acceptodds
Under review as a conference paper at ICLR 2027

GLARE: Global and Local Reconstruction for Time Series Forecasting with Exogenous Variables

Abstract

Exogenous variables can improve time series forecasting through their relationships with the endogenous variable. However, relationships estimated over the full training period may differ from those fitted to recent observations. We propose GLARE, a lightweight forecasting framework that reconstructs the endogenous variable from exogenous variables through global and local relationships. An MLP captures global relationships through a pointwise mapping learned over the full training period. Ridge regression captures local relationships by fitting coefficients to the lookback window of each instance. The global and local reconstructions are blended into a single exogenous reconstruction, and a residual forecaster predicts the remaining residual. On the 12 TSF-X datasets, GLARE achieves state-of-the-art forecasting performance with a single linear residual forecaster and only hundreds to a few thousand additional parameters. Our code is available at: https://anonymous.4open.science/r/GLARE-6350.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2gsZOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

GLARE predicts the target from the exogenous variables in two ways: a small per-time-step MLP trained on the whole training set, and a ridge regression fitted on each lookback window after a learned low-pass filter. The two fits are averaged with a fixed weight, and a linear model forecasts what is left over. On 12 TSF-X datasets it beats DAG, GCGNet, CrossLinear, TimeXer, TiDE and TFT in 22 of 24 settings with a few thousand extra parameters.

Critical strengths
  1. The model is very small and fast, and the paper reports parameters, FLOPs, memory and run time against the baselines.
  2. The writing is clear, and the ablations cover each part: global vs. local fit, fit vs. residual forecaster, filter choice and MLP choice.
  3. Code is released, and GLARE results are averaged over three seeds with standard deviations in Table 7.
Critical weaknesses
  1. Nothing here is new. A per-step regression from covariates to the target, a rolling regression on the recent window (the paper itself cites Brown et al., 1975), and a linear model on the residual: this is regression with ARIMA-style errors. Blending the two fits with one fixed weight picked from five values is not a new idea. Section 4 adds nothing either: it bounds the norm of the ridge operator, which says nothing about accuracy. The motivation is just as thin: Figure 1 shows that a local fit sometimes beats a global fit (in 23% to 94% of windows), which is true of almost any data.

  2. GLARE wins because the benchmark hands it the future. Every method gets the true values of the covariates over the whole forecast window, and the paper never says whether these are forecasts or recorded values. For wind power, if they are the recorded weather, predicting power is just reading it off at the same time step. The ablation shows exactly this: on Sdwpfm1 at horizon 360, the MLP alone matches the full model (MSE 0.331), adding the forecaster makes MAE worse (0.400 to 0.412), and the error hardly grows from 24 to 360 steps (0.301 to 0.331). When future covariates are not given and must be predicted, the setting that matters in practice, GLARE is only "comparable to DAG" (Appendix D.3). The headline result is about the benchmark, not about a better forecaster.

  3. I cannot tell whether the comparison is fair. Some baseline numbers are copied from the DAG paper and some are rerun, and the paper does not say which. GLARE is tuned over 60 settings per case (4 learning rates, 5 values of α, 3 values of λ), with no word on whether they were picked on validation or test data, while copied baselines keep whatever tuning their authors used. GLARE also sets drop_last=False; if the copied numbers used True, the two sides were scored on different test sets. The plugin study compares against an "MLP fusion" that is often worse than using no covariates at all (PatchTST on Energy: 0.108 to 0.239; on Colbun: 0.217 to 0.417), so beating it means little.

  4. The extra benchmarks do not make sense as forecasting tasks. On Electricity and Traffic, the target is one series and the other 320 or 861 series are fed in as known future inputs (Table 6). No real forecaster has that, so Table 8 says nothing about general use. Figure 3 claims an "association" from 24 points split across four dataset types. Table 2 packs 24 numbers into each row. Facts a reader needs, such as where the future covariates come from and how hyperparameters were chosen, are missing.

Questions for authors

See weaknesses.

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

The paper addresses forecasting of a scalar endogenous series when exogenous variables are known over both the lookback window and the forecast horizon (TSF-X). GLARE reconstructs the endogenous series from the exogenous ones in two ways and forecasts what remains: a pointwise one-hidden-layer MLP (width 32) shared across time steps and instances (Sec. 3.2), and a per-instance closed-form ridge regression fitted on the lookback window to spectrally filtered exogenous variables and applied over lookback and horizon (Sec. 3.3, Eqs. 4–7). The two reconstructions are blended with a dataset-level constant α\alpha, and a single linear layer maps the lookback residual to the future residual (Sec. 3.4). Section 4 gives operator-norm bounds on how a perturbation of the endogenous history propagates through the refitted ridge step. On the 12-dataset TSF-X benchmark of DAG, Table 1 reports the lowest MSE and MAE in 22 of 24 settings against DAG, GCGNet, CrossLinear, TimeXer, TiDE and TFT (average MSE reductions over the best baseline of 9.0% at the short and 18.5% at the long horizon); Table 2 uses GLARE as a plugin for four univariate backbones; Tables 3–5 and Figs. 5–7 ablate components and hyperparameters; Table 11 reports cost; Table 10 covers the setting without future exogenous inputs. GLARE's results are three-seed means (Table 7).

Critical strengths

Strength 1. A large, consistent and cheap improvement on a benchmark the field is actively using (Tables 1, 7, 11). At the long horizon GLARE is first on both metrics in all 12 datasets as printed, with MSE reductions over the best listed baseline of 27.1% on NP (0.380 vs 0.521), 45.5% on Energy (0.091 vs 0.167) and 23.7–29.9% on the four wind sets, all far larger than GLARE's three-seed standard deviations. This costs 4.31K parameters and 1.07 ms per batch on NP at L=168L=168 against 0.39M–2.34M parameters and 1.2–9.4 ms for the baselines, with the lowest inference time in all six blocks of Table 11. It indicates that, once future exogenous variables are known, explicit reconstruction of the exogenous contribution is a competitive and inexpensive alternative to attention or graph machinery.

Strength 2. The plugin and ablation experiments are clean and informative (Tables 2–4). In all 192 cells of Table 2 the GLARE interface improves on both the bare backbone and MLP fusion, while MLP fusion, the protocol DAG and GCGNet use for univariate models, is worse than the backbone without future covariates in 53 of 192 cells (e.g. PatchTST on Energy-24: 0.108 → 0.239; for TimeKAN, xPatch and PatchTST the Base and +M cells are GCGNet's published numbers, so only +G is the authors' own). Table 3 shows that neither branch alone is uniformly preferable and the blend wins or ties in 9 of 10 settings; Table 4 shows the spectral filter beating identity, temporal-linear and convolutional alternatives in 9 of 10 settings.

Critical weaknesses

Weakness 1. The comparison in Table 1 numerically matches the DAG paper's baseline table, understates GCGNet, includes TimeXer entries whose use of future exogenous variables is undocumented, and does not state the hyperparameter selection split (Secs. 5.1–5.2, Appendix C.2, Fig. 5).

Every baseline cell in Table 1 matches the per-horizon Table 5 of the DAG paper. Its GCGNet and CrossLinear values are those of GCGNet's own appendix Table 8, a controlled setting "where each model uses a fixed configuration across forecasting horizons"; GCGNet's main-text Table 1 is better in 22 of 24 settings (e.g. Rapel 10 0.140/0.191 vs 0.211/0.230 here; Rapel 30 0.284/0.327 vs 0.401/0.384). Substituting that column, GLARE (0.179/0.223 and 0.304/0.340 on Rapel) loses both Rapel settings on both metrics and falls behind on Colbun-10 (0.057/0.084 vs 0.058/0.093) and BE-24 MSE (0.323 vs 0.328): the "22 of the 24 settings" count becomes 19 and the sentence that GLARE "ranks first on both metrics across all 12 datasets" at the long horizon no longer holds. The long-horizon margins on NP, DE, Energy and the wind sets remain 17–46% after this substitution, so the main result stands, but a matched-protocol comparison is needed.

Sec. 5.1 lists TimeXer among methods that natively support future exogenous variables, yet TimeXer's formulation (Wang et al., 2024b, Eq. 1) conditions only on lookback exogenous values, and the paper's own introduction contrasts TimeXer with methods that "go beyond historical observations". In DAG's Tables 5 and 6, TimeXer's numbers are identical with and without future exogenous inputs in 11 of 24 settings (all eight wind settings, both Rapel settings, BE-24), which happens for no other model, even though DAG's released adapter appends the future values to the exogenous tokens. How future inputs were supplied to TimeXer should be documented.

Appendix C.2 lists four learning rates, five α\alpha and three λ\lambda values (60 configurations) but not the selection split; "validation" appears only in the split description of Appendix C.1, and Fig. 5 marks the chosen values at the minima of test-MSE curves, which is consistent with either validation- or test-based selection. Some narrow wins depend on the choice: on FR-360, α≥0.5\alpha\ge0.5 gives roughly 0.474, behind TFT (0.465) and no better than DAG (0.473); on PJM-360, α=0.1\alpha=0.1 gives about 0.131, behind GCGNet (0.129). Relatedly, the identity-filter variant on Colbun-30 (Table 4, 0.458) is worse than global-only (0.219), local-only (0.181) and H-only (0.225) in Table 3, which suggests the ablation variants were run with the full model's hyperparameters.

Weakness 2. The central explanation, global–local complementarity, is only weakly supported by the paper's own ablation, the simple baselines that would isolate the mechanism are missing, and the covariates are close to oracle values (Table 3, Table 10, Sec. 5.2).

In Table 3 the global-only variant (pointwise MLP plus linear residual forecaster) already beats DAG on NP-24 (0.190 vs 0.202), NP-360 (0.392 vs 0.521) and Sdwpfm1-24 (0.310 vs 0.351), and equals full GLARE on Sdwpfm1-360 (0.331 vs DAG 0.495); local-only beats DAG on DE-24 and DE-360. A single branch thus beats the best listed baseline in 6 of 10 ablated settings; the blend improves on the better branch by 2.8–8.3% in eight settings, ties once and is 7.9% worse on DE-24, with Colbun-30 (29.8%) the exception, and Table 3 omits Energy, the dataset with the largest gain. The association in Fig. 3 between local win rate and gain over DAG may partly reflect horizon, and for the two wind sets in Fig. 5 the chosen α\alpha is 0.7 and 0.9, i.e. mostly global.

The natural controls are absent: a flat ridge or ARX regression from [Xendo;Xexo;Yexo][X^{\text{endo}};X^{\text{exo}};Y^{\text{exo}}] to YendoY^{\text{endo}} with the same tuning budget, LEAR from the Lago et al. (2021) EPF benchmark that supplies five of the datasets (itself an average of ARX models recalibrated on short and long windows), and NBEATSx, which the paper cites. "Reconstruct from covariates, then model the residual" is closely related to regression with ARIMA errors; the per-window closed-form ridge fit is the core of DeepTime (Woo et al., ICML 2023), uncited; AutoGluon-TimeSeries' covariate regressor and TimesFM's in-context covariate mode implement the same two-step structure; DynaME (Hong et al., 2026), cited only as an online method, already combines a stable general expert with a committee of experts fitted to recent periodic patterns; and the complex filter of Eq. 4 closely resembles FilterNet (Yi et al., NeurIPS 2024).

Finally, the Sdwpf covariates are ERA5 reanalysis weather, Energy uses recorded generation of the other sources, and the EPF day-ahead forecasts are used up to 360 hours ahead, so the benchmark rewards contemporaneous regression on near-oracle covariates. GLARE is no longer dominant when covariates must be predicted (Table 10: 8 MSE wins vs DAG's 11; Sdwpfh1-360 moves from 0.355 to 0.911), although Appendix D.3 also changes the pipeline. If the margin comes mainly from regressing on near-oracle covariates, the contribution is an important baseline finding about the TSF-X benchmark, and the claims about the incremental value of global–local blending should be narrowed.

Questions for authors
  1. [clarification] Which cells in Tables 1, 2, 8 and 10 are reused from the DAG or GCGNet papers and which are your reruns? Will you provide a matched-protocol comparison addressing GCGNet's stronger main-table results and revise the 22/24 count and the "all 12 datasets" statement?

  2. [clarification] How were future exogenous variables supplied to TimeXer in Table 1, given its lookback-only formulation and the 11 settings in which DAG's TimeXer numbers are identical with and without future inputs? Your Table 10 TimeXer numbers differ from DAG's in every setting, so you appear to have rerun it there; did you also rerun it with future inputs?

  3. [clarification] On which split were the learning rate, α\alpha and λ\lambda selected for Tables 1–5? If on validation, please add validation curves next to Fig. 5. Please also explain the Colbun-30 identity-filter result (0.458 vs 0.219 for global-only).

  4. [suggestion] Please add a ridge/ARX baseline from [Xendo;Xexo;Yexo][X^{\text{endo}};X^{\text{exo}};Y^{\text{exo}}] to YendoY^{\text{endo}} with matched preprocessing and tuning, and the global-only variant as a row in Table 1 for all 12 datasets; LEAR on the EPF datasets and NBEATSx would strengthen the comparison further. If GLARE clearly beats these, the concern in Weakness 2 is largely resolved.

  5. [suggestion] Please report an experiment with realistically available covariates (forecast error growing with lead time, or only forecasts available at each prediction origin), since Table 10 indicates the advantage is tied to covariate quality.

  6. [clarification] Eqs. 5–6 have no intercept. Is the data standardised with training-set statistics only, did you try a constant column in AA, and does the OLS comparison behind Fig. 1 include an intercept? How does Local Exogenous Regression differ from DeepTime's closed-form ridge inner loop and from DynaME's dynamically fitted experts apart from the choice of regressors?

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

This paper tackles time series forecasting with exogenous variables by separating the forecasting task into global and local exogenous reconstruction followed by residual forecasting. The proposed framework reconstructs the target series by combining a pointwise MLP learned across the training set to capture static global exogenous-to-endogenous relationships, and a ridge regression fitted per instance to the lookback window over spectrally filtered exogenous series to capture local dynamics. The remaining lookback reconstruction error is then passed to a linear residual forecaster to project the future residual. While the framework achieves strong benchmark results when future exogenous variables are known, its performance relies heavily on oracle future covariates and degrades notably in realistic settings where future exogenous series must be estimated.

Critical strengths

Propose an elegant method: The proposed method is efficient and provides a notable performance improvement in scenarios with known covariates.

Effective experimental analysis: Comprehensive experiments and benchmarks are provided to evaluate the method both as a standalone forecaster and an adapter.

Comments post rebuttal: I have discussed this in detail with the student who wrote the first review. Although I agree that the domain that experiments are made in are not an excellent match for exogeneous variable forecasting, as the availability of exogenous variables in those domains are highly suspect, the issue is an appropriate one more generally in many other contexts.

More generally the responses to the questions are good, and we are both satisfied that this has been implemented and tested sensibly. For this reason we have upgraded the score on this paper, and we do recommend acceptance of the paper. The theory still seems a bit decorative, and the authors could consider whether some of that should not make the cut.

Critical weaknesses

Overreliance on known future covariates: The method's performance largely relies on known covariates significantly degrading when these are not available, despite this when used as an adapter it is compared to base methods which do not use this information, or using a simple MLP fusion as opposed to more modern methods like cross attention.

The theoretical analysis in Section 4 does not provide generalization or excess risk bounds. Proposition 1 is a direct consequence of the eigenvalues of the symmetric ridge projection matrix PλP_\lambda lying in [0,1)[0, 1). Furthermore, the bound in Proposition 2 scales with ∥Gfut∥2\Vert{}G_{fut}\Vert{}_2. In long-horizon settings and with small regularization parameters such as λ=0.1\lambda = 0.1, the spectral norm can be large, making the constant factor loose and uninformative in practice.

The learnable spectral filter weights WW directly parameterize the design matrix AA. Because s∗s^* involves the closed-form inversion (A⊤A+λI)−1(A^\top A + \lambda I)^{-1}, backpropagating task loss gradients into WW requires differentiating through this matrix inversion. The paper lacks discussion on whether gradients are tracked through this inversion or detached, and does not evaluate numerical stability/conditioning during early optimization steps.

Questions for authors
  1. How are gradients dealt with for the matrix inversion? Is this stable during early training?

  2. Aside from this benchmark can the proposed method be useful in downstream applications given the significant performance degradation without known exogenous futures?

  3. For the plugin experiments, why was MLP fusion chosen as the primary baseline rather than modern cross-attention or temporal conditioning mechanisms?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.