acceptodds
Under review as a conference paper at ICLR 2027

When Exogenous Features Hurt: Rethinking Feature Selection in Time-Series Forecasting

Abstract

Feature selection plays a critical role in machine learning in improving generalization and interpretability. In time series forecasting, this problem is particularly challenging: predicting a target variable often relies on a diverse set of exogenous features, yet recent studies have shown that incorporating certain features can unexpectedly degrade performance. This counterintuitive phenomenon has led to the growing adoption of univariate or channel-independent approaches, raising a fundamental question: when and why are exogenous features indeed beneficial? In this work, we argue that effective feature selection should be guided not by static relevance (e.g., correlation), but by the temporal stability of the conditional relationship between features and the target. We introduce Conditional Distribution Difference (CDD), a tractable proxy that characterizes temporal instability in the conditional relationship through a bin-wise representation. Based on this representation, we derive CDD-std, the standard deviation of CDD across time, as a feature selection metric that quantifies conditional stability. Our key insight is that features with high CDD-std and unstable conditional relationships are difficult to generalize and tend to harm forecasting performance. We validate this insight on controlled synthetic experiments and real-world datasets, showing that CDD-std provides a more effective indicator of feature usefulness than correlation-based metrics. We develop simple yet effective feature selection strategies building on CDD-std, including hard filtering and TimeLess, achieving up to 18% performance improvement over state-of-the-art forecasting models across diverse settings.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Forecasting, Feature Selection

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_G6dsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies feature selection in time series forecasting. The authors first propose Conditional Distribution Difference (CDD) and examine its properties and implications using simulated data. CDD empirically shown to be a metric that reflects temporal instability, and Lower CDD-std refers to greater gains from exogenous features. Based on this finding, the authors proposed CDD-std hard filtering and Timeless, a CDD-std guided learning method. Various results were conduced and CDD-std based achieve overall the best performance.

Critical strengths

Pros:

  • This paper is well-motivation and easy to follow.
  • The proposed method is simple and effective. The hard filtering method can also easily apply to retrained foundation models.
Critical weaknesses

Cons:

  • In Figure 4b, the test error shows a decrease and then increase pattern with the increase of number of layers. It would be interesting to also explain why this pattern exists.
  • It would be helpful to analyse the computational complexity given Eqs 1 and 5 and how the number of bins affects runtime.
  • The title of sec. 3.2, it seems not appropriate calling the method as meta learning.
Questions for authors

For multivariate forecasting, can you explain how TIMELESS is applied to forecast each variate?

What is the (possible) relationship between TIMELESS performance and correlations with covariates?

2 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer G6ds

Thank you for the valuable assessment and constructive suggestions.

W1: Error trend with network depth

We believe the reviewer refers to Fig. 2(b). The initial decrease in test error reflects improved modeling capacity, while the subsequent increase is mainly due to overfitting. In Fig. 2(b), the varying-frequency trend and added noise make the underlying pattern harder to distinguish, causing deeper networks to overfit noisy patterns. In contrast, the fixed-frequency trend in Fig. 2(a) is easier to learn, making overfitting less pronounced.

W2: Complexity analysis

Let T be the training length, d the number of exogenous features, w and h the window and horizon lengths, s the sliding step, and K the number of temporal bins. With N=(T-w-h)/s windows, directly evaluating Eq. 1 requires comparing all pairs of windows, resulting in O(dN^2) Wasserstein distance computations. Our bin-wise construction compares each window only with the reference of its corresponding bin, reducing the number of comparisons to O(dN). The resulting computational complexity is O(dN(w+h+(m+T/K) log T)).

Increasing K reduces the reference size T/K and therefore does not increase the computational cost. CDD-std is computed only once on the training split, with marginal overhead. For example, on Traffic (862 features), computing dataset-level CDD-std takes less than 2 minutes.

For Eq. 5, the CDD-std vector is precomputed during dataset initialization, before model training. The MLP therefore only introduces a small additional computational overhead. Its architecture and computational cost are independent of the number of temporal bins.

W3: Terminology

We agree that "meta learning" could be inaccurate. We will use "Adaptive Feature Weighting" to better reflect how TimeLess uses CDD-std as feature-level metadata to guide feature weighting.

Q1: Multivariate forecasting

Timeless can be extended to multivariate forecasting following the multi-univariate forecasting paradigm, as also discussed in [1]. First, CDD-std is computed for each target-feature pair to obtain target-specific feature descriptors. Second, the descriptors are organized into a batched tensor with an additional variate dimension (e.g., [batch, n_variates, ...], instead of [batch, ...]) and jointly processed in a single MLP forward pass to generate target-specific feature weights and selection masks. Finally, the resulting weights and masks are applied to the corresponding target's forecasting process.

Q2: TimeLess performance and correlation with covariates

To investigate the relationship between feature-target correlation and TimeLess gains, we analyze the eight datasets in Fig. 1(b) for which TimeLess results are available with the four main models (Tables 11, 12, and 16). We compute the Spearman correlation between each dataset's overall feature-target correlation and the MSE reduction of TimeLess over Base, averaged across horizons and models. The correlation is -0.19 (p=0.66); restricting the analysis to the datasets used in Table 1 yields 0.49 (p=0.36). Neither result shows a clear relationship between feature-target correlation and TimeLess gains.

This is consistent with our intuition in Fig. 1(b): overall feature-target correlation is not a reliable indicator of exogenous feature usefulness, as it does not capture temporal stability. TimeLess instead uses CDD-std to identify and down-weight unstable features.

[1] Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning. ICML

Thank the authors for the explanation, which helped me understand more. I would keep my score.

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

This paper studies why exogenous variables can sometimes degrade rather than improve time-series forecasting. The authors argue that conventional feature-selection criteria based on static dependence, such as correlation, fail to account for temporal changes in feature–target relationships. To address this issue, they introduce Conditional Distribution Difference (CDD), which approximates temporal variation in the conditional relationship between historical exogenous inputs and future targets through normalized joint feature–target interactions and Wasserstein-distance comparisons within temporal bins. CDD-std, defined from the variability of these bin-level discrepancies, is then used as an indicator of temporal instability, with larger values interpreted as evidence that a feature may generalize poorly across time. Building on this metric, the paper proposes two feature-selection strategies. The first applies an empirically calibrated CDD-std threshold, while the second, TIMELESS, uses CDD-std as feature-level metadata for a lightweight learnable weighting module. The framework is evaluated across multiple supervised forecasting architectures and datasets, together with analyses on unseen datasets and models, stable settings, and a foundation-model setting. Overall, the paper aims to establish temporal conditional stability as a useful principle for deciding which exogenous variables should be retained in forecasting systems.

Critical strengths

S1. Important and well-motivated problem. The paper addresses a practically relevant issue that is often overlooked: additional exogenous information can harm forecasting when its relationship with the target is temporally unstable.

S2. Simple and model-agnostic perspective. CDD-std provides an external feature-selection signal that can be applied without redesigning the underlying forecasting architecture.

S3. Clear distinction between static relevance and temporal reliability. The paper provides a useful conceptual perspective by emphasizing that strong dependence alone is insufficient if the relationship does not remain stable over time.

S4. Relatively comprehensive evaluation protocol. The study considers multiple forecasting architectures, unseen datasets and models, stable regimes, ablation analyses, repeated runs, and paired statistical testing, which strengthens the empirical evaluation.

Critical weaknesses

W1. CDD does not directly measure incremental predictive utility. The formulation evaluates feature–target relationships largely feature by feature, whereas the practical usefulness of an exogenous variable should generally be assessed conditional on the target history and other covariates. Consequently, CDD may fail to capture redundancy, complementarity, or synergistic interactions among features.

W2. The bin-wise construction is only an indirect proxy for temporal conditional shift. Each local window is compared with a reference distribution from its own temporal bin, and CDD-std summarizes variability in these within-bin discrepancies. This does not necessarily detect a regime change in which each bin is internally stable but the conditional relationship changes substantially between bins. The connection between the implemented statistic and the intended cross-time conditional-distribution discrepancy therefore deserves stronger justification.

W3. The hard-filtering threshold is partly data-dependent in a way that complicates evaluation. The threshold is obtained by regressing forecasting gains on CDD-std using the development datasets and horizons, while the main evaluation subsequently reuses datasets from the same development analysis. Although the paper includes a separate unseen-dataset evaluation, the main threshold-based results on the development datasets are not fully independent of threshold construction.

W4. Hyperparameter robustness is only partially investigated. The sensitivity analysis focuses mainly on the number of temporal bins, although CDD also depends on choices such as the window, step size, normalization, and downsampling procedure. The paper itself notes that changing the number of bins may require recalibrating the filtering threshold, which raises questions about the portability of the absolute threshold across settings.

Questions for authors

1.Clarify the theoretical status of low CDD-std. The current one-sided guarantee supports large product discrepancy as evidence of instability, but not the converse. Please explain what theoretical or empirical argument justifies interpreting small CDD-std as evidence of feature reliability.

2.Clarify what notion of feature utility CDD is intended to measure. In practical forecasting, an exogenous variable may be useful only conditionally on target history or in combination with other covariates. Please discuss how redundancy and feature interactions affect the proposed metric, and whether a conditional or multivariate extension is possible.

3.Provide a stronger analysis of cross-bin regime shifts. A useful counterexample would be a setting where each temporal bin is internally stable but the feature–target relationship differs substantially between bins. Please clarify whether the current CDD-std construction can detect such shifts.

4.Clarify the separation between threshold development and evaluation. Since the absolute threshold is derived from forecasting gains on the development datasets, please distinguish clearly which results are genuinely out-of-sample with respect to threshold calibration. A leave-one-dataset-out or validation-only calibration would provide stronger evidence.

4 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer dEvd

Thank you for the detailed and constructive review. We clarify the intended role of CDD-std and address the concerns below.

Clarification: CDD-std measures temporal stability, not utility

Our goal is not to measure feature utilities or rank intrinsic predictive utility, but to identify features whose relationships with the target are temporally unstable and may therefore hurt out-of-sample forecasting. A feature may be informative under certain contextual conditions but fail to generalize when the necessary context is unavailable at inference time, as illustrated in Example 2.2 and the Exchange dataset (currency changes are naturally not isolated). CDD-std thus serves as a stability measure criterion, where its selected features are more likely to be effectively generalized by forecasting models.

W1 / Q2: Incremental utility, redundancy, and interactions

CDD-std captures the temporal stability of individual feature-target relationships, complementing conventional feature selection criteria based on relevance and redundancy. In particular, mRMR explicitly accounts for feature redundancy, yet CDD-std achieves consistent lower average error under the same feature budget (top 1/3; Table 2). This suggests that temporal stability provides additional information beyond conventional relevance- and redundancy-based criteria.

We perform incremental feature selection experiments to examine how the inclusion of additional exogenous features affects forecasting performance. We find that adding features in correlation order (33% to 67% to all) reduces average error in only 8 of 16 transitions, suggesting that incorporating more highly correlated features does not necessarily improve forecasting due to potential temporal instability. We further compare TimeLess with correlation-based and random selection as controls. TimeLess achieves better performance in 24/32 and 29/32 configurations, respectively, supporting the value of temporal stability as an additional criterion for selecting exogenous features. Detailed results are available at https://anonymous.4open.science/r/cdd_iclr-4CFD/incremental.pdf.

We acknowledge that modeling context-dependent feature interactions remains challenging, requiring rigorous causal analysis and domain knowledge to distinguish meaningful interactions from spurious associations. We view this as an important direction for future work but currently beyond our scope.

Q1: Theoretical status of low CDD-std

We agree that Prop. B.1 provides a one-sided guarantee: large product discrepancy indicates a change in the joint distribution, but low CDD-std does not formally guarantee reliability. Accordingly, our filtering strategy conservatively removes features with strong evidence of instability rather than treating all retained features as necessarily reliable.

Empirically, CDD-std is strongly associated with the benefits of exogenous features across the 11 development datasets (Spearman ρ = -0.87, p < 0.001). Moreover, removing high-CDD-std features improves forecasting on unseen datasets and models without degrading performance on stable datasets (Tables 1, 3, and 16).

W2 / Q3: Cross-bin regime shifts

We can explain this with our synthetic examples. In Example 2.2, temporal bins are aligned with regime intervals (each bin only contain one frequency), so the feature-target relationship is stable within each bin but differs across bins. Despite this within-bin stability, the changing feature has a CDD-std of 0.018, approximately 2.6× higher than the stable-frequency signal in Example 2.1 (0.007). This comparison shows that CDD-std can capture cross-bin changes in feature--target relationships, even when each bin is internally stable.

W3 / Q4: Threshold calibration

Our threshold is an empirical operating point under the default setting (K=8), rather than a universal or theoretically optimal bound. Validation-only calibration is impractical because the validation periods contain only 1–2 temporal bins. To assess calibration robustness, we conduct leave-one-dataset-out (LODO) evaluation, where the resulting feature selections differ from the default by only 3.7% on average results, detailed results at https://anonymous.4open.science/r/cdd_iclr-4CFD/lodo.pdf.

We acknowledge that the hard-filtering results on development datasets are not fully independent of threshold calibration. Nevertheless, the threshold also transfers to unseen datasets and models (Tables 3 and 4), while TimeLess and fixed-budget feature ranking provide complementary evaluations that do not rely on this threshold.

W4: Hyperparameter robustness

K is the primary hyperparameter introduced by CDD. We use K=8 throughout, with the threshold calibrated accordingly. Although changing K affects the absolute CDD-std scale and may require recalibration, feature rankings remain highly consistent across K values (Spearman ρ = 0.93; Table 8). Thus, ranking-based selection is robust to the choice of K.

Thank the authors for the efforts in preparing the response. The rebuttal provides further clarification and empirical analysis, which help to better explain the intended role of CDD-std. However, my primary concern has still not been fully addressed. In particular, the current evidence is insufficient to establish the reliability and general applicability of CDD-std as a criterion for identifying temporally stable features. The additional synthetic examples and overall comparisons offer useful insights, but their performance still needs to be evaluated more systematically, especially their effectiveness under different regime changes and conditional feature interactions, so as to further support the related claims. Moreover, the robustness of threshold calibration should be assessed more directly through its downstream predictive performance on unseen datasets.

AuthorsOpenReview ↗

Response to Reviewer dEvd

We thank the reviewer for the constructive feedback. We agree that evaluating CDD-std under different regime changes, feature interactions, and threshold calibration settings is important for understanding its reliability and applicability. In response, we have conducted additional synthetic experiments and leave-one-dataset-out (LODO) evaluations. We summarize our findings below.

Systematic evaluation under different regime changes and feature interactions.

We have added more synthetic experiments covering three representative types of temporal regime changes: sign flips, vanishing effects, and changes in effect magnitude. Across all three scenarios, CDD-std consistently achieves higher feature selection accuracy and better downstream forecasting performance than both correlation-based selection and mRMR on these setups. These results provide additional evidence that CDD-std can effectively identify temporally stable feature--target relationships under different types of non-stationarity. We also include a pure interaction scenario to examine the scope of CDD-std. Since CDD-std evaluates temporal stability at the individual-feature level, it cannot identify features whose predictive contributions arise exclusively from their joint interaction, as neither feature exhibits an independently stable feature--target relationship in this setting. In contrast, the joint interaction between the two features remains predictive, allowing mRMR to recover exact feature sets. Detailed experimental settings and results are provided in https://anonymous.4open.science/r/cdd_iclr-4CFD/synthetic.pdf.

Discussion regarding feature interactions.

We appreciate the reviewer for raising the important question of conditional feature interactions. We acknowledge that CDD-std is less effective than mRMR in the synthetic interaction scenario, as well as in settings involving feature redundancy. However, we believe this limitation should be interpreted in the context of the intended objective of CDD-std.

First, many commonly used feature selection approaches, such as correlation-based selection, are evaluated on individual features. CDD-std follows this individual-feature perspective, with a different focus: rather than measuring feature relevance alone, it assesses the temporal stability of individual feature-target relationships. Although mRMR additionally considers feature redundancy, it does not provide a general solution to arbitrary conditional or higher-order feature interactions.

Second, although the synthetic interaction scenario provides a controlled setting where the interacting features are explicitly specified and fully observed, this assumption may not hold in practical time-series datasets. In real-world applications, the observed feature set may not contain all variables involved in the underlying interactions. In other words, even when a feature's predictive contribution depends on interactions with other variables, its all interaction features and partners can hardly be fully included in time-series forecasting datasets. Consequently, feature selection based solely on the observed features may fail to capture the underlying joint relationships. This issue is discussed in [1], where the observed inputs may not contain all information necessary to sufficiently determine the interactions. Therefore, explicitly accounting for feature interactions does not necessarily lead to more reliable feature selection in practical settings. Empirically, our evaluation results show that mRMR, which accounts for feature redundancy, can underperform simpler correlation-based feature selection on part real-world time-series datasets.

Third, despite this limitation, our empirical evaluations demonstrate that CDD-std achieves competitive and often better downstream forecasting performance than both correlation-based selection and mRMR across real-world time-series datasets. This supports the practical value of measuring temporal stability at the individual-feature level, even though it does not explicitly model joint interactions.

Finally, jointly modeling temporal stability and conditional feature interactions is an interesting direction for future research. We consider it complementary to the current objective of CDD-std, which is to provide a simple and effective criterion for identifying temporally stable individual feature-target relationships. Developing a unified framework that accounts for both aspects is beyond the scope of this work.

AuthorsOpenReview ↗

Response to Reviewer dEvd (Contd.)

Robustness of threshold calibration on unseen datasets on downstream forecasting tasks.

We evaluate the downstream predictive performance under these LODO-calibrated thresholds on unseen datasets on iTransformer and TimeMixer. Across 48 dataset-model-horizon comparisons, 28 remain unchanged relative to the default threshold, while 8 of the remaining 20 comparisons show an average relative MSE change of -0.022% across all 48 comparisons, and -0.054% among the 20 comparisons with changed feature selections. A two-sided sign test on the changed comparisons yields p=0.503, showing no systematic tendency toward either improvement or degradation. These results suggest that recalibrating the threshold on different subsets of datasets does not lead to a systematic loss in downstream forecasting performance. Detailed LODO results are updated in https://anonymous.4open.science/r/cdd_iclr-4CFD/lodo.pdf.

[1] Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning. ICML

Reviewer_VfYbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on cases where adding exogenous features (series other than the target) worsens time-series forecasts. It proposes feature selection based on the temporal stability of feature-target relationships. This paper standardizes past values of an exogenous feature and future values of the target and forms a distribution of their products. It computes Conditional Distribution Difference (CDD) to measure the difference between the distribution in each window and a reference distribution for the time interval containing that window. It uses CDD-std as the selection criterion. CDD-std is the standard deviation of the CDD values across time intervals divided by the absolute value of their mean. This paper presents hard filtering, which removes features with large CDD-std, and TimeLess, which takes CDD-std as input to a multilayer perceptron (MLP) to compute a weight for each feature. TimeLess keeps features with weights at or above the mean and applies those weights to the retained features. It jointly trains the MLP and the forecasting model to reduce forecasting error. This paper examines the properties of the measure on synthetic data. It reports forecasting improvements over using all exogenous features on several real datasets and forecasting models.

Critical strengths

Feature selection based on CDD-std reduces forecasting error in many settings with iTransformer, TimeMixer, CrossLinear, and TimeFilter. These improvements are over Base, where the same forecasting model uses all exogenous features. Hard filtering keeps the threshold at 0.223, the value determined on other datasets. Even with this unchanged threshold, hard filtering reduces forecasting error compared with Base on Neuro and SP500. There are also average improvements over existing feature selection methods. Although the comparisons are limited, the results support the practical value of improving input feature selection across several forecasting models.

Critical weaknesses
  • W1 (Additional value specific to CDD-std): Prior work [1–3] uses the stability of feature-target relationships for prediction. However, the main feature selection baselines in this paper do not account for changes over time. There is no comparison with feature selection based on variation in correlation across time intervals. On the six datasets in the main experiments, TimeLess gives a small average improvement over univariate forecasting (forecasting using only the target’s own past values).

  • W2 (Structure of TimeLess and interpretation of the control experiment): It is unclear whether TimeLess applies a shared multilayer perceptron (MLP) to each feature separately or feeds the CDD-std values of all features into the MLP together. This difference affects whether the MLP can learn different weights for different features even with constant input. This makes it difficult to judge what the comparison with constant input shows about the contribution of CDD-std.


References:

  • [1] Rojas-Carulla et al. Invariant Models for Causal Transfer Learning. JMLR 2018.

  • [2] Pfister et al. Stabilizing variable selection and regression. Annals of Applied Statistics 2021.

  • [3] Liu et al. Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning. ICML 2024.

Questions for authors
  • Q1: Could you provide a comparison with feature selection based on variation in correlation across time intervals? I would like to check whether the distribution comparison used by CDD-std gives additional value over a simple stability measure.

  • Q2: Could you clarify whether the MLP in TimeLess takes the CDD-std of each feature separately or the CDD-std values of all features together? Please also state the order of weighting and normalization. This structure affects how much the comparison with constant input can tell us about the contribution of CDD-std.

  • Q3: Could you compare TimeLess with univariate forecasting using TSMixer on Neuro and SP500? I would like to check whether TimeLess improves over univariate forecasting even in settings where it greatly improves over using all exogenous features.

1 comment
AuthorsOpenReview ↗

Rebuttal to Reviewer VfYb

W1 (a) / Q1: Value of CDD-std beyond correlation variability

Our key contribution is to quantify the cross-bin temporal stability of individual feature--target relationships for feature selection. CDD-std first measures deviations of local sliding-window relationships from a reference distribution within each temporal bin, and then quantifies the variability of these deviations across bins. The Wasserstein distance is used for the within-bin comparison because it provides the one-sided theoretical guarantee in Prop. B.1; other discrepancy measures can also be incorporated (App. B.2, Lines 892--899).

We compare CDD-std with correlation-based alternatives under the same temporal bins, windows, and data splits. Across 11 development datasets, CDD-std achieves the strongest association with the gain from exogenous features (Spearman ρ = -0.87, p = 0.0009), compared with static correlation (ρ = 0.13), per-bin correlation (ρ = -0.59), and sliding-window correlation (ρ = -0.72). This highlights the importance of comparing local relationships against a within-bin reference. Detailed results are available here.

W1 (b): Relation to prior works

We thank the reviewer for pointing out these related works. While [1--3] also exploit stability across environments, they do so through invariant subset selection, stabilized regression, or representation learning. In contrast, our work explicitly quantifies temporal stability at the individual-feature level, independently of the forecasting model. This makes CDD-std a model-agnostic feature screening method, applicable even to frozen forecasting models such as Chronos-2 (Table 4).

W1 (c): Comparison with univariate forecasting

Whether univariate forecasting or using all exogenous features performs better varies across datasets and can only be determined by evaluating their test performance. However, in practice, such a decision cannot be made based on test results, as test information is unavailable at deployment time. Our feature selection addresses this challenge with a single adaptive procedure:

  • On stable datasets, we retain the benefits of exogenous features, outperforming univariate forecasting in all 8 model--dataset pairs (Table 16).
  • On the six main datasets with less stable feature-target relationships, We mitigates the degradation caused by using all exogenous features, outperforming Base in 21 of 24 pairs, while remaining comparable to univariate forecasting (12 wins, 7 ties (within 1% relative MSE) out of 24 pairs, Tables 9--12).

Thus, while univariate and multivariate forecasting each have their advantages in different scenarios, always relying on univariate forecasting would forgo the potential benefits of useful exogenous features. Our apporach preserves the benefits of both: it exploits exogenous features when they are helpful while suppressing them when they are harmful.

W2 / Q2: Structure of TimeLess and the constant-input control

TimeLess independently learns a score for each exogenous channel based on its CDD-std features. The channel-wise scores are converted into normalized weights, which are used to generate a learnable feature mask. Each channel is first instance-normalized over the lookback window, and the learned mask is then applied to the normalized features before they enter the backbone.

To isolate the contribution of CDD-std, we replace its input with a constant in the ablation. This removes the information of CDD-std to distinguish feature stability, while we adding a learnable feature mask so that the model can still select a subset of features. This control does not significantly improve over Base, whereas incorporating CDD-std consistently reduces both MSE and MAE. This demonstrates that a learnable feature mask alone for feature selection is insufficient; CDD-std provides essential guidance for effective feature selection.

Q3: TimeLess vs. univariate TSMixer on Neuro and SP500

We conduct additional experiments comparing univariate forecasting (Uni) with CDD-std-guided selection using TSMixer under the same experimental setup. TimeLess consistently outperforms Uni on both Neuro and SP500. Threshold-based filtering also improves over Uni on Neuro. TimeXer is omitted because it does not support univariate forecasting (as discussed in Appendix C.3, Lines 991-992). Detailed results are available at https://anonymous.4open.science/r/cdd_iclr-4CFD/tsmixer.pdf.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.