acceptodds
Under review as a conference paper at ICLR 2027

FLAG: Factor-Lag Relational Context for Adaptive Group Communication in Multivariate Time Series Forecasting

Abstract

Effectively modeling and using cross-variable dependencies is an important route to better multivariate time series forecasting. Prior work has explored several mechanisms for interaction among variables, but it remains challenging to represent lagged relations compactly while adjusting how those relations are used according to the current input. In particular, cross-variable dependencies can recur across observation windows, yet their contribution to a forecast depends on the current state. If this difference is not taken into account, information aggregation can mix cross-variable information that is useful for the current input with information that is not. To address this problem, we propose FLAG, a unified framework that jointly models lagged relations and performs input-adaptive information exchange within the same representation. FLAG first combines shared low-rank latent factors with lag-specific factor interactions, incorporating other variables' histories into each variable's representation and forming a relational context. This design retains lag-specific interaction detail without constructing dense variable-pair matrices. Group communication then produces input-dependent aggregation weights and messages from this context and updates the same representation, so that the shared relations take part in the forecast according to the current state. With the other model and temporal dimensions held fixed, both the number of parameters and the computation grow linearly with the number of variables. Experiments on eight benchmarks show that FLAG achieves competitive forecasting performance. Code is available at https://anonymous.4open.science/r/FLAG-DE06.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Time Series Forecasting, Time Series Analysis, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_F6svOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FLAG, a framework for multivariate time series forecasting that unifies (i) compact modeling of lagged cross-variable relations and (ii) input-adaptive information exchange within a shared representation. The core "factor-lag" operator connects variables through shared low-rank read/write bases U, V in R^{Nxd} and a full, lag-specific factor transition B_l in R^{rxr}, giving the interaction A_l = V B_l U^T with self-terms removed; messages are aggregated over valid patch offsets into a relational context C^(0). A "context-adaptive group communication" module then performs N->K->N soft-group aggregation/redistribution whose weights are recomputed per input/patch/channel, refining C^(0) through a residual update. Parameters and compute grow linearly in the number of variables N for fixed structural hyperparameters (2Nr + Lc r^2 factor params, up to LNK membership params). The authors evaluate on eight benchmarks (Weather, Traffic, Electricity, Solar-Energy, four ETT) with T=96 and H in {96,192,336,720} against eleven baselines, plus a five-dataset component ablation, lag-setting comparisons, factor-transition visualizations, and a look-back study.

Critical strengths
  1. Compact, parameter-efficient lag-relation operator. Representing per-offset interactions via shared read/write bases plus full r x r lag transitions avoids O(N^2) variable-pair matrices while retaining lag specificity, and the zero-initialization lets the model fall back to a channel-independent backbone — enabling a clean ablation.
  2. Unified "construct-then-adapt" design that lets the same relational context both determine aggregation weights and receive updates, tying relation modeling to input-adaptive use under one objective.
Critical weaknesses
  1. Marginal, inconsistent empirical gains without significance evidence. FLAG is best only on Weather (avg MSE) and second-best elsewhere by thin margins; on Traffic it is worse than SOFTS (0.424 vs 0.415) and on Electricity worse than FACT (0.167 vs 0.160). Results are averaged over only three seeds with no standard deviations or significance tests, so several deltas (e.g., Weather 0.237 vs GTR 0.241) are plausibly within seed noise. The Solar-Energy MAE advantage (0.209 vs 0.248, ~15.7%) is anomalously large and inconsistent with MSE, where FLAG is only second — a metric-consistency concern. This directly undercuts the headline claim.
  2. The ablations do not cleanly support the design. On Traffic, factor-lag alone worsens MSE (0.444 -> 0.454); on Weather, adding group communication worsens MSE (0.146 -> 0.150). Contributions are non-additive and dataset-specific, so the value of the "unified" framework is mostly shown on Traffic/Electricity. With no theoretical grounding or identifiability analysis, the central claim that combining relations with adaptive use is beneficial is under-supported.
Questions for authors
  1. Please report per-seed variance / confidence intervals and significance for the main results. Are the headline deltas (Weather, Traffic) statistically significant?
  2. Why is the Solar-Energy MAE improvement so large relative to the small MSE gain? Please confirm the metric/target computation is identical across baselines.
  3. How are r, K, and L chosen, and how sensitive is performance to them? A sensitivity study is missing.
  4. Given factor-lag alone hurts on Traffic and group communication hurts on Weather, how do you substantiate a unified contribution rather than a per-dataset selection effect?
  5. Provide runtime/memory comparisons to support the linear-N efficiency claim, and evidence that the adaptive aggregation weights and memberships actually vary meaningfully across inputs.
3 comments
AuthorsOpenReview ↗

Response to W1, W2

Thank you for your insightful feedback. We respond to each point below.

Response to W1

Revised Table 1 averages over four forecast horizons for each dataset and reports the mean and standard deviation for FLAG over three fixed seeds. Model rankings vary by dataset and metric. FLAG ranks first in 9 of 16 comparisons, including ties. Its average MSE on Traffic is 0.423 versus 0.415 for SOFTS, and on Electricity it is 0.166 versus 0.160 for FACT. On Weather, FLAG has the lowest average MAE (0.262). Its average MSE of 0.241 is higher than TimeFilter’s 0.239 and matches the reviewer’s GTR reference of 0.241 at three decimals. The overall ranking supports FLAG’s performance across tasks. Q1 reports uncertainty and test results, and Q2 addresses Solar-Energy.

Response to W2

FLAG connects two stages of forecasting. Factor-Lag (D) learns lagged relations shared across windows. Group Communication (G) uses the current window to determine which messages receive more weight. To test how these stages work together, we held the encoder, forecasting head, and training settings fixed within each dataset at T=H=96 and compared Base, D alone, G alone, and the full model (Section 4.2, Figure 3). The table reports mean test MSE over three seeds.

DatasetBaseD onlyG onlyFull FLAG
ETTm10.3090.3080.3110.308
Weather0.1720.1540.1580.153
Electricity0.1670.1380.1450.138
Traffic0.4430.4520.3900.389

Traffic illustrates why the two stages should be evaluated together. D alone gives an MSE of 0.452, above Base at 0.443. When G uses D’s relations in the current window, the full model reaches 0.389, slightly below G alone at 0.390. On Weather, D alone already reaches 0.154, and adding G gives 0.153. The same D→G design improves on Base across all four datasets, from 7 variables on ETTm1 to 862 on Traffic. The contribution of each stage varies, while their combination allows one architecture to accommodate datasets with different temporal patterns and variable counts. A unique physical interpretation of individual latent factors would require additional identifiability assumptions.

AuthorsOpenReview ↗

Response to Q1, Q2, Q3

Response to Q1

We quantified variation across training seeds for all 32 forecasting tasks on the eight datasets. For seeds 2021, 2022, and 2023, we first averaged the unrounded results over the four horizons within each seed. We then computed the sample variance and a two-sided 95% Student t confidence interval for the mean (2 degrees of freedom). Table 1 and Appendix K, Table 6 report means and standard deviations. The table below gives the variances and intervals for the four-horizon averages under the selected configurations.

DatasetMean MSEMSE variance (×10⁻⁶)MSE 95% CIMean MAEMAE variance (×10⁻⁶)MAE 95% CI
ETTh10.43340.391[0.4318, 0.4349]0.42550.429[0.4239, 0.4271]
ETTh20.36042.174[0.3568, 0.3641]0.38480.133[0.3839, 0.3857]
ETTm10.37420.695[0.3721, 0.3763]0.38030.376[0.3788, 0.3819]
ETTm20.27500.491[0.2733, 0.2767]0.31690.831[0.3146, 0.3192]
Weather0.24092.070[0.2373, 0.2444]0.26160.557[0.2597, 0.2634]
Solar-Energy0.18811.387[0.1852, 0.1910]0.20823.004[0.2039, 0.2125]
Electricity0.16640.941[0.1640, 0.1688]0.26165.305[0.2558, 0.2673]
Traffic0.42280.385[0.4212, 0.4243]0.27740.963[0.2749, 0.2798]

For the comparisons raised by the reviewer, we treated each reported baseline score as a fixed reference and applied a two-sided exact sign test to the three FLAG seed results. On Weather, 2 of 3 MSEs were below GTR’s 0.241 (p=1.00). On Traffic, all 3 were above SOFTS’s 0.415 (p=0.25). For Weather MAE, all 3 were below TimeFilter’s 0.269 (p=0.25). With three seeds, 0.25 is the smallest attainable two-sided p-value for this test.

In Table 1, FLAG achieves the lowest MAE averaged over the four forecast horizons on six of eight datasets. On the other two, the gap to the best MAE is small, at 0.006 on Electricity (0.262 vs. 0.256) and 0.009 on Traffic (0.277 vs. 0.268). FLAG also achieves the lowest average MSE and MAE on ETTh2, ETTm1, and Solar-Energy. Taken together, these results show that the same architecture forecasts well across datasets with different temporal patterns and widely varying numbers of variables.

Response to Q2

FLAG has the lowest Solar-Energy error on both metrics in revised Table 1. Using the displayed three-decimal values, its average MSE of 0.188 is 5.1% lower than the joint runners-up FACT and TQNet (0.198). Its average MAE of 0.208 is 10.7% lower than the runner-up DUET (0.233). These percentages use different reference models. Against DUET on both metrics, FLAG reduces MSE by 20.7% and MAE by 10.7%. The apparent disparity therefore depends on the comparator.

ModelAvg MSEAvg MAE
FLAG0.1880.208
FACT0.1980.260
TQNet0.1980.258
DUET0.2370.233

FLAG computes MSE and MAE on the same standardized predictions and targets across test windows, forecast steps, and variables. MAE averages absolute errors, whereas MSE averages squared errors and gives larger errors more weight.

We compare baselines at a common input length of 96 and forecast horizons of 96, 192, 336, and 720. TimeFilter scores come from Table 8 of its paper, DUET scores from the authors’ fixed-input-length results, and the remaining baselines from recorded runs at the same forecasting settings. Each source uses its own evaluation implementation.

Response to Q3

r is the number of latent factors, K the number of soft groups, and L the communication depth. We compared a limited set of candidate configurations and chose the best-performing candidate in that comparison for each dataset, using performance across all four horizons (Appendix A, Table 3). In the main experiments, we fixed the lag-bank size at 12 for every dataset. This covers all patch offsets from 0 through 11.

DatasetVariablesrKL
ETTh1 / ETTh2 / ETTm1 / ETTm27422
Weather211624
Solar-Energy1371622
Electricity3211621
Traffic8621612

Appendix B, Figure 7 varies r, K, and L one at a time on five datasets at H=96, averaging each setting over three seeds. On Traffic, increasing r from 4 to 12 lowers MSE from 0.408 to 0.388, while r=32 raises it to 0.427. All five tested K values on ETTm1 give 0.308, and K=1 and K=2 both give 0.389 on Traffic. Communication depth also has a useful range. On Traffic, L=3 gives 0.387 and L=5 gives 0.402.

We did not tune the lag-bank size in the main experiments. The separate lag ablation (Section 4.3, Figure 4) shows that adjusting it can improve individual tasks. On Solar-Energy at H=336, reducing the bank from 12 to 6 lowers three-seed mean MSE from 0.196 to 0.193.

AuthorsOpenReview ↗

Response to Q4, Q5

Response to Q4

FLAG separates the modeling of cross-variable relations from their use in forecasting. All eight datasets in the main experiments use the same architecture. The factor–lag module (D) first constructs the relational context, and the group communication module (G) uses that context adaptively for the current input. Figure 3 in Section 4.2 reports mean test MSE over three seeds at H=96. On Traffic, adding D alone raises MSE from 0.443 for Base to 0.452. G alone gives 0.390, and the full model gives 0.389. On Weather, D alone gives 0.154 and the full model gives 0.153. On ETTm1 and Electricity, the full model ties D alone to three decimal places. Across all four datasets, the same D→G architecture improves on Base and matches or slightly improves on the better single module at the reported precision. These results support our aim of using one architecture across forecasting tasks with different temporal patterns and numbers of variables.

Response to Q5

With r, K, L, lag count, window lengths, and feature widths fixed, the factor relations use 2Nr+Lcr22Nr+L_c r^2 parameters and group memberships use at most LNKLNK. Operations run through factors and groups, avoiding a dense N-by-N variable-pair matrix (Section 3.5). Section 4.5, Figure 6 and Appendix J, Figure 11 report six models on the same H100 at T=H=96, batch size 8, and FP32. Training time is averaged over one complete epoch, memory is peak allocated training memory, and inference latency is measured after warm-up.

ModelETTm1 train (ms/iter)ETTm1 infer (ms/batch)ETTm1 peak (MiB)Traffic train (ms/iter)Traffic infer (ms/batch)Traffic peak (MiB)
FLAG8.471.32124.0823.258.082,519.81
SOFTS6.320.8868.6912.744.12654.60
TimeFilter7.261.05109.9424.999.212,578.71
DUET17.063.49126.1576.6623.636,025.89
TimeMixer18.473.5882.8653.9615.305,744.55
PatchTST9.271.16179.8381.2837.626,452.53

From ETTm1 (7 variables) to Traffic (862), FLAG’s training time per iteration increases 2.75-fold, inference latency 6.14-fold, and peak training memory 20.31-fold. SOFTS remains lighter on both. On Traffic, FLAG uses 69.7% less training time per iteration and 58.2% less peak memory than DUET, with 65.8% lower inference latency and reference MSE of 0.389 versus 0.407. The two measurements describe practical costs under the respective dataset configurations. The linear order follows from the fixed-parameter analysis.

The input-adaptive quantity is the aggregation weight α. Group memberships π are learned soft assignments shared across windows. Each window’s messages determine how much each variable contributes within its groups, so α varies with the input, patch, and message channel (Section 3.4). We froze three H=96 checkpoints per dataset and kept each window’s message values fixed while replacing α with either mean validation-window weights or weights from another test window. At fixed coordinates, weights varied across windows in every tested layer and seed. Both replacements raised MSE for every seed on all four datasets (Appendix F, Table 5). This directly tests the predictive value of matching weights to the current window.

DatasetInput-matched αValidation-mean αWindow-mismatched α
Weather0.1530.1790.185
Solar-Energy0.1710.2550.251
Electricity0.1380.2790.301
Traffic0.3890.7240.743
Reviewer_dyKCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FLAG, a unified framework that jointly models lagged relations and performs input-adaptive information exchange within a shared representation. Prior methods have limitations in exploring lagged relations and dynamically adjusting their utilization based on the current input, often leading to the mixing of relevant and irrelevant variable information. To address these challenges, this paper constructs a relational context by modeling lag-specific factor interactions without explicitly constructing dense variable-pair matrices, and further refines and updates this context using group communication. Experiments on eight public datasets demonstrate competitive performance and confirm the effectiveness of the Factor-Lag Relational Context.

Critical strengths
  1. The paper explores the importance of lagged relations and the fact that such cross-variable dependencies may vary across different observation windows; simply combining variable information may not fully utilize the dynamic changes in these relationships. This perspective is relatively novel and useful for modelling multivariate time series data.
  2. The paper employs two techniques to improve model efficiency and reduce computational costs: using shared low-rank read and write bases with lag-specific factor transitions to avoid explicitly constructing dense variable-pair matrices, and employing group communication to facilitate information exchange among variables.
  3. Through extensive experiments on eight public datasets, the paper demonstrates strong performance and validates the effectiveness of the Factor-Lag Relational Context module.
Critical weaknesses
  1. In Table 1, the paper achieved the best results in only six out of the 16 evaluation metrics across the eight datasets; in particular, on the ECL and Traffic datasets, which are relatively high-dimensional and involve complex cross-variable dependencies, there was a significant gap compared with other baselines. These results weaken the empirical evidence supporting the proposed method and indicate that there is considerable room for improvement in its design.
  2. Although the Factor-Lag Relational Context module aims to improve computational efficiency, it essentially relies on low-rank factorization; furthermore, the aggregation and redistribution mechanism employed by the Context-Adaptive Group Communication module has been explored in previous studies, such as Crossformer and SOFTS. These similarities raise concerns about the novelty of the proposed method.
  3. The paper does not provide specific hyperparameter configurations nor analyze the impact of these hyperparameters on the model. For example, it does not specify how to select the value of rr in the transition matrix: how should rr be chosen for the ETT datasets, which contain only seven variables, and how should it be chosen for the high-dimensional ECL and Traffic datasets? Furthermore, it does not specify how to select the value of KK in the Context-Adaptive Group Communication module.
  4. While the complexity is analyzed in Section 3.5, the experimental section fails to report efficiency metrics such as training time, memory cost and scalability with respect to the number of variables. The lack of empirical efficiency evaluations makes it difficult to assess the practical computational advantages of the proposed method.
Questions for authors
  1. The paper constructed the Factor-Lag Relational Context only once, , followed by multiple layers Context-Adaptive Group Communication to exchange variable information. Have the authors attempted to stack these two modules alternately across multiple layers? Such a design would allow the aggregated features across variables to participate once again in cross-lag interactions, thereby potentially enhancing the overall modelling capability.
  2. In Figure 5 of the Appendix, the transition matrices on Electricity dataset show very small differences. Does this suggest that the model learns similar interaction patterns across different lags? Could the authors further explain whether these lag-specific transition matrices capture meaningful cross-variable dependencies and how much they contribute to forecasting performance?
  3. As shown in Figure 6 in the Appendix, the forecasting performance on the Weather and Solar-Energy datasets does not consistently improve as the input length increases, and even decreases for some longer input windows. However, intuitively, as the input length increases, the model would have access to more historical information and thus achieve better performance. Could the author explain why the model does not consistently benefit from longer historical inputs on these datasets?
3 comments
AuthorsOpenReview ↗

Response to W1, W2, W3

We sincerely thank you for your thoughtful review. We respond to each point below.

Response to W1

In revised Table 1 (which adds FLAG results for the fixed seeds 2021, 2022, and 2023), FLAG ranks first in 9 of 16 comparisons, including ties, the most of any method. The absolute gaps to the best scores on the two high-dimensional datasets are small. Averaged over four horizons, FLAG’s MSE and MAE exceed the respective best scores by 0.008 and 0.006 on Electricity and by 0.016 and 0.009 on Traffic.

We observe that the relative performance of the models varies across datasets. TimeFilter has the lowest average MSE on Electricity and Traffic, while FLAG has the lowest average MSE and MAE on ETTh2, ETTm1, and Solar-Energy. These differences are consistent with the datasets’ distinct temporal patterns and cross-variable dependencies. FLAG’s strength lies in its performance across tasks. Appendix K, Table 6 gives the per-horizon results.

Response to W2

FLAG separates learning cross-variable relations from deciding how to use them. Crossformer routes information through a small set of vectors, and SOFTS aggregates series into a shared core. As described in Sections 3.3–3.4, FLAG first learns a transition matrix for each lag in shared read and write bases. The resulting relational context then determines both the messages and their pooling weights for the current window. The same learned relations can therefore support different exchanges across inputs.

Appendix E, Table 4 reports three-seed mean test MSE at T=H=96 as we vary transition structure and pooling independently. On Traffic, replacing full with diagonal transitions raises MSE from 0.389 to 0.395 with adaptive pooling fixed. Replacing adaptive with static weights raises it to 0.443 with full transitions fixed. The first three configurations tie to three decimals on ETTm1. The contribution of both choices is clearest on Traffic.

DatasetFull + adaptive (FLAG)Diagonal + adaptiveFull + staticDiagonal + static
ETTm10.3080.3080.3080.309
Weather0.1530.1560.1560.155
Electricity0.1380.1390.1390.139
Traffic0.3890.3950.4430.436

Response to W3

Appendix A, Table 3 lists the configuration shared across all four horizons for each dataset. The factor count r sets the size of each lag transition matrix to r × r. K is the number of soft groups, and L is the number of communication layers.

DatasetVariablesrKL
ETTh1 / ETTh2 / ETTm1 / ETTm27422
Weather211624
Solar-Energy1371622
Electricity3211621
Traffic8621612

In Appendix B, Figure 7, we vary r, K, and L one at a time at H=96, holding the other parameters at the settings above. The table gives the main configuration’s three-seed mean test MSE alongside the best settings in each sweep and their MSE. Ties at three decimals are included.

DatasetMain config. MSEr / min MSEK / min MSEL / min MSE
ETTm10.3086 / 0.3061, 2, 3, 4, 6 / 0.3084 / 0.306
Weather0.15312, 16 / 0.1531 / 0.1511, 2, 6 / 0.152
Solar-Energy0.1712 / 0.1682 / 0.1712, 6 / 0.171
Electricity0.1388 / 0.1364, 32 / 0.1372 / 0.136
Traffic0.38912 / 0.3881, 2 / 0.3893 / 0.387

The useful capacity depends on the dataset. On Traffic, increasing r from 4 to 12 lowers MSE from 0.408 to 0.388, but increasing it to 32 raises MSE to 0.427. On ETTm1, all five tested group counts from 1 to 6 give 0.308, indicating little gain from additional groups. On Traffic, MSE is 0.387 at L=3 and 0.402 at L=5. Excess rank may admit redundant or noisy interactions, and extra groups or layers offer limited benefit once performance levels off. The main configurations span four horizons, so they can differ from the minimum at H=96.

AuthorsOpenReview ↗

Response to W4

Response to W4

We measured six models on ETTm1 and Traffic, the datasets with the fewest and most variables. Section 4.5, Figure 6 reports training time, peak training memory, and reference MSE. Appendix J, Figure 11 reports inference latency. SOFTS uses less time and memory than FLAG on both datasets. On Traffic, FLAG has lower reference MSE, training time, inference latency, and peak training memory than DUET, TimeMixer, and PatchTST.

All measurements use the same H100, FP32, batch size 8, and T=H=96. Training time is averaged over one complete epoch, inference is timed after warm-up, and memory is peak allocated training memory. From ETTm1 (7 variables) to Traffic (862), FLAG’s time per iteration grows 2.75-fold and peak memory 20.31-fold, the second-smallest increases among the six models on both measures. Against DUET on Traffic, FLAG yields reference MSE of 0.389 versus 0.407 with 69.7% less training time per iteration and 58.2% less peak memory.

ModelETTm1 train (ms/iter)ETTm1 infer (ms/batch)ETTm1 peak (MiB)Traffic train (ms/iter)Traffic infer (ms/batch)Traffic peak (MiB)
FLAG8.471.32124.0823.258.082,519.81
SOFTS6.320.8868.6912.744.12654.60
TimeFilter7.261.05109.9424.999.212,578.71
DUET17.063.49126.1576.6623.636,025.89
TimeMixer18.473.5882.8653.9615.305,744.55
PatchTST9.271.16179.8381.2837.626,452.53
AuthorsOpenReview ↗

Response to Q1, Q2, Q3

Response to Q1

Thank you for this constructive suggestion. We tested alternating stacks in which each subsequent factor–lag module receives the representations updated by the preceding group-communication layer. This lets the model revisit cross-lag interactions after variables have exchanged information. Appendix C, Figure 8 compares 2, 4, 6, 8, 16, and 32 independently parameterized stages with Pre-RMS residual updates and a final RMSNorm. We fixed T=H=96, r=8, and K=2 and report three-seed mean test MSE. The seven variables in ETTm1 cap its effective rank at 7.

The non-alternating control uses one factor–lag module followed by two group-communication layers, with the original normalization. The control and alternating models use the same rank, group count, and training settings.

DatasetMatched non-alternating control2 stages4 stages6 stages8 stages16 stages32 stages
ETTm10.3080.3070.3080.3080.3080.3070.305
Weather0.1530.1520.1500.1520.1520.1490.151
Solar-Energy0.1750.1800.1770.1730.1810.1840.179
Electricity0.1370.1410.1400.1400.1370.1390.143

On ETTm1, both architectures have effective rank 7. The 32-stage alternating model achieves a three-seed mean test MSE of 0.305, compared with 0.308 for the matched non-alternating control.

Weather improves from 0.153 to 0.149 at 16 stages and Solar-Energy from 0.175 to 0.173 at 6 stages. On Electricity, 8 stages match the control at three decimals. Alternation can therefore improve forecasting, with different depths preferred on different tasks. We thank the reviewer for pointing us toward this direction for further study.

Response to Q2

On Electricity, the lag matrices have similar magnitudes, yet the coefficient for the same factor pair can change sign across lags (Figure 5). To measure the forecasting benefit of delayed interactions, we varied the lag-bank size on four datasets at H=96 and H=336. The table reports mean test MSE over three seeds. Size 1 permits only synchronous interactions (Section 4.3, Figure 4).

DatasetH1 (sync)36912
Weather960.1540.1530.1530.1540.153
Weather3360.2740.2670.2650.2670.263
Solar-Energy960.1860.1760.1770.1700.171
Solar-Energy3360.2040.1980.1930.1940.196
Electricity960.1380.1370.1370.1380.138
Electricity3360.1680.1680.1690.1720.170
Traffic960.4030.3920.3910.3880.389
Traffic3360.4430.4340.4300.4300.428

Solar-Energy and Traffic benefit most consistently. Every bank that includes nonzero lags improves on the synchronous control at both horizons. Weather benefits mainly at H=336, where MSE falls from 0.274 to 0.263. On Electricity, the largest gain at H=96 is 0.001, and the synchronous control ties for best at H=336. In a separate matched run on Electricity at H=96 with seed 2021, sharing one matrix across all positive lags gave 0.136 versus 0.138 for independent matrices. Different lags can learn different factor combinations, but the evidence for an additional forecasting benefit from independent matrices on Electricity is limited.

Response to Q3

A longer input window improves forecasting accuracy by providing more history, but the gains level off. Once older observations mainly repeat information already in the window or add noise unrelated to the forecast, extending the window offers little further benefit. In Appendix I, Figure 10, both Weather and Solar-Energy show their largest improvement when the input grows from 48 to 96 steps. Weather MSE then fluctuates between 0.150 and 0.153. Solar-Energy MSE reaches its minimum of 0.165 at 192 steps before edging up to 0.167. SOFTS, TQNet, and FACT also show intervals where longer windows increase error. Overall, FLAG achieves lower test MSE than SOFTS, TQNet, and FACT on both Weather and Solar-Energy at each of the five input lengths tested from 96 to 336 steps.

Reviewer_KKgXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FLAG for multivariate time series forecasting. It combines shared low-rank factors with lag-specific transitions to construct cross-variable relational context, which then guides input-dependent aggregation through latent groups. The architecture models synchronous and delayed interactions without constructing dense variable-pair matrices, with linear complexity in the number of variables when other dimensions are fixed. Experiments demonstrate competitive forecasting performance, alongside analyses of component contributions and lagged interactions.

Critical strengths
  1. The distinction between recurring cross-variable relations and their usefulness for a particular input provides a clear motivation for combining structured lag modeling with adaptive communication.

  2. The factor-lag formulation offers a compact way to represent delayed interactions. Sharing read and write bases across lags retains interactions between latent factors while avoiding explicit pairwise relation matrices.

  3. The evaluation covers several application domains and recent baselines. The component and lag ablations provide useful insights, and the discussion acknowledges cases where combining the modules does not improve performance.

Critical weaknesses
  1. The novelty over structured low-rank temporal models and the aggregation mechanism in SOFTS [1] remains insufficiently established. The component ablations do not isolate the distinctive choices, such as full versus diagonal lag transitions or adaptive versus static pooling, leaving the contribution of the proposed coupling unclear.

  2. The evidence for explicit lag modeling is limited. On Weather, the lag settings perform similarly under the standard prediction head, while the larger separation appears with a restricted head. Broader tests across datasets and forecasting horizons would help establish when delayed interactions improve the main model.

  3. The baseline suite omits TimeFilter [2] and DUET [3], although both are cited when motivating the need to filter unhelpful cross-variable information. Direct comparisons would better establish FLAG’s contribution relative to existing approaches addressing this concern.

  4. The scalability argument is supported only by asymptotic analysis. Training time, inference latency, peak memory, and scaling with the number of variables are not measured, so the practical efficiency advantage over compact models such as SOFTS remains unclear.

[1] SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion

[2] TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting

[3] DUET: Dual Clustering Enhanced Multivariate Time Series Forecasting

Questions for authors

How does FLAG compare with SOFTS in runtime and peak memory as the number of variables grows, and how does its forecasting accuracy compare with TimeFilter and DUET under comparable tuning budgets?

3 comments
AuthorsOpenReview ↗

Response to W1, W2

We sincerely appreciate your thoughtful and insightful review. We respond to each point below.

Response to W1

FLAG separates how cross-variable relations are learned from how they are used. It learns a full transition matrix for each lag in a shared low-rank factor space to capture relations across windows. SOFTS aggregates and redistributes information through a single shared core. FLAG uses the learned lag relations together with the current input to determine what each variable contributes to its group and how much weight its message receives (Sections 3.3–3.4).

We evaluated all four combinations of full or diagonal lag transitions and adaptive or static pooling on four datasets (Appendix E, Table 4). The input and forecast lengths were both 96. We report mean test MSE over three seeds.

DatasetFull + adaptive (FLAG)Diagonal + adaptiveFull + staticDiagonal + static
ETTm10.3080.3080.3080.309
Weather0.1530.1560.1560.155
Electricity0.1380.1390.1390.139
Traffic0.3890.3950.4430.436

Full transitions with adaptive pooling yield the lowest MSE on Weather, Electricity, and Traffic and tie for best on ETTm1 at three decimals. This combination delivers its largest gain on Traffic, where replacing static with adaptive pooling while keeping full transitions reduces MSE from 0.443 to 0.389.

Response to W2

We compared five lag-bank sizes using the main model’s standard forecasting head (Section 4.3, Figure 4). The table reports mean test MSE over three seeds. A bank of size 1 permits only synchronous interactions.

DatasetH1 (sync)36912 (FLAG)
Weather960.1540.1530.1530.1540.153
Weather3360.2740.2670.2650.2670.263
Solar-Energy960.1860.1760.1770.1700.171
Solar-Energy3360.2040.1980.1930.1940.196
Electricity960.1380.1370.1370.1380.138
Electricity3360.1680.1680.1690.1720.170
Traffic960.4030.3920.3910.3880.389
Traffic3360.4430.4340.4300.4300.428

Explicit lag interactions help the main model most consistently on Solar-Energy and Traffic. At both horizons, every bank that includes nonzero lags improves on synchronous interactions alone. On Weather, the gain is clearest at H=336, where the fixed 12-position bank reduces MSE from 0.274 to 0.263.

AuthorsOpenReview ↗

Response to W3, W4, Q1

Response to W3

We added comparisons with TimeFilter and DUET (Table 1 and per-horizon results in Appendix K, Table 6). Entries are MSE / MAE averaged over horizons 96, 192, 336, and 720 at T=96.

DatasetFLAGTimeFilterDUET
ETTh10.433 / 0.4250.420 / 0.4280.443 / 0.436
ETTh20.360 / 0.3850.364 / 0.3970.372 / 0.397
ETTm10.374 / 0.3800.377 / 0.3930.390 / 0.393
ETTm20.275 / 0.3170.272 / 0.3210.280 / 0.324
Weather0.241 / 0.2620.239 / 0.2690.251 / 0.273
Solar-Energy0.188 / 0.2080.223 / 0.2500.237 / 0.233
Electricity0.166 / 0.2620.158 / 0.2560.172 / 0.258
Traffic0.423 / 0.2770.407 / 0.2680.451 / 0.269

FLAG ranks first most often in Table 1, with 9 wins across 16 dataset–metric comparisons, including ties. On Electricity and Traffic, the two datasets with the most variables, FLAG has lower horizon-averaged MSE than DUET and higher MSE than TimeFilter.

Response to W4 / Q1

We profiled six models on ETTm1 (7 variables) and Traffic (862) using the same H100, batch size 8, FP32, and T=H=96. Section 4.5, Figure 6 reports training time and peak training memory. Appendix J, Figure 11 reports inference latency.

ModelETTm1 train (ms/iter)ETTm1 infer (ms/batch)ETTm1 peak (MiB)Traffic train (ms/iter)Traffic infer (ms/batch)Traffic peak (MiB)
FLAG8.471.32124.0823.258.082,519.81
SOFTS6.320.8868.6912.744.12654.60
TimeFilter7.261.05109.9424.999.212,578.71
DUET17.063.49126.1576.6623.636,025.89
TimeMixer18.473.5882.8653.9615.305,744.55
PatchTST9.271.16179.8381.2837.626,452.53

From ETTm1 to Traffic, training time per iteration, inference latency, and peak training memory increase by factors of 2.75, 6.14, and 20.31 for FLAG and 2.02, 4.67, and 9.53 for SOFTS. SOFTS has lower costs on all three measures on both datasets. Among the six models, FLAG has the second-smallest increases in training time and peak memory. On Traffic, FLAG has lower costs on all three measures and lower reference MSE at H=96 than DUET, TimeMixer, and PatchTST. Compared with DUET, FLAG reduces reference MSE from 0.407 to 0.389 and the three costs by 69.7%, 65.8%, and 58.2%, respectively.

For forecasting accuracy, we use the results reported by the TimeFilter and DUET authors under their respective configurations with a fixed input length of 96 (table in W3). Averaged over four forecast horizons, FLAG has lower MSE than DUET on all 8 datasets. Compared with TimeFilter, FLAG has lower MSE on 3 datasets and lower MAE on 6. Across the eight datasets, FLAG ranks first most often on forecasting accuracy (9 of 16 comparisons, including ties) and, on high-dimensional Traffic, trains faster per iteration, has lower inference latency, and uses less peak training memory than DUET, TimeMixer, and PatchTST.

Thank you for the detailed response and additional experiments. While I appreciate the authors’ efforts, my concerns regarding novelty, predictive performance, and computational efficiency have not been sufficiently addressed. I therefore maintain my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Area Chair and Reviewers

Dear Area Chair and Reviewers,

We sincerely appreciate the time and effort you have devoted to evaluating our manuscript, as well as your thoughtful and constructive feedback. Your comments have helped us strengthen the experimental evaluation and clarify the design and contributions of FLAG.

We have provided point-by-point responses and revised the manuscript accordingly. The main additions include:

  1. Richer experimental evaluations: We added three-seed results with standard deviations, detailed statistical analyses, and direct comparisons against TimeFilter and DUET.
  2. Deeper ablations: We expanded our ablation studies to cover full versus diagonal lag transitions, adaptive versus static pooling, lag-bank size, hyperparameter sensitivity, alternating stacking, and input-adaptive aggregation weight diagnostics.
  3. Real-world efficiency profiling: We measured training time, inference latency, and peak GPU memory across datasets with substantially different numbers of variables.

We have also clarified FLAG’s core contribution: cleanly separating the modeling of cross-variable relations from their adaptive use in forecasting. We hope this makes our two key components—Factor-Lag Relational Context and Context-Adaptive Group Communication—much clearer.

We truly appreciate your continued guidance. Please let us know if you have any remaining questions or if there is anything else we can clarify in the revised draft.

Thank you again for your time and thoughtful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.