acceptodds
Under review as a conference paper at ICLR 2027

MeraTSF: Event-Aware Residual Modeling for Sparse Exogenous Time Series Forecasting

Abstract

Incorporating exogenous variables into time series forecasting remains challenging, as real-world exogenous information is often sparse and temporally heterogeneous in its effects. Unlike conventional covariates, sparse exogenous events occur infrequently and abruptly, inducing transient disturbances, horizon-dependent responses, and long-term distribution shifts. These effects can be viewed as scale-dependent deviations from the underlying normal temporal dynamics, which are difficult to capture through direct correlation-based modeling. To address this challenge, we propose MeraTSF, a multi-scale residual refinement framework that models exogenous effects as structured residual deviations from normal temporal patterns. MeraTSF decomposes temporal dynamics across multiple scales and uses exogenous information to adaptively refine the corresponding residual representations, enabling the model to capture heterogeneous temporal effects induced by sparse exogenous events. Extensive experiments on multiple real-world datasets show that MeraTSF consistently outperforms state-of-the-art methods across diverse datasets and forecasting settings. Moreover, MeraTSF can be seamlessly integrated into existing forecasting architectures, highlighting its strong plug-and-play capability and broad applicability.

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

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_YZ15OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses time series forecasting with exogenous variables (TSF-X) when the exogenous information consists of sparse, abrupt events (weather, holidays, large-scale events) whose effects differ across temporal scales. MeraTSF decomposes the endogenous input with three moving-average kernels into trend and residual components, each patch-encoded separately. Historical and known-future exogenous variables are injected only into the residual tokens through a gated MLP correction, and scale-specific forecasts are fused with instance-dependent softmax weights. Experiments cover a large ride-hailing dataset with three event scenarios and 12 smaller TSF-X benchmarks, plus ablations, plug-in experiments with four backbones , zero-shot transfer, robustness studies and a setting without future exogenous variables.

Critical strengths
  1. The problem is practically important and the empirical study is broad. The paper uses a large real-world ride-hailing dataset (14M observations, 200 areas) with scenario-level breakdowns (Table 1) and covers deployment-relevant settings: zero-shot areas (Table 8), sparsified events (Table 9), missing endogenous values (Table 10) and inference cost (Fig. 5). It also honestly reports a negative result when future exogenous variables are unavailable (Table 11).
  2. The observation that exogenous conditioning helps more on the decomposition residual than on the trend is useful. In Table 4, the full model (WMAPE 0.1027) improves over both "w/o exo" (0.1102) and "on Trend" (0.1051).
Critical weaknesses

The mechanism is not specific to the stated problem, and the evidence that the gain comes from the key idea is limited. The method contains no sparsity-specific component, and the gated MLP (Eq. 7-8) is itself a learned endogenous-exogenous interaction; near-zero Pearson correlations for binary sparse indicators (Fig. 7-8) do not show that attention- or graph-based models fail. In Table 4, "on Trend" (0.1051) already recovers most of the gain, while "w/o future exo" (0.1122) is worse than no exogenous input at all, and MeraTSF loses to DAG on most datasets without future exogenous variables (Table 11).

Questions for authors
  1. Why does a method without any sparsity-specific component help specifically with sparse exogenous events? Please add a control that injects exogenous variables by simple concatenation into the undecomposed representation. 2.Report seed means ± std for results.
13 comments
AuthorsOpenReview ↗

Rebuttal Part 1

W1&Q1: The method contains no sparsity-specific component, and the gated MLP (Eq. 7-8) is itself a learned endogenous-exogenous interaction; Why does a method without any sparsity-specific component help specifically with sparse exogenous events?

We thank the reviewer for the valuable comments on the modeling of sparse exogenous events and the interaction between endogenous and exogenous variables. We apologize for not sufficiently clarifying these design choices in the original manuscript. We would like to clarify that MeraTSF employs learnable categorical embeddings to represent sparse discrete events, while its gated MLP performs exogenous-conditioned feature interactions without explicitly constructing correlation structures between endogenous and exogenous variables. Our detailed response is as follows.

Modeling of Sparse Discrete Exogenous Events. MeraTSF employs learnable embedding matrices based on discrete variable values, allowing the same event category to share a learnable feature representation across different temporal positions. This avoids treating discrete event categories as continuous numerical values and enables the model to learn reusable category representations from limited event observations.

Importantly, learnable embeddings are not the only way to encode sparse discrete events. We also consider One-hot encoding, which represents each event category using a distinct one-hot vector without introducing additional learnable encoding parameters.

To ensure fair comparisons, we evaluated exogenous-aware baselines, including DAG and CrossLinear, using both embedding and one-hot encoding strategies. For each encoding strategy, we further evaluated five different learning rates and selected the configuration with the best validation performance for reporting test results.

Endogenous–Exogenous Feature Interaction. We agree that the gated MLP in Eqs. (7)–(8) is inherently a learnable interaction mechanism between endogenous and exogenous features. However, we would like to clarify that MeraTSF does not aim to eliminate interactions between endogenous and exogenous variables, but rather to avoid explicitly constructing correlation structures between them.

Unlike methods that characterize endogenous–exogenous relationships through correlation matrices or dedicated dependency modeling modules, MeraTSF incorporates exogenous information as conditional signals to adaptively interact with multi-scale residual representations. Therefore, the key distinction is not whether feature interactions exist, but the representation space in which they occur and how they are modeled.

In summary, MeraTSF handles sparse discrete events through shared categorical representations and incorporates exogenous information through conditional interactions with multi-scale residual features. These design choices distinguish its modeling approach from methods that rely on explicitly constructed dependency structures, while maintaining fair comparisons through alternative encoding strategies and systematic hyperparameter selection.

AuthorsOpenReview ↗

Rebuttal Part 2

W2: near-zero Pearson correlations for binary sparse indicators (Fig. 7-8) do not show that attention- or graph-based models fail.

We thank the reviewer for this valuable observation and fully agree with the concern. We apologize for not sufficiently clarifying the relationship between sparse exogenous events and dependency modeling in the original manuscript. We would like to emphasize that weak Pearson correlations do not imply that dependency-based methods are incapable of learning linear or nonlinear relationships. Rather, sparse events may make it more challenging to learn stable dependency structures from the overall data distribution. We clarify this distinction using two representative methods, DAG and GCGNet.

Dependency Modeling under Sparse Events. DAG focuses on discovering and incorporating temporal and inter-channel dependencies, whereas GCGNet emphasizes maintaining graph structural consistency between predictions and ground-truth observations.

The relatively weaker performance of DAG and GCGNet under sparse-event scenarios does not necessarily indicate an inability to learn correlations. Instead, their dependency modeling objectives may be affected by the highly imbalanced distribution of sparse events, making it more challenging to capture stable dependency patterns or maintain graph structural consistency for rare event conditions.

Distinction from MeraTSF. In contrast, MeraTSF does not require explicitly establishing stable cross-variable dependency structures. Instead, it directly incorporates exogenous variables as conditioning signals to adaptively refine residual representations at different temporal scales.

Consequently, even when exogenous events occur infrequently, their information can still influence predictions through conditional feature interactions, without requiring strong global correlations between endogenous and exogenous variables.

In summary, our argument is not that dependency-based methods cannot model sparse events or nonlinear relationships, but that sparse event distributions may pose additional challenges for learning stable dependency structures. MeraTSF provides an alternative approach by directly conditioning multi-scale residual representations on exogenous information, rather than relying on explicitly constructed cross-variable dependencies.

AuthorsOpenReview ↗

Rebuttal Part 3

W3: In Table 4, "on Trend" (0.1051) already recovers most of the gain.

We thank the reviewer for the valuable comments on the choice of residual refinement over trend refinement. We would like to clarify that exogenous events may affect not only the overall level of a time series but also its fine-grained temporal dynamics, which motivates our choice of exogenous-conditioned residual refinement. Our detailed explanation is as follows.

Motivation for Residual Refinement. Incorporating exogenous information only into coarse-grained temporal patterns can capture global changes induced by external events, such as shifts in mean demand or overall demand levels. However, exogenous events may also alter fine-grained intraday patterns.

For example, a large-scale concert may cause a concentrated surge in ride-hailing demand around the venue after the event ends, changing the magnitude and distribution of intraday demand peaks rather than simply increasing the average demand throughout the day.

Such localized variations may be smoothed out in coarse-grained temporal components but can be better preserved in residual representations. Therefore, MeraTSF incorporates exogenous information into the residual branch to conditionally refine fine-grained temporal variations across multiple scales, enabling the model to better capture complex temporal dynamics associated with exogenous events.

In summary, our choice of residual refinement is motivated by the need to capture localized, event-related temporal variations that may not be adequately represented by coarse-grained trend modeling alone.

W4: while "w/o future exo" (0.1122) is worse than no exogenous input at all, and MeraTSF loses to DAG on most datasets without future exogenous variables (Table 11).

We thank the reviewer for the valuable comments on the performance differences between MeraTSF and DAG under different exogenous information settings. We would like to clarify that MeraTSF is designed to leverage future exogenous information for conditional forecasting, whereas DAG focuses more on modeling dependencies among variables. This distinction explains both the advantages and limitations of MeraTSF. Our detailed response is as follows.

Performance without Future Exogenous Variables. In non-sparse scenarios, MeraTSF does not explicitly establish stable dependency structures between endogenous and exogenous variables. Consequently, when future exogenous information is unavailable, its forecasting performance can be inferior to DAG, which emphasizes dependency modeling. We have already acknowledged this limitation in the revised manuscript.

Performance with Future Exogenous Variables. When future exogenous information is available, MeraTSF outperforms DAG by leveraging the additional contextual information to adaptively refine its predictions. This allows the model to better account for changes in external conditions that may not be fully captured by historical temporal dependencies alone.

The importance of future exogenous information is particularly evident in sparse-event scenarios. For example, if a large-scale event is scheduled to take place in a particular area, this information can help the model anticipate a potential surge in ride-hailing demand around the event. Without such information, the demand surge may be difficult to predict from historical observations alone.

In summary, MeraTSF is designed to exploit future exogenous information rather than relying primarily on explicitly learned dependency structures. While this may limit its performance when future exogenous variables are unavailable, it enables more accurate conditional forecasting when such information is provided, particularly under sparse exogenous events.

Reviewer_P9E5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MeraTSF addresses time-series forecasting with sparse and heterogeneous exogenous variables whose effects may be transient, horizon-dependent, or persistent. The method normalizes the endogenous history, decomposes it into trend and moving-average residual components at multiple temporal scales, encodes historical and known-future exogenous variables, applies a gated exogenous-conditioned correction to the residual representations, and combines scale-specific forecasts through instance-level weighting. It is evaluated on a four-year, 200-area Ride-Hailing dataset under weather, holiday, and large-scale event scenarios, together with 12 smaller datasets. The experiments include comparisons with exogenous-aware and generic forecasting baselines, kernel and component ablations, backbone integration, zero-shot area forecasting, exogenous sparsification, missing-value tests, seed sensitivity, and a setting without future exogenous variables. The reported results show modest but generally favorable forecasting improvements, especially on the Ride-Hailing benchmark.

Critical strengths
  1. A modular and operationally plausible architecture. The gated additive correction in Equations (7) and (8) provides a clear mechanism for injecting exogenous information into multi-scale residual tokens. Tables 3 to 5 show that removing kernels or major components generally reduces performance, and the module improves several backbones on the Ride-Hailing dataset.
  2. Broad empirical coverage with an explicit limitation analysis. The paper evaluates 13 datasets and includes scenario-specific forecasting, zero-shot area testing, sparsification, missing-value robustness, backbone integration, and random-seed analysis. The authors also openly report that MeraTSF does not retain its advantage when future exogenous variables are unavailable, which is a useful and credible limitation.
Critical weaknesses
  1. The central residual and event-separation interpretation is not identified by the method. Equation (2) defines R_s as the difference between the normalized input and a moving average. This component necessarily contains endogenous high-frequency dynamics, noise, and model mismatch; it is not a forecasting error and is not shown to isolate exogenous effects. Therefore, statements that the method "decouples" normal dynamics from event effects or that residual refinement explains exogenous deviations are stronger than the evidence supports. Tables 3 and 4 establish predictive ablation effects, but not the proposed mechanistic or causal interpretation. Figure 4(a) further uses an undefined "information gain" quantity and makes uncertainty-related claims without specifying an estimator or uncertainty calculation.
  2. The strongest empirical claims rely on insufficiently documented future-exogenous information and uncertainty. The main model receives exact future exogenous inputs, and removing future exogenous variables causes a substantial degradation. Yet the paper does not specify whether weather, event, and scenario variables are observed, scheduled, forecasted, or derived after the forecast origin. The sparse-event results are concentrated in one Ride-Hailing benchmark, where large-scale events occur at only 0.23-0.30% rates, but no event-window counts, overlap rules, repeated masks, confidence intervals, or paired significance tests are reported. The overall improvements in Table 1 are often only about 1-2%, so their reliability cannot be assessed from the current presentation.
Questions for authors
  1. Are the future exogenous variables truly available at prediction time, and are the scenario factors free from target leakage? Please report results with realistic forecasted or noisy exogenous inputs.
  2. What evidence shows that R_s represents exogenous effects rather than generic high-frequency variation? Please also clarify how the horizon-dependent weights in Figure 3 and the information gain in Figure 4 are computed.
  3. How many test windows belong to each event scenario? Please report repeated-seed results and confidence intervals, especially for the rare large-scale event cases.
  4. How are unseen Area IDs encoded in the zero-shot experiment, and what exact procedures and parameter budgets are used to adapt the baselines?
5 comments
AuthorsOpenReview ↗

Rebuttal Part 1

W1&Q2: Equation (2) defines R_s as the difference between the normalized input and a moving average. This component necessarily contains endogenous high-frequency dynamics, noise, and model mismatch; it is not a forecasting error and is not shown to isolate exogenous effects. Therefore, statements that the method "decouples" normal dynamics from event effects or that residual refinement explains exogenous deviations are stronger than the evidence supports. What evidence shows that R_s represents exogenous effects rather than generic high-frequency variation?

We thank the reviewer for the valuable comments on the consistency between the mechanistic interpretation and mathematical formulation of MeraTSF. We agree that our original claim of "decoupling normal temporal dynamics from exogenous event effects" was not sufficiently precise. We would like to clarify that MeraTSF does not directly separate normal temporal dynamics from exogenous event effects, but instead structurally decouples coarse-grained temporal pattern modeling from exogenous-conditioned residual refinement. Accordingly, we have revisited and refined the methodological motivation and mechanistic description in the revised manuscript. Our detailed response is as follows.

Mathematical Definition and Interpretation of Residual Components. The moving average operation in Eq. (2) aims to extract coarse-grained temporal structures at different scales, while the residual components represent fine-grained variations relative to these structures. Importantly, the residual components do not exclusively represent exogenous event effects. Instead, they contain endogenous high-frequency fluctuations, exogenous event-related variations, and other information not captured by the trend components. Therefore, we do not interpret residuals as isolated exogenous event effects, but rather as fine-grained representations encompassing multiple sources of temporal variation.

Actual Modeling Mechanism of MeraTSF. The core idea of MeraTSF is to structurally decouple coarse-grained temporal pattern modeling from exogenous-conditioned residual refinement.

Specifically, the trend branch primarily captures relatively smooth temporal structures, while the residual branch preserves fine-grained variations relative to these structures. The model then incorporates exogenous variables as contextual conditions to adaptively refine residual representations at different temporal scales, thereby leveraging exogenous information more effectively to improve forecasting performance.

Thus, MeraTSF does not aim to directly isolate exogenous event effects from observed time series. Instead, it focuses on leveraging exogenous information to selectively refine residual representations that encompass multiple sources of temporal variation.

To provide a more intuitive illustration of this modeling mechanism, we further visualize the forecasting results and different prediction components, as shown in the following figure: https://anonymous.4open.science/r/MeraTSF-main/figure/r1.png

As illustrated, without weather factors, the model struggles to respond effectively to weather disturbances, whereas incorporating weather information enables it to better adjust its predictions. Meanwhile, the trend branch primarily captures relatively smooth, coarse-grained temporal patterns, while the residual branch adaptively refines fine-grained temporal variations through exogenous conditioning. These observations are consistent with the design principle of MeraTSF.

We have accordingly revised the methodological motivation and mechanistic description in the manuscript, explicitly distinguishing residual components from exogenous event effects to ensure consistency between the mechanistic interpretation and mathematical formulation.

In summary, MeraTSF does not assume that residual components independently isolate exogenous event effects. Instead, through multi-scale decomposition and exogenous-conditioned residual refinement, it adaptively refines fine-grained temporal dynamics using exogenous information while modeling coarse-grained temporal structures. This structural decoupling mechanism constitutes the core design principle of MeraTSF.

AuthorsOpenReview ↗

Rebuttal Part 2

W1&Q2: Figure 4(a) further uses an undefined "information gain" quantity and makes uncertainty-related claims without specifying an estimator or uncertainty calculation. Please also clarify how the horizon-dependent weights in Figure 3 and the information gain in Figure 4 are computed.

We thank the reviewer for the valuable comments on the scale weights and information gain analysis of MeraTSF. We apologize for not sufficiently clarifying the definitions and calculations of these quantities in the original manuscript. We would like to clarify that Figure 7 illustrates the relative prediction contributions of different scales across forecasting horizons, rather than directly learned horizon-dependent weights, while Figure 8(a) uses a Gaussian information gain proxy to quantify the additional predictive information provided by exogenous variables for trend and residual components. We have clarified these definitions and estimation procedures in the revised manuscript. Our detailed response is as follows.

Clarification of Scale Weights and Prediction Contributions. Although the instance-level scale fusion weights defined in Eq. (11) are shared across forecasting horizons, the scale-specific predictions vary across forecasting steps. To quantify the relative contribution of each scale at different forecasting horizons, we compute the relative absolute prediction contribution from the weighted predictions:

\frac{\sum _{i=1}^{N} |\omega _{i,s}\bar{Y} _{i,s,h}|} {\sum _{j=1}^{S}\sum _{i=1}^{N} |\omega _{i,j}\bar{Y} _{i,j,h}|+\epsilon}.$$ Here, $\omega _{i,s}$ denotes the instance-level fusion weight at scale $s$, $\bar{Y} _{i,s,h}$ denotes the corresponding prediction at forecasting horizon $h$, and $N$ denotes the number of endogenous variables included in the analysis. Importantly, **$w _{s,h}$ is not a directly learned horizon-dependent weight, but a relative absolute contribution computed from the weighted predictions.** Its variation across forecasting horizons arises from the scale-specific predictions rather than the scale fusion weights themselves. **Definition and Estimation of Information Gain.** To quantify the additional predictive information provided by exogenous variables for different temporal components, we decompose the future ground-truth target series into trend and residual components using the same decomposition kernel at each scale $s$. For each component $U \in \{T,R\}$, we train two Ridge regression probes: - The first probe uses historical observation statistics and periodic features $C$. - The second probe additionally incorporates exogenous variables $Z$. Both probes use identical target representations and region-disjoint cross-validation splits to obtain out-of-fold prediction errors. We then project the future target components into a $d=12$ dimensional PCA space, estimate the prediction error covariance matrices, and define a regularized Gaussian information gain proxy as follows: $$\widehat{IG} _{s}^{(U)}=\frac{1}{2d} \log \frac{\det(\widehat{\Sigma} _{C,s}^{(U)}+\lambda I _d)} {\det(\widehat{\Sigma} _{C,Z,s}^{(U)}+\lambda I _d)}, \qquad \lambda=10^{-4}.$$ Both covariance matrices are estimated from out-of-fold prediction errors. We further perform 1,000 area-level bootstrap resamples to assess the sampling variability of this quantity. Importantly, **this metric measures the change in the prediction error covariance volume after incorporating exogenous variables. It serves as a proxy for predictive information gain under a Gaussian assumption, rather than a measure of causal information gain.** The purpose of Figure 8(a) is to compare the incremental predictive information provided by $Z$ for trend and residual components under the same probing protocol. **In summary,** we have explicitly distinguished the directly learned scale fusion weights from the relative prediction contributions computed from model outputs. We have also provided the mathematical definition, estimation procedure, and uncertainty assessment of the information gain proxy. These revisions clarify the interpretations of Figure 7 and Figure 8(a) and provide a more rigorous empirical basis for analyzing the multi-scale modeling mechanism of MeraTSF and the role of exogenous information.
AuthorsOpenReview ↗

Rebuttal Part 3

W2&Q1: The main model receives exact future exogenous inputs, and removing future exogenous variables causes a substantial degradation. Yet the paper does not specify whether weather, event, and scenario variables are observed, scheduled, forecasted, or derived after the forecast origin. Are the future exogenous variables truly available at prediction time, and are the scenario factors free from target leakage? Please report results with realistic forecasted or noisy exogenous inputs.

We thank the reviewer for the valuable comments on the availability of future exogenous variables, potential target leakage, and the practical applicability of MeraTSF. We apologize for not sufficiently clarifying these aspects in the original manuscript. We would like to clarify that holiday and large-scale event information can generally be obtained in advance, whereas future weather information requires forecasts. Our main experiments use observed future weather factors under controlled conditions, and we further evaluate robustness to inaccurate weather inputs. Our detailed response is as follows.

Availability of Future Exogenous Variables. The exogenous variables in the Ride-Hailing dataset can be categorized into three types:

  • Holiday information: Holiday schedules can be obtained in advance from calendars.
  • Large-scale event information: Events such as concerts and marathons typically announce their schedules and locations in advance, making this information generally available at prediction time.
  • Weather information: Future weather conditions cannot be precisely known in advance and must be obtained through weather forecasts in practical applications.

Furthermore, the large-scale event information in Ride-Hailing was independently collected and organized rather than derived from the target time series. This ensures that the event annotations themselves are not constructed from future target values.

Use of Future Weather Information in the Main Experiments. In our main experiments on Ride-Hailing, we use future weather factors derived from actual weather observations rather than historical weather forecasts. This controlled setting aims to evaluate the potential predictive value of weather information without interference from weather forecast errors.

The experimental results demonstrate that MeraTSF effectively utilizes weather information to improve forecasting performance. In particular, removing weather factors increases WMAPE from 0.1217 to 0.1379 under the Weather Disturbance scenario, corresponding to a 13.31% relative increase.

Robustness to Errors in Future Weather Information. Since actual weather forecasts inevitably contain errors, we further conduct robustness experiments by injecting Gaussian noise at levels of 5%, 15%, and 25% into the weather factors to simulate uncertainty in future weather information. The results are presented below.

Table 5 (Excerpt): WMAPE under different levels of weather-factor perturbation (seed 2024).

Weather-factor SettingOverallWeather DisturbanceHoliday EffectLarge-scale Event Impact
w/o weather factor0.10670.13790.13470.1106
0% noise0.10270.12170.12610.1069
5% noise0.10290.12200.12630.1069
15% noise0.10320.12300.12700.1073
25% noise0.10410.12600.13000.1084

The results show that forecasting errors gradually increase with higher noise levels, while MeraTSF maintains relatively stable performance. Even under 25% Gaussian noise, WMAPE increases by only 1.36%–3.55% across different scenarios compared with the noise-free setting. Specifically, Overall WMAPE increases from 0.1027 to 0.1041, while WMAPE under Weather Disturbance increases from 0.1217 to 0.1260.

In summary, we have clarified the availability of different future exogenous variables and the use of observed weather information in our main experiments. The additional robustness experiments demonstrate that MeraTSF effectively leverages future weather information while maintaining relatively stable forecasting performance under moderate input errors, providing further empirical support for its potential applicability in real-world forecasting scenarios.

Reviewer_v6HCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MeraTSF, a multi-scale residual refinement framework for time series forecasting with sparse exogenous variables. The method decomposes endogenous time series into trend and residual representations at multiple temporal scales, uses historical and future exogenous variables to conditionally refine the residual representations, and combines scale-specific predictions through adaptive weighting.

Critical strengths

S1. This paper addresses the challenge of incorporating sparse and heterogeneous external events into forecasting models. S2. The proposed residual refinement approach provides a reasonable architectural strategy for handling exogenous influences. S3. This paper evaluates the proposed method against a substantial set of baselines and includes comparisons across multiple datasets.

Critical weaknesses

W1. The proposed framework combines several established techniques, including multi-scale trend-residual decomposition, gated feature interaction, residual connections, and adaptive scale mixing. the methodological distinction from existing decomposition-based and exogenous-aware forecasting methods is insufficiently established, and appear to be combination of existing techniques.

W2. A central claim is that MeraTSF separates normal temporal dynamics from exogenous-event-induced deviations. However, Eq. 2 defines the residual component as the difference between the observed series and its moving-average trend. Since both components are derived from observations potentially affected by external events, the current formulation and empirical analyses do not sufficiently substantiate the claimed disentanglement of normal dynamics and event-induced effects.

W3. Eq. 11 defines instance-level scale weights without an explicit forecasting-horizon dimension, suggesting that each scale receives a single weight across the prediction horizon. However, Figure 3 presents scale weights that vary substantially across individual forecasting horizons. It is unclear how these horizon-dependent weights are obtained from the proposed formulation.

W4. the experimental evidence directly addressing sparse exogenous events primarily comes from the Ride-Hailing dataset. the generalizability of the proposed framework to independently collected sparse-event datasets, different event distributions, and other application domains remains insufficiently established.

W5. MeraTSF incorporates future exogenous variables into the forecasting process, but the paper does not sufficiently clarify whether these variables are genuinely available at prediction time or correspond to subsequently observed values. These raise concerns about the practical applicability of the method and the extent to which its improvements depend on potentially unavailable future information.

Questions for authors

NA

6 comments
AuthorsOpenReview ↗

Rebuttal Part 1

W1: The proposed framework combines several established techniques, including multi-scale trend-residual decomposition, gated feature interaction, residual connections, and adaptive scale mixing. the methodological distinction from existing decomposition-based and exogenous-aware forecasting methods is insufficiently established, and appear to be combination of existing techniques.

We thank the reviewer for the careful assessment of MeraTSF's methodological novelty. We would like to clarify that the core contribution of MeraTSF does not lie in introducing entirely new individual modules, but in proposing a multi-scale residual refinement framework designed to address the temporal heterogeneity of sparse exogenous event effects. By structurally decoupling coarse-grained temporal pattern modeling from exogenous-conditioned residual refinement, MeraTSF enables adaptive utilization of exogenous information across different temporal scales. We provide further clarification below.

Core Design of MeraTSF. MeraTSF structurally decouples coarse-grained temporal pattern modeling from exogenous-conditioned residual refinement. Specifically, the trend branch primarily captures relatively smooth temporal structures, such as daily and weekly periodic patterns, while the residual branch represents fine-grained variations relative to these structures, including both endogenous temporal fluctuations and exogenous event-related changes.

By introducing exogenous-conditioned refinement into residual representations at different temporal scales, MeraTSF adaptively adjusts these representations according to temporal context and exogenous information, while maintaining an independent trend modeling pathway. This allows the model to capture dominant temporal patterns while selectively incorporating exogenous information to refine its forecasts.

Distinctions from Existing Methods. The main distinctions between MeraTSF and existing methods lie in their modeling objectives and how exogenous information is incorporated.

  • Compared with decomposition-based methods such as DLinear and TimeMixer: These methods primarily focus on modeling and integrating different temporal components. In contrast, MeraTSF introduces exogenous conditioning into multi-scale residual refinement, emphasizing the adaptive adjustment of fine-grained dynamics according to external conditions.

  • Compared with dependency modeling methods such as DAG and GCGNet: DAG focuses on temporal and channel dependencies, while GCGNet emphasizes graph structural consistency. In contrast, MeraTSF does not rely on explicitly constructing cross-variable dependency structures. Instead, it incorporates external information through interactions between multi-scale residual representations and exogenous conditions.

AuthorsOpenReview ↗

Rebuttal Part 2

W2: A central claim is that MeraTSF separates normal temporal dynamics from exogenous-event-induced deviations. However, Eq. 2 defines the residual component as the difference between the observed series and its moving-average trend. Since both components are derived from observations potentially affected by external events, the current formulation and empirical analyses do not sufficiently substantiate the claimed disentanglement of normal dynamics and event-induced effects.

We thank the reviewer for pointing out the inconsistency between the mechanistic interpretation and mathematical formulation of MeraTSF. We agree that our original claim of "decoupling normal temporal dynamics from exogenous event effects" was not sufficiently precise. The core idea of MeraTSF is not to directly separate normal temporal dynamics from exogenous event effects, but to structurally decouple coarse-grained temporal pattern modeling from exogenous-conditioned residual refinement. We have clarified these concepts and revised the motivation and mechanistic description accordingly. Our detailed response is as follows.

Definition and Interpretation of Residual Components. MeraTSF employs multi-scale decomposition to extract coarse-grained temporal structures at different scales, while the residual components represent fine-grained variations relative to these structures. Importantly, the residual components do not exclusively represent exogenous event effects. Instead, they contain endogenous high-frequency fluctuations, exogenous event-related variations, and other information not captured by the trend components. Therefore, we do not interpret the residuals as isolated exogenous event effects, but rather as fine-grained representations encompassing multiple sources of temporal variation.

Empirical Evidence. To provide a more intuitive illustration of the modeling mechanism of MeraTSF, we further visualize the different forecasting components, as shown in the following figure: https://anonymous.4open.science/r/MeraTSF-main/figure/r1.png

As illustrated, the trend branch primarily captures relatively stable coarse-grained temporal patterns, while the residual branch, conditioned on exogenous information, adaptively refines fine-grained temporal dynamics. These observations are consistent with the design principle of MeraTSF, which structurally decouples coarse-grained temporal pattern modeling from exogenous-conditioned residual refinement.

AuthorsOpenReview ↗

Rebuttal Part 3

W3: Eq. 11 defines instance-level scale weights without an explicit forecasting-horizon dimension, suggesting that each scale receives a single weight across the prediction horizon. However, Figure 3 presents scale weights that vary substantially across individual forecasting horizons. It is unclear how these horizon-dependent weights are obtained from the proposed formulation.

We thank the reviewer for pointing out the ambiguity in the equation and figure terminology. The scale fusion weights in Eq. (11) are shared across forecasting horizons within each prediction instance. The horizon-dependent quantities are the scale-specific predictions, rather than the scale weights themselves. Therefore, the revised Figure 7 (Figure 3 in the original discussion) does not visualize directly learned horizon-dependent scale weights. Instead, it presents the relative absolute prediction contribution of each scale, computed from the weighted predictions:

ws,h=∑i=1N∣ωi,sYˉi,s,h∣∑j=1S∑i=1N∣ωi,jYˉi,j,h∣+ϵ.w _{s,h}=\frac{\sum _{i=1}^{N}\left|\omega _{i,s}\bar{Y} _{i,s,h}\right|}{\sum _{j=1}^{S}\sum _{i=1}^{N}\left|\omega _{i,j}\bar{Y} _{i,j,h}\right|+\epsilon}.

Here, Yˉi,s,h\bar{Y} _{i,s,h} denotes the prediction of the ii-th evaluated instance at scale ss and forecasting horizon hh. The horizon dependence of ws,hw _{s,h} arises from the scale-specific predictions, rather than from the scale fusion weights defined in Eq. (11).

Comments on the paper

1 comment
Area_Chair_yLVeOpenReview ↗

Reminder: Reviewer–Author Discussion

Thank you to the reviewers who have already engaged in the discussion and provided valuable feedback. As the discussion deadline is approaching, I would kindly encourage the remaining reviewers to review the authors' responses and share their follow-up comments, particularly on whether their concerns have been adequately addressed.

Authors are also encouraged to respond promptly to any further questions. Thank you all for your time and contributions to a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.