acceptodds
Under review as a conference paper at ICLR 2027

Dual-Context Analog Retrieval for Time-Series Forecasting

Abstract

Most long-term time-series forecasting models map the look-back window directly to the full horizon in a single pass. While efficient, this design does not explicitly determine which historical state is most relevant to each future segment or consider what followed that state. Analog forecasting can provide such evidence by retrieving a past state similar to the present and examining its observed continuation. However, relying on a single nearest match can be unreliable, and overlapping patches may yield highly similar or near-duplicate candidates. To address these limitations, we propose DuoTS, a Dual-Context Time Series forecasting model, which leverages retrieved evidence without depending on it exclusively. DuoTS first generates a base forecast using a fully parallel patch encoder and a linear prediction head, and then progressively refines this forecast one future patch at a time. Each refinement considers two complementary views. The first view, the current context, attends to recent tokens and captures the latest dynamics of the series. The second view, the detail context, provides distinct retrieved analogs together with the trajectories that followed them. Since the refinement is performed separately for each future patch, the model weighs the two views differently across the forecast horizon and associates each segment of the prediction with the evidence most appropriate to its temporal distance from the present. Extensive experiments on multiple real-world time-series datasets show that DuoTS achieves state-of-the-art performance, while component-wise ablation studies verify the contribution of each context. Moreover, the refinement mechanism is independent of the remaining DuoTS architecture. It requires only the encoded look-back window and the position of the future patch, allowing it to be applied as a model-agnostic module to improve existing forecasting models.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, retrieval-augmented forecasting, patch-based forecasting

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_2vFjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DuoTS adds a horizon-dependent correction to a patch forecaster, combining recent context with retrieved historical states and their observed successor changes. Experiments cover eight datasets, four horizons, and transfer to three backbones.

Critical strengths

The forecast–retrieve–revise design is sensible. Horizon-specific queries combine recent context with historical continuations, while the exclusion preference reduces duplicate matches from overlapping patches. Encoding successor changes provides a reasonable way to represent what followed a matched state.

The results are competitive across eight datasets and four horizons. The main experiments report three-run variability, and the transfer and perturbation studies provide additional evidence. In Appendix F, random candidate selection worsens 25 of 27 affected settings, supporting the usefulness of learned ranking.

Critical weaknesses

The model has many interacting components, but the gain from explicit analog-successor retrieval over a simpler corrector is not isolated. Table 3 masks contexts only at inference, and Table 11 compares outputs of the same jointly trained model. Neither tests a retrained corrector with comparable capacity that attends to the same history.

Appendix D describes 20–50 Optuna trials per dataset, look-back candidate, and horizon, using validation MSE. Comparable tuning budgets and window-selection rules are not documented for the main baselines. Since several gains are small, this uncertainty affects the strength of the comparative claims.

Questions for authors

Could you compare the full correction with one simpler attention-based corrector on representative datasets, keeping the backbone, model capacity, training budget, and tuning budget comparable? This would help identify the benefit of explicitly retrieving successor information.

Which main-table baselines were rerun, and what tuning budgets and look-back selection rules were used? If these differed substantially from DuoTS, a matched rerun of the strongest baseline would make the comparison more convincing.

1 comment
AuthorsOpenReview ↗

Simpler attention corrector and baseline rerun protocol

Thank you for your concrete reviews. For more detailed results, please refer to the revised pdf.

W1 / Q1. We retrained a generic attention corrector that reads the same history tokens with the same horizon query, without block retrieval, Top-K selection, diversity, confidence, or successors. It has the same parameter count as DuoTS and keeps its backbone, current context, router, and decoder. We also retrained a variant without successors (Table 4, Table 14).

DatasetNo successors ΔMSE (×10⁻³)Generic attention ΔMSE (×10⁻³)
Weather+1.28 (2/4)+1.72 (4/4)
ETTh1+11.85 (4/4)+21.09 (4/4)
ETTm1+4.97 (4/4)+5.72 (3/4)

Entries are alternative minus DuoTS, with the number of horizons where DuoTS is better. DuoTS is better than both at 15 of the 16 ETTh1 and ETTm1 comparisons, by up to 6.5% MSE, and is lowest or tied at every Weather horizon. We will add the remaining datasets to the paper as the runs complete.

W2 / Q2. All six baselines in Table 1, SRSNet (Wu et al., 2025), TimeKAN (Huang et al., 2025), Amplifier (Fei et al., 2025), iTransformer (Liu et al., 2024b), TimeMixer (Wang et al., 2024), and PatchTST (Nie et al., 2023), are now rerun through TFB (Qiu et al., 2024) under the DuoTS protocol. The look-back is searched over {96, 336, 512} jointly with each baseline's hyperparameters in 50 Optuna trials (Akiba et al., 2019) (Appendix D.2). The results are essentially unchanged. DuoTS still has the lowest MSE in 22 and the lowest MAE in 24 of 32 settings (22 and 23 in the submission). We also rerun RAFT (Han et al., 2025), the closest retrieval-based method, and DuoTS is more accurate in three of four cells (Appendix E, Table 11). We will add the remaining datasets to the paper as the runs complete.

Reviewer_1HfrOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors focus on the limitation that conventional one-shot forecasting models do not explicitly determine which historical states are most relevant to different future segments or exploit the trajectories that followed those states. To address this issue, the paper proposes DuoTS, which first generates a base forecast and then refines it patch by patch using two complementary contexts. The current context captures recent dynamics near the forecasting boundary, while the detail context retrieves similar historical patches together with their observed successors. The retrieved evidence is combined with the base forecast through horizon-dependent view weights and correction gates. The reported results are generally strong: DuoTS achieves the best MSE in 22 and the best MAE in 23 of 32 dataset–horizon settings, and the correction module improves 15 of 24 settings when transferred to other backbones. Nevertheless, I have several concerns regarding the fairness of the experimental comparison, the actual contribution of the retrieval mechanism, and the claimed model generality.

Critical strengths
  1. The motivation of retrieving historical states together with their observed continuations is intuitive and practically meaningful. 2) The correction-based design avoids relying entirely on retrieved analogs and preserves a direct base forecast. 3) The main table reports results for every dataset and prediction horizon, together with standard deviations over three runs. 4) The experiments include complete high-dimensional Electricity and Traffic datasets rather than reduced channel subsets. 5) The paper provides extensive ablations, retrieval perturbations, transfer experiments, and computational-cost analysis. 6) The method achieves strong overall performance, particularly on ETTh2, Weather, Traffic, and Solar. 7) The implementation and hyperparameter configurations are described in considerable detail.
Critical weaknesses
  1. My main concern is that the research topic has already been extensively studied, while the proposed method does not demonstrate sufficiently strong methodological novelty. Most of its components are based on existing techniques, and the contribution mainly lies in their integration.
  2. The experimental comparison may not be fully fair. DuoTS performs a separate Optuna study for every dataset, look-back candidate, and prediction horizon, with up to 50 trials per study, whereas it is unclear whether the competing methods receive a comparable tuning budget.
  3. The look-back length and a large number of architectural and retrieval hyperparameters are selected separately for each dataset–horizon setting. Such extensive per-setting optimization may overfit the validation sets and limits the practicality of the method.
  4. The claim that the correction module is model-agnostic is somewhat overstated. It requires patch embeddings and joint training with the forecasting backbone, and it improves only 15 of the 24 reported transfer settings.
  5. The context-component study removes the current or detail context only at inference time. This introduces a distribution mismatch and is less convincing than retraining the corresponding ablated models.
  6. The paper does not directly demonstrate that the retrieved latent patches correspond to meaningful or interpretable historical analogs. More qualitative cases showing the retrieved states and their continuations would strengthen the mechanism claim.
  7. DuoTS contains many interacting modules and hyperparameters, including two backbone paths, sparse channel mixing, temporal differences, hierarchical retrieval, confidence calibration, view routing, and correction gates. The resulting complexity makes it difficult to identify the essential source of the performance gain.
  8. The computational overhead is substantial in some settings. On ETTh2, DuoTS uses more parameters, FLOPs, memory, and inference time than all of the reported comparison methods.
Questions for authors

as stated in the weakness

3 comments
AuthorsOpenReview ↗

Novelty, re-tuning, retrained ablations, and corrected cost table

Thank you for the detailed reviews. For more detailed results, please refer to the revised pdf.

W1. DuoTS differs from prior retrieval methods in five ways. It retrieves from the current look-back window instead of the training set or a global memory as in RAFT (Han et al., 2025) and TS-RAG (Ning et al., 2025), each future patch issues its own query, each analog is paired with its successor changes, coarse-to-fine retrieval avoids near-duplicate matches, and the evidence enters as a gated correction that keeps the direct forecast. Table 4 and Appendix F show that these choices matter. Removing the successors or replacing retrieval with a parameter-matched attention reader makes DuoTS worse at 15 of 16 ETTh1 and ETTm1 comparisons.

W2. All six baselines are now re-tuned through TFB (Qiu et al., 2024) under the DuoTS protocol, with 50 Optuna trials (Akiba et al., 2019) per study and the look-back searched over {96, 336, 512} (Appendix D.2). The results are mostly unchanged. DuoTS still has the lowest MSE in 22 and the lowest MAE in 24 of 32 settings (22 and 23 in the submission, Table 1).

W3. Following SRSNet (Wu et al., 2025), the look-back is selected per setting, and the same per-setting search is now applied to every method in Table 1. The gain is consistent across horizons, since DuoTS is best at every horizon of ETTh2, Traffic, and Solar on both metrics (Table 1).

W4. We replaced "model-agnostic" by "plug-in module for patch-based forecasters" throughout. With the re-tuned baselines, the module improves 14 of 24 transfer settings for SRSNet (Wu et al., 2025), TimeKAN (Huang et al., 2025), and PatchTST (Nie et al., 2023), including all six on ETTh2 (Section 4.3, Table 5). We will add the remaining datasets to the paper as the runs complete.

W5. We added two measurements that do not remove a context at inference. Retraining without successors or with a parameter-matched generic attention reader makes the model worse at 15 of 16 ETTh1 and ETTm1 comparisons (Table 4), and we will add the remaining datasets as the runs complete. Compared with the base forecast of the same model, the correction lowers MSE at every horizon of Weather, Solar, Electricity, Traffic, and ETTm2 (Appendix G.1, Table 15). Context removal now also covers all eight datasets (Appendix G.2, Table 16).

W6. We evaluate the retrieved analogs in the raw series (Appendix G.4). Judged by how well its continuation matches the true future, the retrieved match ranks better than random in 29 of 30 cells, although retrieval is never trained for this. Selected lags lie near a period multiple about as often as chance (0.22 vs 0.19), so retrieval is not a seasonal lookup and follows the principle of analog forecasting (Lorenz, 1969).

W7. Table 2 shows that each backbone module contributes. The backbone alone already outperforms every baseline on ETTh2, and the correction lowers MSE by up to 14.6% on Weather (Appendix G.1). The perturbations and the retrained alternatives attribute the correction gain to the retrieved analogs and their successors (Appendix F, Table 4).

W8. We have fixed the original ETTh2 row, which was profiled with the ETTh1 configuration. With its own ETTh2 configuration, DuoTS has 0.27 M parameters, needs fewer FLOPs than every model except DLinear (Zeng et al., 2023), and is about seven times faster in inference than PatchTST and TimeKAN (Appendix H, Table 18).

Dear Authors,

Please accept my apologies for the delay in submitting my review. Thank you for addressing my comments and revising the manuscript. I have carefully read the authors' responses and reviewed the revised manuscript. Overall, I am satisfied with the revision, and the authors have addressed most of my concerns. I have two remaining suggestions.

  1. A good overview figure should allow readers to understand the authors’ main idea from the figure and its caption alone. Figures 2 and 3 clearly fall short of this standard.

  2. In reviewing this work, I also looked through some of the related literature on retrieval. I found the name of the proposed method quite reminiscent of KUP-BI. Is there a connection between “Dual” and the idea of “bidirectional” retrieval? The manuscript does not discuss this possible connection.

Thank you for considering these comments. I wish you all the best with the revision.

Best regards, The Reviewer

Dear Reviewer 1Hfr,

Thank you for carefully reviewing our manuscript and for raising your score. We really appreciate your constructive feedback, which has helped us substantially improve the paper, particularly through the re-tuned baselines, retrained ablations, corrected cost analysis, and clearer description of the plug-in module.

We agree that Figures 2 and 3 should be understandable independently of the main text. We have redesigned both figures and revised their captions accordingly. Figure 2 now clearly illustrates the backbone paths, their fusion into patch tokens, and the plug-in correction module. Figure 3 details the horizon-specific correction process, including the current and detail contexts, coarse-to-fine retrieval, observed successors, view weights, and gated correction. We have also added a legend and clarified how the correction varies across future patches.

Thank you also for pointing us to KUP-BI (Chong et al., ICML 2026). The name DuoTS refers to its two complementary contexts, current and detail, rather than bidirectional inference. While both methods use historical continuations, their retrieval mechanisms differ. KUP-BI retrieves from training-set trajectories using Pearson correlation and integrates the retrieved evidence into backbone features. In contrast, DuoTS retrieves analogs and their observed successors from the current look-back window, uses horizon-specific learned queries and coarse-to-fine retrieval, and applies the evidence as a gated correction to the base forecast. We have clarified the naming in the Introduction and added KUP-BI and this distinction to Related Work.

Thank you again for your engagement and kind wishes.

Best regards,
The Authors

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

This paper proposes DuoTS, a multivariate time-series forecasting model combining a direct patch-based forecaster with a horizon-conditioned correction module. For each future patch, the correction combines attention over recent historical tokens with retrieved historical representations and their observed successor changes. Retrieval is restricted to the current input window and uses coarse-to-fine selection, a temporal diversity constraint, and confidence weighting. The backbone additionally incorporates sparse channel mixing and explicit first- and second-order patch differences. Across eight benchmarks and four horizons, DuoTS reports the best MSE in 22 of 32 settings and the best MAE in 23, including ties. Jointly training the correction with three alternative backbones improves MSE in 15 of 24 settings.

Critical strengths
  1. DuoTS integrates recent dynamics and retrieved historical continuations through a horizon-conditioned residual correction. This is a reasonable way to exploit recurring patterns while retaining a direct forecasting path. The correction can also be integrated with different backbones under joint training, giving the design potential usefulness.

  2. The evaluation covers eight datasets and four forecast horizons, with three-run statistics for the main DuoTS results. The paper goes beyond aggregate accuracy by reporting context removal, retrieval perturbations, backbone transfer, and computational costs. It acknowledges nonuniform transfer, the limited effect of diversity, and the diagnostic nature of the perturbation experiments. This transparency helps establish both the method's practical promise and its limitations.

Critical weaknesses
  1. The evidence does not sufficiently isolate the contribution of the proposed analog-continuation mechanism. The full model introduces both a new backbone and a correction branch. Table 3 and Table 11 disable components or compare outputs of an already jointly trained model; they therefore do not establish the advantage over independently trained, comparably tuned simpler alternatives. The evaluation findings support some dependence on retrieval, but leave unclear how much of the gain specifically requires aligned successor evidence and the proposed retrieval machinery rather than a generic additional prediction branch. Since retrieval of similar histories and their continuations has established precedents, this distinction is central to the significance of the contribution.

  2. The experimental protocol leaves uncertainty about the robustness of the relatively modest accuracy advantage. Appendix D describes separate hyperparameter studies for each dataset, horizon, and candidate look-back length, usually with 50 trials per study. However, the corresponding tuning budgets and rerun protocols for the Table 1 baselines are not comparably clear. Many improvements over SRSNet are small on average across datasets, and the manuscript does not report baseline variability, making this uncertainty consequential. The RAFT/SARAF comparison is explicitly cross-paper and uses different look-back lengths, so it does not resolve the issue for the closest retrieval-based alternatives.

Questions for authors
  1. Can the retrieval and successor perturbations in Appendix F be evaluated across multiple seeds, with the variability of the paired changes reported? In particular, does correctly aligned successor information provide a reproducible benefit, and is there a setting in which the diversity constraint consistently helps?

  2. Equation 13 mixes information across the entire historical patch axis before retrieval. Consequently, a retrieved token and its “successor” can both encode global window information. How should we interpret the analog-continuation interpretation under this mixing, and what evidence distinguishes the proposed mechanism from selectively reading contextualized representations globally?

1 comment
AuthorsOpenReview ↗

Retrained alternative correctors, re-tuned baselines, and successor evidence

We appreciate the thoughtful reviews. For more detailed results, please refer to the revised pdf.

W1. We retrained two alternatives with the DuoTS backbone, a parameter-matched generic attention corrector and a variant without successors (Table 4). DuoTS is better than both at 15 of the 16 ETTh1 and ETTm1 comparisons, by up to 6.5% MSE, and is lowest or tied at every Weather horizon. We will add the remaining datasets to the paper as the runs complete. In addition, replacing the retrieved candidates or their successors raises MSE in 23 to 26 of the 32 trained models (Appendix F), and the correction lowers MSE at every horizon of Weather, Solar, Electricity, Traffic, and ETTm2 (Appendix G.1).

W2. All six baselines are now rerun through TFB (Qiu et al., 2024) under the DuoTS protocol, with 50 Optuna trials (Akiba et al., 2019) per study and the look-back searched over {96, 336, 512} (Appendix D.2). The results are essentially unchanged. DuoTS still has the lowest MSE in 22 and the lowest MAE in 24 of 32 settings (22 and 23 in the submission). It lowers the average MSE by 9.1% on ETTh2, 8.2% on Solar, and 5.4% on Traffic (Table 1). We also re-tuned RAFT (Han et al., 2025) with its official implementation, and DuoTS is more accurate in three of four cells, with 11.2% lower MSE on ETTh2 at H=720 (Appendix E, Table 11). We will add the remaining datasets to the paper as the runs complete.

Q1. We apply each perturbation to all 32 independently trained models (Tables 12 and 13). Zeroing the successor changes raises MSE in 23 of 32 cells, and reading the patches before each match instead of after it raises MSE in 26. The retrained successor-free variant is also worse at all eight ETTh1 and ETTm1 horizons (Table 4). The diversity radius is active in 7 of the 32 cells. There it changes 32% of the selections and spreads the selected patches from 4.47 to 6.09 positions apart at no measurable accuracy cost (Appendix F).

Q2. The mixed representation enters the retrieval tokens only through the learned fusion weight of Eq. 14. On ETTh1, ETTh2, and ETTm2 the tokens stay close to their local patch encodings, with a relative distance of 0.05 (Appendix G.3, Table 17). Across all datasets, reading the patches before each match instead of after it raises MSE in 26 of 32 cells (Table 13). A window-wide summary would change little under this reversal. The retrained generic attention corrector, which reads all contextualized tokens, is worse than DuoTS at 11 of 12 horizons (Table 4). Finally, the retrieved matches rank better than random against the true future in 29 of 30 cells (Appendix G.4).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.