+RESE and KVAE+RESE on the exchange-rate and COVID-19 tasks using the settings and metrics of the existing comparisons. Present the standalone and ESE values from ESE alongside the new RESE values; verify comparability before reusing those numbers.

ModelBase RMSE / MAERESE RMSE / MAERESE runtime (min)
Exchange rates (16 systems)FiLM5.93/5.545.52/5.270.70
Exchange rates (16 systems)KVAE6.05/5.595.45/5.160.51
COVID-19 (20 systems)FiLM76.84/73.8572.83/69.840.86
COVID-19 (20 systems)KVAE70.85/64.4473.42/68.360.67
COVID-19 (79 systems)FiLM56.21/51.3452.03/50.464.31
COVID-19 (79 systems)KVAE53.12/52.7151.93/50.274.12
COVID-19 (320 systems)FiLM5.74/4.964.49/4.1358.45
COVID-19 (320 systems)KVAE5.29/4.944.57/4.2057.97

Q2. Are the reliability weights constant in time? Would placing more weight on recent history change the results?

Regarding the time dependence of the weights, your interpretation holds when the validation set and relation scale are unchanged. In the selection-locked evaluation protocol, reliability weights are calibrated using pre-test data and retained during test forecasting. However, the relation models are refitted at each forecast origin using the available history (Appendix G.1, lines 1097–1109), so their predictions and the resulting equilibrium state continue to evolve. The framework also permits recalibration of reliability weights when new historical information becomes available. We have clarified the evaluation protocol and the notation in Eq.6.

Regarding the influence of recent data, Eq.5 gives equal weight to errors across the selected validation origins, but this does not mean that all past observations receive equal consideration. The implementation supports sliding historical windows and limits on the number of recent validation origins (Appendix G.1, lines 1097–1109). Thus, older observations can be excluded from both model fitting and reliability calibration. Assigning progressively greater weight to recent validation errors is another possible approach. It may change the selected models, reliability weights, and forecasts, although its effect on prediction accuracy has not yet been evaluated.

Q3. Could selecting the top-kk most similar systems reduce computation without materially changing the equilibrium estimates?

Thank you for this suggestion. Selecting a smaller set of system pairs is a reasonable way to reduce the computational cost of RESE.

The equilibrium estimation in Eq.7 does not require all possible pairwise relations. As discussed in Appendix C.1–C.5 (lines 897–959), the weighted reconciliation can operate on a subset of relations, with additional constraints when the resulting graph is disconnected.

Selecting the top-kk most similar systems before relation estimation could reduce the number of pairwise model fits from O(N2)O(N^2) to O(Nk)O(Nk), although the selection step introduces its own computational cost. The effect on equilibrium estimates depends on which relations are retained, since similarity does not necessarily imply predictive reliability.

We consider sparse relation selection a useful direction for improving scalability and have included it in our discussion of future work.


We hope the above response has fully addressed your comments. If so, we would greatly appreciate it if you could reconsider your rating.

Thank you for your response, the revised manuscript has sufficiently addressed my concerns regarding the paper's clarity and readability. My main concern about the computational complexity remains, however, as I feel that requiring relational estimates between every pair is unnecessarily crude and dramatically impacts the scalability of the method. I will therefore keep my score unchanged.

AuthorsOpenReview ↗

Thank you and further clarification on complexity and scalability

Thank you for your follow-up and for acknowledging the improvements in the manuscript.

The use of a complete relation graph is a modeling choice rather than a requirement of RESE. As established in Appendix C.2–C.3, a connected graph is sufficient to identify the relative coordinates up to a common offset. In particular, a spanning tree requires only N−1N-1 edges, compared with N(N−1)/2N(N-1)/2 in a complete graph. If such a graph is selected efficiently before model fitting, the number of pairwise estimation tasks can be reduced from O(N2KR)O(N^2KR) to O(NKR)O(NKR), making linear scaling with NN possible for this stage.

We would also like to clarify that estimating all pairwise relations is a deliberate modeling choice rather than a simple and crude approach. Cross-system relationships can be complex, heterogeneous, and time-varying, making it difficult to determine in advance which pairs are informative. RESE therefore estimates these relations individually and uses historical validation errors to assign reliability weights, allowing more predictable relations to contribute more strongly to the equilibrium estimation. The additional relations also provide more information for reconciling potentially inconsistent forecasts. Using fewer relations may affect estimation accuracy.

Quadratic computational growth is not unique to RESE. Explicit cross-variable modeling in methods such as iTransformer - O(N2d)O(N^2d) [1], TPGNN - O(N2c+qN2d)O(N^2c+qN^2d) [2], and TQNet - O(N2L+NL2)O(N^2L+NL^2) [3] also involves operations that scale quadratically with the number of variables. Other methods face computational costs that grow quadratically along different dimensions, including PatchTST - O(NP2d)O(NP^2d) [4], GTR - O(NL2+Nd2+NLd+NHd)O(NL^2+Nd^2+NLd+NHd) [5], and RAFT - O(M2)O(M^2) [6] for historical retrieval preprocessing). These expressions describe the relevant computational components rather than the full training complexity. Thus, the quadratic computational cost arises from the complete relation graph used in the current implementation, rather than being an inherent requirement of the RESE equilibrium formulation. Using fewer relations may affect estimation accuracy, since the additional edges provide more information for reconciliation.

[1] Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. (2024). iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. International Conference on Learning Representations (ICLR 2024).

[2] Yijing Liu, Qinxian Liu, Jian-Wei Zhang, Haozhe Feng, Zhongwei Wang, Zihan Zhou, and Wei Chen. (2022). Multivariate Time-Series Forecasting with Temporal Polynomial Graph Neural Networks. Advances in Neural Information Processing Systems 35 (NeurIPS 2022).

[3] Shengsheng Lin, Haojun Chen, Haijie Wu, Chunyun Qiu, and Weiwei Lin. (2025). Temporal Query Network for Efficient Multivariate Time Series Forecasting. Proceedings of the 42nd International Conference on Machine Learning (ICML 2025), Proceedings of Machine Learning Research, 267:37797–37814.

[4] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. (2023). A Time Series Is Worth 64 Words: Long-Term Forecasting with Transformers. International Conference on Learning Representations (ICLR 2023).

[5] Fanpu Cao, Lu Dai, Jindong Han, and Hui Xiong. (2026). Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval. International Conference on Learning Representations (ICLR 2026).

[6] Sungwon Han, Seungeon Lee, Meeyoung Cha, Sercan O. Arik, and Jinsung Yoon. (2025). Retrieval Augmented Time Series Forecasting. Proceedings of the 42nd International Conference on Machine Learning (ICML 2025), Proceedings of Machine Learning Research, 267:21774–21797.

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

RESE extends Equilibrium State Estimation (ESE) by removing the dependence on external system attributes. For N systems, RESE forms a pairwise relation series r_{ij,t} = T(s_{i,t}) - T(s_{j,t}) (typically log-ratio), forecasts each pair independently using several relation models selected by horizon-matched validation, weighs each forecast by a reliability score, and reconciles all pairwise forecasts into a latent coordinate via weighted least squares with a zero-sum constraint. The equilibrium shares are combined with a separately forecast aggregate to recover individual forecasts. The method is evaluated on synthetic data, exchange rates, COVID-19, and seven multivariate benchmarks.

Critical strengths

The paper’s main strength is its simple and modular pairwise-to-global forecasting framework. It models pairwise relations directly from observed time series, weights them according to reliability, and reconciles them into a globally consistent relative state that can be combined with different aggregate predictors. It has competitive forecasting performance and provides flexibility across different forecasting settings.

Critical weaknesses

The main weakness is the use of "equilibrium", which is quite misleading. On P3-P4, lines 115-121 and 179-185, "...global relational equilibrium..." and "...The equilibrium state is obtained by...": The use of the term "equilibrium" needs further justification. The cited ESE paper gives "statistical equilibrium" an operational long run interpretation and uses cointegration as an equilibrium criterion. In RESE, however, eq (7) appears to provide a weighted least squares reconciliation of independently forecast pairwise relations, while eq (8) merely maps the reconciled log ratio coordinates to a unit sum share vector. It's hard to come up with an analogous equilibrium condition showing why this reconciled state should be regarded as an equilibrium rather than simply a globally consistent relational configuration. Please formally define what "equilibrium" means in RESE and clarify which property distinguishes Gamma_t^* from an ordinary reconciled/normalised relative state.

Switch to another angle, if RESE uses only the observed multivariate target histories without the attributes, in what sense is it an equilibrium state estimation method rather than a structured multivariate forecasting/reconciliation method? What does the equilibrium interpretation provide beyond forecasting pairwise cross series relations, projecting them onto a globally consistent set of coordinates, and reconstructing the individual series from an aggregate forecast?

The reliability weight w_{ij,t} = exp(-eta sqrt(L*{ij}) / s{ij}) has a free hyperparameter eta >= 0 whose role is not analysed. There is no sensitivity study, no theoretical justification for the specific functional form, and no comparison with simpler alternatives such as inverse validation error or softmax over negative losses.

Questions for authors

P2, lines 96-100, "...let s_{i,t} denote the observed value of system S_i, and define the aggregate as M_t = \sum_{i=1}^N s_{i,t}.": Is s_{i,t} required to be positive, i.e., s_{i,t} \in \mathbb{R}_{+}?

P2, lines 97-101, "...M_t = \sum_{i=1}^N s_{i,t}. The state of the system S_i is represented by its share of the aggregate, \gamma_{i,t} = s_{i,t}/M_t...": What if M_t = 0?

P3, lines 115-117, "...\Gamma_t^* = (\gamma_{1,t}^, \ldots, \gamma_{N,t}^)^\top, \sum_{i=1}^N \gamma_{i,t}^* = 1...": Why is the constraint \sum_{i=1}^N \gamma_{i,t}^* = 1 required to hold?

P3, line 155, "At each validation point, each pairwise relation is modelled using its preceding history.": What is the order or length of the preceding history? What is the computational complexity of this procedure?

P4, lines 167-171, "...RESE jointly reconciles all estimated relations according to their reliability by solving \hat{\mathbf{u}}t = \arg\min{\mathbf{u}} \sum_{i<j} w_{ij,t}(u_{i,t}-u_{j,t}-\hat{r}{ij,t})^2...": Is this optimisation performed while holding w{ij,t} and \hat{r}_{ij,t} fixed?

P4, lines 173-174, "The common offset is therefore fixed by setting \sum_{i=1}^N u_{i,t}=0.": What does this sentence mean? What offset?

P4, lines 180-181, "Under the log-ratio representation, the difference between two coordinates determines their equilibrium share ratio, such that \log(\gamma_{i,t}^/\gamma_{j,t}^) = \hat{u}{i,t}-\hat{u}{j,t}.": In this case, is the transformation T assumed to be the logarithm?

4 comments
AuthorsOpenReview ↗

Response to Weaknesses 1 - 2

We sincerely thank you for the comments. Please see our responses below.

W1. Definition and interpretation of equilibrium

ESE and RESE differ mainly in how equilibrium is estimated. ESE relies on system attributes and uses cointegration as a long-run equilibrium criterion. RESE instead estimates the relative state from historical cross-system dynamics without explicitly modelling on attribution relations.

RESE adopts an optimization-based interpretation of equilibrium, where the global state is characterized by the marginal balance of a collective objective. Related approaches have been studied in machine learning through optimization-defined equilibrium representations [1] and in economics through first-order equilibrium conditions in potential games [2]. In RESE, the first-order condition of Eq.7 is

B⊤W(Bu^−r^)=0,B^\top W(B\widehat{\mathbf{u}}-\widehat{\mathbf{r}})=\mathbf{0},

which means that the reliability-weighted relational discrepancies balance at each system, consistent with the zero-divergence property of graph-based least-squares solutions [3]. The statistical component comes from estimating the relations and their reliability using historical observations. Thus, RESE defines an statistical equilibrium.

[1]Bartunov, S., Fuchs, F. B., & Lillicrap, T. P. (2022). Equilibrium aggregation: Encoding sets via optimization. Proceedings of the 38th Conference on Uncertainty in Artificial Intelligence, PMLR 180, 139–149.

[2]Sandholm, W. H. (2001). Potential games with continuous player sets. Journal of Economic Theory, 97(1), 81–108.

[3]Jiang, X., Lim, L.-H., Yao, Y., & Ye, Y. (2011). Statistical ranking and combinatorial Hodge theory. Mathematical Programming, 127(1), 203–244.

We have added a new figure, Figure 1 in Section 2, to illustrate how the equilibrium state is inferred from the historical evolution of cross-system relations, rather than obtained by simply normalizing the observed system values.

W2. Distinction from structured multivariate forecasting and reconciliation

The main difference lies in the information used to estimate the relative state. ESE relies on system attributes, whereas RESE uses the observed dynamics of the systems themselves. These trajectories reflect the combined effects of factors influencing system behaviour, allowing RESE to estimate cross-system relationships without explicitly modelling those attributes.

As described in Section 3 (lines 125–199), RESE forecasts pairwise relations, reconciles them into a common relative state, and combines this state with a separately forecast aggregate to reconstruct individual system forecasts (Eq. 9). This retains ESE's aggregate–allocation representation while replacing attribute-based estimation with historical relational information.

AuthorsOpenReview ↗

Response to Weakness 3 & Question 1

W3. The reliability-weighting function and the role of η\eta need justification and sensitivity analysis.

In RESE, the reliability weights are estimated from historical validation errors, while η\eta controls their sensitivity to these errors. We use η=1\eta=1 as a natural reference setting, since the validation errors are already normalized by the historical standard deviation of each relation. Under this choice, a one-unit increase in normalized error reduces the relative reliability weight by a factor of e−1e^{-1}. This gives the weighting scheme an interpretable decay rate without introducing an additional scale adjustment or requiring dataset-specific tuning.

We examined (η∈0,0.25,0.5,1,2,4\eta\in{0,0.25,0.5,1,2,4}) on the 16-currency exchange-rate dataset, using the one-step forecasting setting of Table 2. We also compared the exponential weighting scheme with uniform, inverse-error, and softmax weighting.

Weighting schemeRMSE*MAE
Uniform (η=0\eta=0)1.2070.612
Exponential (η=0.25\eta=0.25)1.1790.582
Exponential (η=0.5\eta=0.5)1.1740.597
Exponential (η=1\eta=1)1.1800.589
Exponential (η=2\eta=2)1.2230.633
Exponential (η=4\eta=4)1.2740.676
Inverse-error weighting1.2460.648
Softmax over raw losses1.1860.573

The RMSE* values vary within a narrow range for η∈[0,1]\eta\in[0,1] whereas η=2\eta=2 and η=4\eta=4 produce higher errors. The results suggest that stronger reliability weighting does not necessarily improve forecasting accuracy.

For the alternative weighting schemes, let eij=L∗ij/sije_{ij}=\sqrt{\mathcal L^*{ij}}/s{ij} denote the normalized validation error, where sijs_{ij} is the historical standard deviation of the relation. Inverse-error weighting is defined as

wijinv=1eij,w_{ij}^{\mathrm{inv}}=\frac{1}{e_{ij}},

where (e_{ij>0) prevents division by zero. It performs worse than the default exponential scheme on both metrics. The exponential function gives positive, bounded weights that decrease smoothly with validation error, while inverse-error weighting can assign much greater relative influence to relations with near-zero errors.

For softmax weighting, let (E) denote the set of pairwise relations retained after the persistence-based screening described in Appendix G.3. We apply softmax to the negative validation losses: wijinv=1eij,w_{ij}^{\mathrm{inv}}=\frac{1}{e_{ij}}, where eij>0e_{ij}>0 prevents division by zero. It performs worse than the default exponential scheme on both metrics. The exponential function gives positive, bounded weights that decrease smoothly with validation error, while inverse-error weighting can assign much greater relative influence to relations with near-zero errors.

For softmax weighting, let EE denote the set of pairwise relations retained after the persistence-based screening described in Appendix G.3. We apply softmax to the negative validation losses: wijsoftmax=exp⁡(−Lij∗)∑(a,b)∈Eexp⁡(−Lab∗)w_{ij}^{\mathrm{softmax}}=\frac{\exp(-L_{ij}^*)}{\sum_{(a,b)\in E}\exp(-L_{ab}^*)} Softmax gives a slightly higher RMSE* than the default exponential scheme (1.186 versus 1.180), but a lower MAE* (0.573 versus 0.589). Unlike RESE's exponential weighting, which uses normalized validation errors, this alternative uses raw losses and is therefore sensitive to their scale.

These results show that RESE is relatively insensitive to moderate changes in η\eta in terms of RMSE*, while larger values can reduce accuracy. The alternative weighting schemes do not consistently outperform the default, although the best configuration depends on the evaluation metric.

The sensitivity results show that moderate changes in η\eta have a limited effect on RMSE*, while larger values can reduce accuracy. Although η=1\eta=1 is not necessarily optimal, it provides a consistent reference configuration across datasets.


Q1. Is si,ts_{i,t} required to be positive?

Strict positivity is not a general requirement of RESE. The definition in Section 2 describes the multiple systems ensemble and its aggregate, rather than imposing a nonnegativity constraint on the RESE estimator. In the multiple systems experiments considered in this paper, the observed targets are nonnegative. The synthetic and exchange-rate series take positive values, while the COVID-19 data consist of daily case counts that may include zeros.

More generally, RESE defines pairwise relations as rij,t=T(si,t)−T(sj,t).r_{ij,t}=\mathcal{T}(s_{i,t})-\mathcal{T}(s_{j,t}).

The relation estimation and reconciliation steps do not require the original observations to be nonnegative, provided that the chosen transformation is defined for those observations. This also allows RESE to be applied to conventional multivariate forecasting benchmarks, where variables may contain negative values. As stated in Section 4.2 (lines 299–323), we use the signed-log transformation for these benchmarks. Other representations, including asinh, are described in Appendix A.3 (lines 733–746), with the reconstruction procedure in Appendix I.1 (lines 1182–1225).

AuthorsOpenReview ↗

Responses to Questions 2 - 5

Q2. What happens if Mt=0M_t=0?

In the nonnegative multiple-system setting, Mt=0M_t=0 means that all systems have zero observations at time t. In this case, the observed system shares (γi,t=si,t/Mt)(\gamma_{i,t}=s_{i,t}/M_t) in Eq.2 are undefined.

However, RESE does not estimate the future equilibrium state by directly dividing individual observations by the current aggregate. Instead, it forecasts the pairwise relations from their historical trajectories and reconciles these forecasts to obtain the equilibrium coordinates. As described in Appendix A.1 (lines 713–721), log⁡(s+ϵ)\log(s+\epsilon) is used to handle zero-valued observations. The predicted relations can therefore still be constructed and estimated when the current aggregate is zero.

The equilibrium shares are subsequently recovered from the reconciled coordinates through Eq.8. Hence, a zero current aggregate does not necessarily prevent RESE from estimating a future equilibrium state.

The final forecasts are obtained by combining this state with the separately predicted aggregate:

s^t+h=M^t+hΓt+h∗.\hat s_{t+h} = \hat M_{t+h}\Gamma_{t+h}^*.

This reconstruction is described in Section 3.5 (lines 189–199) and Appendix D.3 (lines 1014–1025). The alternative expression involving θt+hMt\theta_{t+h}M_t applies when the aggregate ratio is defined; Appendix D (lines 963–968) explicitly specifies Mt>0M_t>0 for this ratio.

Q3. Why is ∑iγi,t∗=1\sum_i\gamma^*_{i,t}=1 required?

This constraint is inherited from the multiple-system formulation introduced in ESE (Xu et al., 2026a), rather than being introduced specifically for the reconciliation step in RESE.

In ESE, the systems are considered as members of a defined ensemble, with each system represented by its proportion of the aggregate. As stated in Section 2, Definition 2 of ESE, a change in one system's proportion must be offset by the net changes in the proportions of the remaining systems:

Δγi,t=−∑j≠iΔγj,t.\Delta\gamma_{i,t}=-\sum_{j\ne i}\Delta\gamma_{j,t}.

This follows from the requirement that the proportions of all systems in the ensemble sum to one. It expresses the complementary nature of their relative changes, rather than a restriction on their absolute values.

RESE retains this formulation. Section 2 (lines 095–103) defines the observed system state through its shares of the aggregate, while lines 115–120 extend the same normalization to the estimated equilibrium state.

The difference lies in how the equilibrium shares are estimated. Whereas ESE uses system attributes, RESE infers the relative configuration from predicted pairwise relations. Under the log-ratio representation, Eq. (8) reconstructs the equilibrium shares through softmax normalization, preserving the same ensemble-level constraint.

This also ensures that the reconstructed system forecasts sum to the predicted aggregate, as described in Section 3.5 (lines 189–199).

Q4. Length of history & complexity

RESE supports both expanding and fixed-length sliding windows (Eq.48). For a fixed input length (L), the same (L) observations are used at each validation origin and during the final refit. The input lengths are reported in the experimental tables (e.g., (L=20) for the synthetic dataset and (L=100) for exchange rates).

Regarding computational complexity, there are N(N−1)/2N(N-1)/2 candidate pairs for NN systems. Given KK candidate relation models and RR validation origins, the procedure requires O(N2KR)O(N^2KR) candidate trainings, excluding model-specific fitting costs (Appendix G.5, Eq.54). These training are independent across pairs and validation origins and can be parallelized.

Table 5 in Section 4.3 reports the runtime of relation estimation and equilibrium reconciliation separately. The main computational cost comes from pairwise model fitting rather than the subsequent equilibrium reconciliation.

Q5. Are wij,tw_{ij,t} and r^ij,t\hat r_{ij,t} held fixed during Eq.7?

The pairwise relation forecasts and their reliability weights are obtained from historical relation modelling and validation, as described in Section 3.2 (lines 148–163) and Appendix B.2–B.3 (lines 858–885). They are treated as fixed inputs during each individual solution of Eq.7, where the optimization estimates the global coordinate vector uu (Section 3.3, lines 167–176).

However, Eq.7 is not solved only once. As specified in Appendix G.4 (lines 1140–1146), RESE produces a sequence of relation forecasts and performs reconciliation at each prediction horizon (h=1,…,Hh=1,\ldots,H). The relation forecasts may therefore differ across prediction steps. At subsequent forecast origins, the selected relation models are also refitted using the available historical observations (Appendix G.1, lines 1099–1109). The reliability weights are determined by historical validation and may be updated when that validation is repeated, but they are not re-estimated within each reconciliation solve. Thus, the fixed-input assumption applies to an individual optimization, rather than to the entire forecasting procedure.

AuthorsOpenReview ↗

Responses to Questions 6 - 7

Q6. What does “common offset” mean?

The common offset refers to an additive constant shared by all relational coordinates. In Eq.7, the estimated pairwise relations depend only on the differences ui,t−uj,tu_{i,t}-u_{j,t}. Adding the same constant (c) to every coordinate leaves these differences unchanged: (ui,t+c)−(uj,t+c)=ui,t−uj,t.(u_{i,t}+c)-(u_{j,t}+c)=u_{i,t}-u_{j,t}. Consequently, the pairwise relations determine only the relative positions of the systems, not the absolute level of their coordinates. We impose ∑iui,t=0\sum_i u_{i,t}=0 to remove this ambiguity and obtain a centered representation. This is consistent with the centered coordinate definition in Eq.4, Section 3.1 (lines 138–144). The corresponding condition is established in Appendix C.2 (lines 910–920) for a connected relation graph with positive weights.

Under the log-ratio representation, this centering does not affect the estimated equilibrium shares, since the normalization in Eq.8 is invariant to a common additive shift, as explained in Appendix A.2 (lines 723–731).

Q7. Does Eq.8 assume T(s)=log⁡s\mathcal T(s)=\log s?

Yes. Note that the logarithmic transformation is used in the main multiple-system forecasting formulation, rather than being a general requirement of RESE. Since RESE was developed primarily for this setting, the main text focuses on the log-ratio representation, which provides a direct relationship between pairwise relations and system shares. Specifically, with T(s)=log⁡s\mathcal T(s)=\log s, the differences between reconciled coordinates correspond to the log-ratios of equilibrium shares, allowing the use of Eq.8, as explained in Section 3.4 (lines 180–185) and Appendix A.1–A.2 (lines 711–731).

The underlying RESE procedure is not restricted to this logarithmic transformations. Eq.3 defines pairwise relations using a general transformation T\mathcal T, and the reconciliation in Eq.7 operates on these relations. As described in Appendix A.3 (lines 733–746), alternative representations include asinh, signed-log, and additive coordinates. These representations use the same relation estimation and reconciliation framework, but require a different reconstruction procedure from the softmax formulation in Eq.8, as detailed in Appendix I.1 (lines 1182–1225).

This distinction is reflected in Section 4.2 (lines 299–323), where RESE is applied to conventional multivariate forecasting by treating each variable as a system. The signed-log transformation is used in these experiments to accommodate zero and negative observations.


We hope the above response has fully addressed your comments. If so, we would greatly appreciate it if you could reconsider your rating.

Reviewer_AMBxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce RESE (Relational Equilibrium State Estimation), a method for forecasting related time series. It forecasts the ratio between each pair of series and adjusts conflicting predictions so that they agree. It converts the adjusted predictions into shares of a separately forecast total, then multiplies each share by that total to obtain individual forecasts. Unlike ESE, RESE uses historical observations without external attributes. The reported accuracy is competitive, but fitting a model for every pair becomes expensive as the number of series grows.

Critical strengths

The method achieves competitive forecasting accuracy using conventional statistical models to predict pairwise relations, without requiring external attributes. The evaluation provides evidence that the approach can be useful across different settings. The separation between relation estimation and aggregate forecasting also allows different aggregate predictors to be used within the same framework. In addition, the reconciliation step has an explicit objective for resolving conflicting pairwise predictions, and reconstruction ensures that individual forecasts sum to the predicted aggregate.

Critical weaknesses

The meaning of “equilibrium” is not clearly explained especially if reader is unfamilier with ESE. the experiments do not isolate the gains from each step( forecasting relative relationships, model selection, reconciliation, or aggregate forecasting.) Important steps are distributed across the appendices, making the method difficult to understand. There is no clear explanations of why forecasts should improve. The paper needs to support when pairwise forecasting should help and when it is likely to fail and how it compares to residual forecasting.

Questions for authors

Why forecast every pair’s ratio rather than each series’ share of the total? What evidence supports this choice?

What evidence shows that reliability weighting improves forecasting accuracy beyond the pairwise predictions themselves?

3 comments
AuthorsOpenReview ↗

Response to Weaknesses 1 - 3

We sincerely thank you for the review. Please see our responses below.

W1. The meaning of “equilibrium” is not clearly explained, especially if the reader is unfamiliar with ESE.

The meaning of equilibrium could be explained without requiring prior knowledge of ESE. In RESE, equilibrium is also a statistical estimation of equilibrium state, inferred based on pairwise relations between systems rather than attribution relations between systems in ESE.

As described in Section 2 (lines 063–067 and 115–120), the equilibrium state represents the system-wide relative configuration implied by the relational structure. Section 3 (lines 125–132) further clarifies that this state is estimated from historical relations and may change over time.

More specifically, individual pairwise forecasts may be mutually inconsistent because they are estimated separately. RESE resolves these inconsistencies through the reliability-weighted least-squares formulation in Eq.7, obtaining a common set of relative coordinates. Under the log-ratio representation, these coordinates are converted into equilibrium shares through Eq.8.

The mathematical basis of this procedure is detailed in Appendix C, particularly the graph formulation and cycle-consistency property (lines 897–932). The latter explains why relations derived from a common coordinate vector are globally consistent.

We have also added a new illustration (Figure 1 in Section 2) to clarify the distinction between the observed system state (ST) and the estimated equilibrium state (ES), and to illustrate how historical cross-system relations inform equilibrium estimation.

W2. The experiments do not isolate the gains from each step (forecasting relative relationships, model selection, reconciliation, or aggregate forecasting).

We appreciate this insightful point. The four steps serve different purposes in RESE, and we clarify their roles with supporting evidence below. Experiments are added to show isolated effects of the four steps in our responses to your Question 2.

  1. Pairwise relation forecasting. This step models the evolving relative dynamics between systems. We have conducted an additional comparison against share persistence, with results presented in our response to Q1 below. RESE achieves lower RMSE at (h=5) and (h=10) across all five forecasting origins, supporting the benefit of forecasting changes in relative relationships.

  2. Model selection. We considered possible ablation settings to evaluate model selection separately, but neither provides a clean separation from pairwise forecasting. Fixing a common model across all pairs would restrict their ability to capture different dynamics, affecting both the relation forecasts and their estimated reliability. Removing pairwise forecasting altogether would eliminate the relations on which model selection operates. Model selection is therefore closely integrated with relational forecasting rather than functioning as an independent processing step.

  3. Reconciliation. Independently predicted pairwise relations may be inconsistent. Eq.7 combines them into a common relative state. To examine the role of reliability weighting in this process, we conducted an additional ablation comparing reliability and uniform weights while keeping the pairwise predictions and aggregate forecasts unchanged. The results, reported in our response to Q2 below, show improved forecasting accuracy with reliability weighting.

  4. Aggregate forecasting. This step estimates the overall system scale, while the equilibrium shares determine its allocation across systems (Eq. 9). Tables 1–3 compare RESE using different aggregate predictors, demonstrating the flexibility of this component without changing the relative-state estimation procedure.

W3. Important steps are distributed across the appendices, making the method difficult to understand.

Section 3 describes the main forecasting procedure. The details of the relation-model library, historical model selection, reconciliation details, and aggregate predictor selection are explained respectively in Appendices B, C, and D.

We have revised Section 3 to describe these steps in a more connected manner, particularly the relationship between pairwise model selection, reliability estimation, global reconciliation, and aggregate forecasting. The revision clarifies the existing method without changing the formulation.

AuthorsOpenReview ↗

Response to Weaknesses 4 - 5

W4. There are no clear explanations of why forecasts should improve.

Our hypothesis of RESE improving forecast is that the historical relations among systems may contain useful information about their future relative configuration. Rather than forecasting every system individually, we separate the forecasting task into two parts in RESE: (1) estimating the future relative state and (2) forecasting the aggregate.

The former is obtained from predicted pairwise relations, while the latter determines the overall scale of the reconstructed forecasts. This decomposition is detailed in Eq.9, Section 3.5 (lines 189–199).

There are at least two reasons why the relational structure is useful.

First, different system pairs may exhibit different temporal dynamics. RESE models these relations separately and allows each pair to select a forecasting model based on historical validation performance (Section 3.2, lines 148–163; Appendix B.1–B.2, lines 758–871).

Second, independently predicted relations don't have to agree. The reconciliation step combines them into a common relative state, with greater influence assigned to relations receiving higher reliability weights. This provides a consistent estimate of the system-wide relative configuration, as described in Section 3.3 (lines 167–176).

These are the rationale behind our RESE method. Whether a separate system improves out-of-sample forecasts depends on the predictability of the relations and the quality of the aggregate forecasts introduced by the system.

The existing experiments provide evidence for the proposed framework. For example, Table 1 (lines 219–235) reports the lowest RMSE for standalone RESE on the 10-system synthetic dataset. Table 2 (lines 246–265) shows improvements over most standalone baselines on the exchange-rate dataset, while Table 3 (lines 270–297) reports results for the 79-city COVID-19 dataset.

W5. The paper needs to support when pairwise forecasting should help and when it is likely to fail and how it compares to residual forecasting.

W5(a). When pairwise forecasting should help or fail

The effectiveness of RESE depends on whether the historical cross-system relations provide useful information for forecasting future relative states. We have investigated this question through the controlled system-dependent experiment in Section 4.3, Table 8 (lines 415–433).

As the realised mean pairwise correlation increases from 0.44 to 0.87, standalone RESE's RMSE changes from 0.294 to 0.234. At the highest correlation level, RESE achieves an RMSE of 0.234, compared with 0.238 for DLinear. However, DLinear performs better at the lower correlation levels. The results therefore suggest that RESE can benefit from stronger cross-system dependence, while its advantage becomes less evident when the shared relational information is weaker.

This observation is consistent with the limitations already acknowledged in Section 6 (lines 478–482): when cross-system relations are weak, unstable, or supported by limited history, the estimated equilibrium state may provide little additional benefit over individual forecasting.

We have also examined the sensitivity of RESE to observation noise, missing observations, and heterogeneous local shocks in Section 4.3. Table 6 (lines 381–396) shows that RESE remains relatively stable under moderate noise and missing values, whereas stronger and more widespread perturbations lead to higher errors. Similarly, Table 7 (lines 398–412) shows that mild local shocks have limited effects, while stronger and longer-lasting shocks gradually degrade forecasting performance.

Together, these results provide empirical evidence about the conditions under which RESE is more useful and where its performance may deteriorate.

W5(b). Comparison with residual forecasting

RESE and residual forecasting differ in their prediction targets and reconstruction procedures. Residual forecasting typically models deviations from a baseline prediction or a common trend to correct the final forecasts. RESE, by contrast, predicts the evolving relationships between systems and reconciles them into a consistent relative state. Individual system forecasts are then obtained by combining the estimated equilibrium shares with the aggregate forecast (Eq. 9).

In addition, our experiments do involve methods related to residual forecasting, including NLinear and DLinear. NLinear predicts changes relative to the last observed value, while DLinear separately models trend and remainder components. Both are evaluated as standalone baselines and in combination with RESE (Tables 1–3).

AuthorsOpenReview ↗

Responses to Questions 1 & 2

Q1. Why forecast every pair's ratio rather than each series' share of the total? What evidence supports this choice?

To address this question, we considered three possible approaches to obtaining future system shares without explicitly forecasting pairwise ratios.

  1. One approach is to forecast each system independently and calculate its share from the predicted values:

γ^i,t+h=s^i,t+h∑js^j,t+h.\hat\gamma_{i,t+h}=\frac{\hat s_{i,t+h}}{\sum_j\hat s_{j,t+h}}.

If the aggregate is also obtained from these predictions, reconstruction gives (M^t+hγ^i,t+h=s^i,t+h\hat M_{t+h}\hat\gamma_{i,t+h}=\hat s_{i,t+h}), where (M^t+h=∑js^j,t+h\hat M_{t+h}=\sum_j\hat s_{j,t+h}). Thus, this approach simply reproduces the original baseline forecasts.

  1. Another approach is to calculate historical shares and forecast each one independently, for example, using separate ARIMA models. However, each share depends on the aggregate of all systems, so its variation may reflect changes in other systems rather than its own. This makes the underlying dynamics harder to identify from individual share histories, whereas RESE models pairwise relations separately, allowing different system pairs to provide complementary information about the evolving relative state. Moreover, independently predicted shares may not sum to one. Subsequent normalization can propagate prediction errors across systems: an inaccurate prediction for one system may alter the final shares of others, even when their original predictions were accurate.

  2. We also conduct an additional experiment following the naive share-based approach in ESE (Xu et al., 2026a, Appendix N.1). This approach retains the observed shares at each forecast origin, while RESE forecasts future pairwise relations.

SampleMethodDateh=1h=2h=5h=10
1Direct Share14-Jun-246.446.9412.4732.84
Pairwise Ratio (RESE)14-Jun-245.576.2411.3714.83
2Direct Share17-May-246.228.5216.7736.54
Pairwise Ratio (RESE)17-May-246.387.9510.8415.17
3Direct Share22-Apr-245.529.3220.9724.74
Pairwise Ratio (RESE)22-Apr-245.678.4611.4114.48
4Direct Share22-Mar-246.347.5614.7435.02
Pairwise Ratio (RESE)22-Mar-245.938.2411.3714.98
5Direct Share23-Feb-245.948.6429.5739.41
Pairwise Ratio (RESE)23-Feb-245.486.3110.6415.37

The results are shown in the table above as well as in the updated manuscript. RESE achieves lower RMSE in all five samples at (h=5) and (h=10), while results at shorter horizons are mixed. These results indicate that forecasting pairwise dynamics can improve upon retaining the current share distribution, particularly at longer forecast horizons.

Q2. What evidence shows that reliability weighting improves forecasting accuracy beyond the pairwise predictions themselves?

Pairwise relations can differ substantially in their predictive reliability, and some may provide limited information about future relative states. RESE therefore uses historical validation errors to determine how much influence each relation should have in the global reconciliation. As defined in Eq.6, relations with lower normalized validation errors receive higher weights, while less reliable predictions have reduced influence in Eq.7.

To assess whether this weighting improves forecasting accuracy, we conducted an additional ablation on the Exchange Rates dataset. We compared the original RESE with a uniform weight, keeping the pairwise predictions, aggregate forecasts, and reconciliation framework unchanged. Any uniform weight value yields the same solution, as uniformly scaling the weights only multiplies the least-squares objective in Eq.7 by a constant and does not change its minimizer.

MethodRMSEMAERMSE*MAE*
RESE (Uniform Weights)6.2376.0491.2070.612
RESE (Reliability Weights)5.5725.3311.1800.589

The results show that pairwise relations differ in predictive reliability, and treating them equally can give less reliable forecasts undue influence during reconciliation. Reliability weighting reduces this influence, leading to more accurate forecasts on the Exchange Rates dataset.


We hope the above response has fully addressed your comments. If so, we would greatly appreciate it if you could reconsider your rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.