acceptodds
Under review as a conference paper at ICLR 2027

Complementary Retrieval Geometries for Enhancing Multivariate Time Series Forecasting

Abstract

Retrieval-augmented forecasting provides multivariate time-series forecasters with additional evidence by reusing trajectories that follow similar historical states; its effectiveness depends on how predictive historical similarity is defined. Existing methods typically organize retrieval around one principal criterion, causing a target variable's temporal shape and its dynamic role in a multivariate system to share the same similarity definition. We propose CoReG, a complementary retrieval framework for multivariate time-series forecasting. CoReG constructs two semantically distinct retrieval geometries: TimeGeo matches the target variable's multiscale temporal shape, while RoleGeo matches its system role using stable cross-variate relations and time-aligned role responses. The two paths independently retrieve the target variable's own historical future segments, form separate views relative to the same frozen backbone forecast, and produce the final prediction through symmetric merging. Across ten datasets and six forecasting backbones, CoReG improves both MSE and MAE in 58 of 60 paired settings, with mean relative reductions of 14.94% and 10.52%, respectively. Anonymous code is available at https://anonymous.4open.science/r/CoReg-3CCD/.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time-Series Forecasting, Retrieval-Augmented Forecasting, Forecasting Enhancement, Cross-Variate Relations

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_KyqsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes CoReG, a retrieval-augmented framework for multivariate time-series forecasting. TimeGeo retrieves histories with similar multiscale temporal shapes, while RoleGeo uses representations informed by stable cross-variable relations. Each path retrieves target-owned training futures and combines them with a frozen forecasting backbone. The two views are then averaged. Experiments cover ten datasets and six backbones.

Critical strengths

CoReG frames retrieval augmentation around complementary predictive similarities, separating multiscale temporal-shape matches from relation-informed system-role matches, and implements them as two plug-in paths around a frozen forecaster. The evaluation spans ten datasets and six backbones, reports horizon-averaged MSE improvements in 58 of 60 dataset–backbone settings, and includes ablations, data-sparsity, multi-seed, and accuracy–latency analyses.

Critical weaknesses
  1. Innovation: The necessity of independent retrieval geometries is not established. The paper argues that a single space mixes evidence, but does not compare against a capacity-matched joint-space retriever. TimeGeo Only, RoleGeo Only, and Shared Readout do not provide this comparison, so the claim that separate spaces are necessary remains unsupported.
  2. Experiments: Table 2 lacks time-series forecasting backbones proposed in 2026.
  3. Presentation: The method and training procedure are difficult to follow. The architecture and training stages are scattered across sections, while key encoder details, including the contrastive objective and positive-pair selection, are deferred to Appendix B. A concise training flow in the main text would improve clarity and reproducibility.
Questions for authors

Could you compare CoReG with a capacity-matched single retriever that jointly embeds temporal shape and system-role information, and include a 2026 forecasting backbone in Table 2?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Kyqs

Thank you for identifying these gaps. We have added the requested capacity-matched joint retrieval comparison on seven datasets, DTAF (2026) on all ten datasets, and a concise main-text training algorithm with explicit encoder supervision and positive selection. All revisions are highlighted in blue in the revised manuscript. Revisions to figures and tables are indicated by blue captions.

1. Capacity-matched joint retrieval

Joint-CM uses one encoder to jointly learn temporal shape and system role through equally weighted shape-predictive and role-predictive contrastive losses. It selects neighbors once in a shared 64-dimensional space and forms one retrieval forecast for fusion with the frozen DLinear Anchor. Joint-CM-Large increases memory from 1,000 to 2,000 and Top-K from 50 to 200, matching FULL's total memory-entry budget and nominal Top-K selection count. Both variants match FULL's added-module parameter count within 0.04% at each horizon, excluding the shared Anchor. The comparison uses historical paired FULL results and the corresponding frozen Anchor checkpoints for the new variants. Table R2 reports four-horizon mean errors across all seven datasets.

Table R2. Capacity-matched joint retrieval comparison. Each cell gives four-horizon mean MSE / MAE. Values are displayed to four decimals.

DatasetFULLJoint-CMJoint-CM-Large
ETTh10.5020 / 0.50850.5034 / 0.50980.5036 / 0.5099
ETTh20.2932 / 0.37720.3020 / 0.38250.3049 / 0.3850
ETTm10.3810 / 0.41550.3827 / 0.41630.3837 / 0.4155
ETTm20.2010 / 0.29570.2089 / 0.30230.2090 / 0.3021
Weather0.2282 / 0.27470.2297 / 0.27780.2285 / 0.2762
Exchange0.2929 / 0.40120.3540 / 0.42730.3117 / 0.4131
PEMS080.2902 / 0.28370.2885 / 0.28490.2943 / 0.2913

FULL has lower mean MSE on six of seven datasets and lower MAE on all seven compared with Joint-CM. Compared with Joint-CM-Large, FULL has lower mean errors for both metrics on all seven datasets, as determined from unrounded results. Joint-CM achieves lower MSE on PEMS08. These results support the predictive benefit of independent retrieval under the tested capacity-matched conditions, including the enlarged single-retriever budget.

Revision locations: Section 4.3 and Table 3 (p. 8, lines 404-431). Appendix E.2 (p. 53, lines 2808-2827, Table E.4, and p. 54, lines 2863-2882, Table E.5).

2. DTAF (2026) in Table 2

We have added DTAF [1] to Table 2, extending the evaluation to seven forecasting backbones. The added experiments cover all ten datasets at horizons 96, 192, 336, and 720. Table R3 reports the relative reduction in each dataset's four-horizon mean error from the paired DTAF baseline.

Table R3. Relative error reductions after adding CoReG to DTAF (2026). Positive values indicate improvement.

DatasetMSE reduction (%)MAE reduction (%)
ETTh11.460.82
ETTh21.501.12
ETTm12.031.15
ETTm20.050.29
Weather1.892.54
Exchange6.903.88
Electricity7.146.44
PEMS0439.9224.23
PEMS0821.0123.28
Solar6.0413.77
Mean across datasets8.797.75

CoReG improves both dataset-mean metrics on all ten datasets, with mean relative reductions of 8.79% in MSE and 7.75% in MAE.

Revision locations: Section 4.1 (p. 7, lines 360-362), Section 4.2 (p. 7, lines 375-377), and Table 2 (p. 8, lines 379-401). Appendix E.1, Table E.1 (p. 49, lines 2602-2639) and Table E.2 (p. 50, lines 2655-2691).

3. Training flow in the main text

Algorithm 1 now presents the complete training sequence in one place. It trains and freezes the Anchor, trains and freezes the predictive encoders, constructs the per-variable memories, and then optimizes the fusion modules using the final forecasting loss. The Anchor, encoders, and memories remain fixed during fusion training. The main text now specifies the supervision signature, positive selection, valid negatives, exact contrastive objective, shared encoder architecture, normalization, and retrieval protocol. Together, these additions make the training procedure and its required definitions directly accessible in the main text.

Revision locations: Section 3.6 (p. 7, lines 324-346, Algorithm 1 and Eq. 22), Section 3.4 (p. 6, lines 288-304, Eqs. 16-18), and Section 3.2 (pp. 4-5, lines 199-227, Eqs. 2-5).

References

[1] Junkai Lu, Peng Chen, Chenjuan Guo, Yang Shu, Meng Wang, and Bin Yang. Towards Non-Stationary Time Series Forecasting with Temporal Stabilization and Frequency Differencing. Proceedings of the AAAI Conference on Artificial Intelligence, 40(29), 24070-24078, 2026. Publisher

Thank you for the detailed response and revisions. I will increase my Rating from 2 to 3.

  1. Independent retrieval geometries. The added capacity-matched joint retrieval baselines (Joint-CM, Joint-CM-Large, Concat) address my concern. With parameters, memory, and Top-K aligned, FULL consistently performs better, supporting the effectiveness of independent retrieval geometries.

  2. 2026 backbone. Adding DTAF (Lu et al., AAAI 2026) as the seventh backbone, with the ten datasets and four horizons unchanged, meets my original request. Table 2 and the abstract have been updated accordingly.

  3. Structure and readability. Section 3.6 and Algorithm 1 in the main text, together with the key encoder details moved from the appendix, make the paper easier to follow and reproduce.

  4. Reference formatting. The reference citations would benefit from careful checking and unification: (1) URL consistency: some references with stable links lack URLs, e.g., Fei et al., Amplifier: Bringing attention to neglected low-energy components in time series forecasting (AAAI 2025) and Lai et al., Modeling long- and short-term temporal patterns with deep neural networks (SIGIR 2018). Please unify the citation format. (2) Mixed ordinal styles: "The Fourteenth International Conference on Learning Representations" and "The 41st International ACM SIGIR Conference…" coexist. Please adopt a single style. (3) Missing venues: some references omit the publication venue, e.g., Stationarity-aware retrieval-augmented time series forecasting was published at KDD 2026. Please add the missing venues.

AuthorsOpenReview ↗

Response to Reviewer Kyqs

Thank you very much for increasing your rating and for recognizing the effectiveness of the capacity-matched retrieval comparison, the addition of DTAF, and the clearer training presentation. We also especially appreciate your careful checking of the reference formatting.

1. URL consistency

We added stable publisher or proceedings URLs for published papers that previously listed only DOIs, including Amplifier: Bringing Attention to Neglected Low-Energy Components in Time Series Forecasting and Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks. We applied the same rule to the other published entries with available stable links.

2. Conference-name formatting

We unified ordinal conference names using numeric forms. For example, the revised bibliography consistently uses “The 11th,” “The 12th,” and “The 14th International Conference on Learning Representations,” together with “The 41st International ACM SIGIR Conference” and “The 42nd International Conference on Machine Learning.”

3. Missing venues and record types

We updated Stationarity-Aware Retrieval-Augmented Time Series Forecasting with its formal KDD 2026 publication entry, including the conference venue, page range, DOI, and publication link. We also removed the volume-number conflict in the RAST entry and updated the relevant proceedings links.

The revised bibliography now uses consistent record types, venue information, DOI and URL fields, and conference-name formatting. The manuscript compiles successfully with zero BibTeX warnings.

Revision location: References.

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

This paper introduces CoReG, a retrieval framework designed to enhance multivariate time series forecaster. The method constructs two independent retrieval paths: TimeGeo decomposes the historical data into multi-scale components; RoleGeo leverages stable cross-variable relationships during training to construct role responses, using both the future of the target and the future of role responses in contrastive learning to formulate an alternative form of historical similarity. Both paths retrieve historical-future segments of the target variable itself, which are then fused with predictions from a shared frozen base model via GeoRead, and the enhanced forecasts from both paths are finally averaged. The method shows significant improvements over state-of-the-art methods on 10 benchmark datasets with comprehensive analysis.

Critical strengths

The way cross-variable information is used in RoleGeo is novel and effective. Previous methods did not carefully design for cross-variable information which is important for multivariate time series forecasting.

Critical weaknesses

None

Questions for authors

The current ablations do not clearly disentangle the contributions of target-future supervision (ft,if_{t,i}), role-response supervision (et,ie_{t,i}), and the variable-specific weighting scheme (cic_i) within RoleGeo. Could the authors compare the original formulation with the following three controlled variants?

  • Target-future supervision only: ci=0c_i = 0, using only ft,if_{t,i}.
  • Role-response supervision only: ci=1c_i = 1, using only et,ie_{t,i}.
  • Equal weighting: a fixed ci=0.5c_i = 0.5 for all variables.
3 comments
AuthorsOpenReview ↗

Response to Reviewer zQDR

Thank you for your positive assessment. We especially appreciate your recognition of RoleGeo's novel and effective use of cross-variable information. We have completed all three supervision variants you requested. All revisions are highlighted in blue in the revised manuscript. Revisions to figures and tables are indicated by blue captions.

1. Supervision ablations

We compare FULL with Target-Future Only (ci=0c_i=0), Role-Response Only (ci=1c_i=1), and Equal Weighting (ci=0.5c_i=0.5) using DLinear on eight datasets and horizons 96, 192, 336, and 720. Each variant retrains RoleEncoder, rebuilds its memory, and trains the fusion modules, keeping the remaining architecture and paired Anchor fixed. Table R1 reports changes in four-horizon mean errors from FULL, which uses variable-specific coherence.

Table R1. RoleGeo supervision ablations. Each cell gives ΔMSE / ΔMAE (%). Positive values indicate higher error than FULL.

DatasetTarget-Future OnlyRole-Response OnlyEqual Weighting
ETTh1+0.03 / +0.00+0.14 / +0.09+0.02 / +0.01
ETTh2+0.58 / +0.24+0.09 / +0.04+0.14 / +0.00
ETTm1+0.07 / +0.06-0.09 / -0.17-0.01 / +0.03
ETTm2+0.51 / +0.11-0.16 / -0.16+0.49 / +0.12
Weather+0.11 / +0.12+0.12 / +0.14+0.07 / +0.07
Exchange+2.94 / +2.18+0.85 / +0.15+2.79 / +2.03
PEMS08+0.29 / +0.01+0.43 / +1.03-0.75 / -0.46
PEMS04+0.24 / +0.12+0.82 / +0.56+0.07 / -0.02

Target-Future Only has higher mean MSE on all eight datasets, supporting the contribution of role-response supervision. Role-Response Only and Equal Weighting each improve mean MSE on two datasets, while FULL gives lower mean MSE on the other six. Equal weighting improves PEMS08 but increases Exchange MSE and MAE by 2.79% and 2.03%. The effects vary across datasets, with several small differences, and variable-specific coherence gives lower errors in most comparisons.

Revision locations: Appendix D.6, RQ6 (p. 37, lines 1947-1974, Table D.4), and Appendix E.7 (p. 65, lines 3458-3482, Table E.20).

Thanks for the response. Please also include more analysis why Equal Weighting is much better on PEMS08. I will keep my postive score. And please make sure your new analysis will be included in the final version.

AuthorsOpenReview ↗

Response to Reviewer zQDR

Thank you for maintaining your positive assessment and for encouraging a closer analysis of PEMS08. Equal Weighting reduces four-horizon mean MSE and MAE by 0.75% and 0.46% relative to FULL, with lower horizon-averaged MSE on 153 of 170 channels.

Our diagnostics identify a distinction between stable relation strength and the effective balance of predictive supervision. PEMS08 contains 143 channels with coherence below 0.5, with a mean of 0.077. FULL therefore places most of the supervision emphasis on target-future shape for these channels. Equal Weighting gives role-response shape an equal contribution and reduces their group-level mean MSE by 0.74%. Both single-source variants have higher group-level MSE than FULL, supporting the value of combining the two signals. Fixing the coefficient also removes distance differences caused by channel-specific coefficients, giving cross-channel positive selection a common supervision geometry. These findings support a dataset-specific mismatch between relation strength and supervision utility, explaining why equal weighting is effective on PEMS08 and identifying a limitation of coherence-based supervision weighting.

We have added this analysis to Appendix D.6 (RQ6) in the revised manuscript, with the new text highlighted in blue. This analysis will be retained in the final version.

Revision locations: Appendix D.6, RQ6 (p. 37, lines 1975-1984), following Table D.4 (p. 37, lines 1955-1967).

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

This paper proposes CoReG, a plug-and-play retrieval-augmented framework for multivariate time-series forecasting that explicitly separates historical similarity into two complementary “retrieval geometries.” TimeGeo retrieves candidates by multiscale temporal-shape similarity via an additive block-averaging decomposition, while RoleGeo retrieves by “system role,” using stable cross-variate relations and time-aligned role responses to construct a role-predictive key space. Retrieved futures from both paths are fused with a frozen backbone forecast via path-specific readouts and a simple symmetric merge. Across 10 datasets and 6 backbones, the authors report consistent error reductions, and provide ablations and resource trade-off analyses suggesting the two geometries are complementary and computationally practical.

Critical strengths
  1. The motivation for complementary retrieval geometries is communicated clearly with intuitive figures and a consistent terminology.

  2. The central idea of decomposing retrieval into two semantically distinct geometries is clear, well-motivated, and novel relative to the single-criteria retrieval common in prior work.

  3. Retrieval-augmented forecasting is an active area. Separating predictive analogies by semantics (shape vs role) is a timely and impactful conceptual contribution.

Critical weaknesses
  1. The RoleGeo encoder training is under-specified in the main text (e.g., contrastive loss form, negatives, data augmentation, architecture choices), making it hard to fully assess or reproduce without deep appendix reading.

  2. The exact mechanics of adjusted Top-K scoring, memory construction (window overlap, horizon alignment), and normalization are not all specified in the main body.

  3. Recent retrieval-augmented spatio-temporal frameworks that decouple spatial/temporal factors and use dual memories (e.g., RAST) are conceptually close; a direct discussion/comparison is missing.

  4. Frequency-domain retrieval (e.g., SpecReTF) explores alternative similarity geometries (amplitude/phase); a comparison or discussion would help position TimeGeo beyond time-domain pattern similarity.

Questions for authors

Please detail the RoleGeo encoder training in the main manuscript: exact contrastive loss, negative sampling strategy, architecture of the encoder, and any data augmentations used.

1 comment
AuthorsOpenReview ↗

Response to Reviewer xeXW

Thank you for your constructive review and recognition of the motivation and novelty of separating temporal-shape and system-role retrieval. All revisions are highlighted in blue, with figure and table revisions indicated by blue captions.

1. RoleGeo encoder training

RoleEncoder is a two-layer MLP with dimensions 720→64→16720\rightarrow64\rightarrow16, ReLU activation, and L2-normalized output keys. Training uses no additional data augmentation. Cross-variable information enters through the supervision signature, which combines the independently standardized target future fˉt,i\bar f_{t,i} and aligned role-response future eˉt,i\bar e_{t,i} using variable-specific coherence cic_i.

zt,i=concat(1−ci fˉt,i,ci eˉt,i).z_{t,i}=\mathrm{concat}\left(\sqrt{1-c_i}\,\bar f_{t,i},\sqrt{c_i}\,\bar e_{t,i}\right).

For minibatch sample aa with window start tat_a, valid candidates are J(a)={b:∣ta−tb∣≥L+H}\mathcal J(a)=\{b:|t_a-t_b|\ge L+H\}, where LL and HH are the history and forecast lengths. This excludes overlapping history-plus-future intervals. The positive is p(a)=arg⁡min⁡b∈J(a)∥za−zb∥2p(a)=\arg\min_{b\in\mathcal J(a)}\|z_a-z_b\|_2, and all remaining valid candidates serve as negatives. Candidates can come from different variables. With normalized history encoding qaq_a and temperature τ=0.05\tau=0.05, the predictive contrastive loss is

LPCL=−1∣I∣∑a∈Ilog⁡exp⁡(qa⊤qp(a)/τ)∑b∈J(a)exp⁡(qa⊤qb/τ).\mathcal L_{\mathrm{PCL}}=-\frac{1}{|\mathcal I|}\sum_{a\in\mathcal I}\log\frac{\exp(q_a^\top q_{p(a)}/\tau)}{\sum_{b\in\mathcal J(a)}\exp(q_a^\top q_b/\tau)}.

Here, I\mathcal I contains samples with at least one valid candidate. Batches with no valid samples are skipped. Role responses guide positive selection, allowing RoleEncoder to learn relation-informed keys from target histories. Predictive encoders are trained for 80 epochs with batch size 256 and learning rate 0.001.

Revision locations: Section 3.2 (p. 5, lines 223-227), Section 3.4 (p. 6, lines 288-304, Eqs. 16-18), and Appendix C.2 (p. 23, lines 1190-1192 and 1201-1211, Table C.2).

2. Retrieval protocol

Adjusted Top-K first selects legal candidates by query-key dot-product similarity ama_m. A confidence network takes the normalized current history and candidate future and produces gm∈(0,1)g_m\in(0,1). The adjusted score is rm=amgmr_m=a_mg_m, and candidate futures are aggregated with αm=exp⁡(rm)/∑j=1Kexp⁡(rj)\alpha_m=\exp(r_m)/\sum_{j=1}^{K}\exp(r_j). The selected candidates retain their original similarity ordering, and their adjusted scores form the GeoRead input.

Memory construction uses stride-1 windows, each comprising a length-LL history immediately followed by a length-HH future, entirely within the training split. Memories are built separately for each variable, geometry, and forecast horizon. Within each contiguous time stratum, the actual window closest to the encoding centroid is retained with its history key, normalized target future, and start time. Training retrieval masks overlapping complete intervals.

The dataset-level StandardScaler is fitted on the training split and reused for validation and testing. Each retrieval path independently standardizes its history and future windows. The retrieval output is restored using the current history's mean and standard deviation.

Revision locations: Section 3.2 (pp. 4-5, lines 199-227, Eqs. 2-5), Section 3.3 (p. 5, lines 254-258, Eq. 10), Section 3.4 (p. 6, lines 306-308), and Appendix C.2 (p. 23, lines 1201-1225, Table C.2).

3. RAST and SpecReTF

RAST [1] separates spatial and temporal representations, retrieves historical pattern vectors from two stores, and fuses them through cross-attention for traffic prediction. CoReG separates temporal shape and relation-supervised system role, reads target-variable future segments, and independently fuses each retrieval forecast with the shared frozen Anchor. We have added a direct discussion of the shared dual-retrieval structure and the distinct information represented by each path.

SpecReTF [2] measures STFT amplitude and phase similarity with recency-weighted frame scores. TimeGeo uses additive multiscale time-domain components and future-supervised encodings, while RoleGeo adds stable-relation supervision to define a second retrieval geometry. The added discussion distinguishes their similarity representations, supervision sources, and forecasting evidence.

Revision locations: Section 2.2 (p. 3, lines 149-154) and Table 1 (p. 3, lines 137-146). Appendix A.4 (p. 16, lines 815-823) and Table A.1 (p. 16, lines 825-839).

References

[1] Weilin Ruan, Xilin Dang, Ziyu Zhou, Sisuo Lyu, and Yuxuan Liang. A Retrieval Augmented Spatio-Temporal Framework for Traffic Prediction. Proceedings of the AAAI Conference on Artificial Intelligence, 40(46), 39163-39172, 2026. Publisher

[2] Huu Hiep Nguyen, Minh Hoang Nguyen, Dung Nguyen, and Hung Le. Spectral Retrieval-Augmented Time-Series Forecasting. arXiv:2606.19412, 2026. arXiv

Comments on the paper

1 comment

Dear reviewers,

The authors have submitted their response to your review comments. Please kindly take a moment to read their response and assess whether your concerns have been adequately addressed. Further discussion with the authors to reach a deeper mutual understanding would be greatly appreciated.

Thanks, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.