acceptodds
Under review as a conference paper at ICLR 2027

Most-Recent Anchoring with Recurrent Ordering for Time-Series Forecasting

Abstract

Long-term forecasting models commonly process all patches in a look-back window using the same fixed stack. Older contextual patches and recent evidence therefore receive the same computational depth. Yet the information closest to the forecast and the more distant context do not contribute equally. Uniform processing leaves this distinction unexpressed in the architecture. We propose MARO, a ost-Recent nchoring with ecurrent rdering model that processes the lookback window from the most recent patch to the oldest. The most recent patch serves as the anchor. It initializes the latent state and conditions each subsequent step, so older patches are folded into a representation that remains centered on recent evidence. A single shared module is reused at every step, so extending the scan further into the past introduces no additional parameters. Intermediate states retained during the scan allow the forecast head to weigh short and long portions of the history separately. This expresses recency through the order of recurrent refinement. Extensive experiments across multiple real-world time series datasets show that MARO achieves state-of-the-art performance on both long-term and short-term forecasting tasks. Ablation studies examine the contribution of the main architectural components.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, Recurrent neural networks, patch-based forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vxpzOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a time-series forecasting model that processes historical data patches in reverse order. The model initializes its latent state using the most recent patch, treating it as a persistent anchor, while integrating earlier contextual information through a shared recurrent module. The model maintains separate context and prediction states, fuses intermediate states from varying look-back depths, and performs additional refinement prior to making a prediction. A weight-sharing mechanism ensures that the number of recurrent parameters remains independent of the look-back length. Experimental results demonstrate that the model achieves competitive performance.

Critical strengths
  1. The authors have released the source code, facilitating reproducibility and further research.
  2. The paper is clearly written and well structured, making the motivation, methodology, and experimental results easy to follow.
Critical weaknesses
  1. The introduction primarily relies on experimental results to support the near-to-far scanning order, but it fails to explain in depth why a uniform computational depth limits the utilization of historical information, and why reverse scanning alleviates this limitation.
  2. The methodological novelty is limited, as a backward RNN or a reverse-order variant of SegRNN can also achieve asymmetric information aggregation.
  3. The update rules in Eqs. (3), (5), and (9) rely heavily on concatenation followed by MLPs, but the rationale for the specific input combinations is unclear.
  4. The comparison is unfair, as models use different look-back lengths for prediction. Please evaluate all models at the same look-back lengths.
  5. The introduction shows an incomplete result of the benefits of reverse scanning. Figure 2 shows that recent-to-past scanning performs better, but Table 13 in the appendix shows that old-to-recent scanning performs better in some cases.
Questions for authors

The relative benefits of recent-to-old and old-to-recent scanning may depend on history length. How does MARO’s forecasting performance vary with different look-back lengths?

4 comments
AuthorsOpenReview ↗

Recency as an inductive bias, backward SegRNN, and fixed look-back comparison

We appreciate the detailed comments. Please see the revised pdf for details.

W1. The reason is an inductive bias. The near future of a series depends most on its latest state, which is why exponential smoothing (Hyndman et al., 2008), NLinear (Zeng et al., 2023), and DeepAR (Salinas et al., 2020) weight recent observations more. A fixed-depth mixer treats every patch alike, so it must learn this from data, and longer history can dilute the recent evidence. Appendix G shows this directly. On ETTh1 at H=96 and 192, extending the look-back from 96 to 512 lowers MARO's MSE by 0.0371 on average, against 0.0099 for TimeMixer, 0.0076 for PatchTST, and 0.0066 for iTransformer (Table 22). At H=96, iTransformer and TimeMixer even get worse, while MARO is best of the four at L=512 on both horizons. MARO avoids this dilution because the most recent patch is carried through every update and older patches add increments with non-increasing weights (Proposition 1, Appendix E.2). The recent anchor accounts for most of the gain (Table 18).

W2. A backward RNN only starts from the most recent patch. MARO also feeds this anchor into every update and fuses log-spaced checkpoints, which yields the recency prior of Proposition 1. A parameter-matched GRU in place of the MARO update raises MSE in nine of twelve settings (Table 19). We also reversed SegRNN (Lin et al., 2023). Backward SegRNN lowers the average MSE of SegRNN on ETTh2 and Weather with both L=96 and L=512 (Table 21), so recent-to-old order helps beyond MARO. MARO still has lower or equal MSE than Backward SegRNN at every horizon, and its average MSE is lower than SegRNN, xLSTMTime (Alharthi & Mahmood, 2024), and xLSTM-Mixer (Kraus et al., 2024) on all eight datasets (Table 12).

W3. Eq. 3 uses the matching features [h ∥ U ∥ U ⊙ h] of Mou et al. (2016) and Conneau et al. (2017) to compare the new patch with the accumulated context. In Eq. 5, h gives the temporal summary and ρ_s what the new patch adds, while the fixed anchor and future calendar keep recent evidence available at every depth. Eq. 9 reuses this update with ρ set to zero, since no new patch is consumed (Section 3.4). Table 3 and Table 19 support these choices.

W4. All models select the look-back from {96, 336, 512} with the same 50 Optuna trials (Akiba et al., 2019), since a single fixed length would favour some models (Abdelmalak et al., 2026). As requested, we also retuned all seven models with the look-back fixed to 512 on all eight datasets (Appendix H). MARO gives the lowest MSE in 25 of 32 settings and the lowest MAE in 22, while no baseline leads in more than four (Tables 23 and 24). On ETTh1, MARO is also the best of four models at L=512 (Table 22).

W5. Appendix F.2 now reports the controls on all five datasets (Table 20), and the introduction points to it. With causal attention, recent-to-old gives lower or equal MSE on all five. A plain RNN already keeps its last input closest to the final state, so its results are mixed. MARO instead favours the patches it scans first (W1). Comparing the two complete traversals, which differ only in anchor and direction, recent-to-old wins nine of ten settings (Table 17), and the order effect is non-negative in every factorial setting (Table 18).

Q. The benefit of recent-to-old scanning holds for short and long history. Backward SegRNN lowers the average MSE of SegRNN on ETTh2 and Weather by 1.2% and 0.5% with L=96 and by 1.3% and 1.4% with L=512 (Table 21). With L=512, MARO's recent-to-old traversal also wins nine of ten settings (Table 17). The same MARO comparison at L=96 and 336 is running, and we will add it as soon as it completes. MARO's average MSE decreases from L=96 to 336 to 512 on all seven datasets in Figure 6, by 9.4 to 19.7% (Appendix C.3).

We will add the remaining datasets of all partial studies to the paper as the runs complete.

Reviewer_vxpzOpenReview ↗

Response to rebuttal

Thank you for the response. However, my concerns have not been addressed. In particular, the explanation remains insufficient to establish why recent-to-old processing improves forecasting performance. I am also unclear about why the dataset examples in Figure 2 were changed and why the revised results differ from those reported in Table 13 of the original manuscript, including a reversal in the relative performance of the two scanning directions. Additionally, the full-model results in the ablation study have improved while the ablated variants remain unchanged. These unresolved issues limit my confidence in the experimental conclusions. I therefore maintain my original score.

AuthorsOpenReview ↗

Clarification of Experimental Results and Additional Analysis of the Recency Prior

Thank you for your follow-up. We understand your decision to maintain the original score. Nevertheless, we would like to provide additional experimental evidence addressing your concerns. All revisions are highlighted in the updated PDF.

1. Correction of Figure 2.

We acknowledge the inconsistency between Figure 2 and the original appendix results. During manuscript revision, we inadvertently replaced Figure 2 with results from different controls while retaining the original appendix comparisons. This was our mistake, and we sincerely apologize for the confusion.

We have restored Figure 2 to the original ETTh1, ETTm1, and Weather results, consistent with Table 13 of the original submission (now Table 20). The figure and appendix are now consistent.

The Introduction also clarifies that causal attention gives lower or equal MSE with recent-to-old scanning on all five datasets, while the plain RNN favors old-to-recent on ETTh2 and ETTm2. We no longer claim that scan reversal universally improves performance.

2. Correction of the ablation results in Table 3.

We also acknowledge the inconsistency in Table 3. During manuscript revision, the full-model row was mistakenly changed while the ablated variants remained unchanged.

We have restored the original full-model values corresponding to the reference runs used for the ablation study. The caption now clarifies this.

The corrected full model still remains best or tied best in every column. Thus, the conclusions of Section 4.2 remain unchanged.

3. Additional experimental evidence explaining why recent-to-old scanning helps.

We agree that the original explanation did not sufficiently establish why recent-to-old processing improves forecasting. Thus, we have added Appendix I, which examines the recency effect through four complementary analyses.

(i) Forecast-horizon sensitivity (Table 25). Recent-to-old scanning yields its largest improvement in the first 10% of the forecast horizon in 14 of 16 settings. These steps depend most directly on the latest state of the series. The advantage generally decreases at longer horizons, suggesting that the recency prior is particularly useful for near-future predictions.

(ii) Anchor versus scan order (Table 26). Most of the improvement comes from the patch used to initialize the recurrence. With the recent anchor fixed, reversing the remaining patches changes MSE by at most 0.7%. This separates the contribution of anchoring from scan direction.

(iii) Input-sensitivity analysis (Tables 28 and 29). At L=512, recent-to-old MARO assigns 0.33 to 0.54 of its input-gradient sensitivity to the newest 96 steps, compared with 0.19 to 0.22 for old-to-recent MARO and 0.25 to 0.33 for PatchTST. Integrated gradients confirm the same ordering across all settings. This directly demonstrates greater sensitivity to recent observations.

(iv) Controlled synthetic experiments (Table 30). We constructed synthetic series with controlled predictive-signal locations. When the signal is concentrated in the newest 96 steps, recent-to-old scanning improves MSE by approximately 4.0%. When it lies at the oldest lags, the advantage reverses and old-to-recent performs better.

These findings clarify both the benefits and limitations of recent-to-old processing. The scan introduces a recency prior that helps when predictive information is concentrated near the forecast boundary, rather than universally outperforming chronological processing. Table 25 suggests that this condition is common in the evaluated benchmarks.

We also clarified Proposition 1. Its non-increasing coefficients describe realized recurrent updates under checkpoint fusion, not the end-to-end influence of raw input observations. The latter is examined empirically in Appendix I.3.

We hope these corrections and additional analyses clarify the experimental evidence and MARO's recency prior. We appreciate your careful review and the opportunity to address these concerns.

Best regards,
Authors

Dear Reviewer vxpz,

We would like to briefly follow up on your remaining concerns.

To make the reported numbers easier to trace, Appendix L (Table 35) now lists the runs and random seeds behind each results table. We also reran the configurations of Table 10 with three new seeds (Table 32), and the horizon-averaged MSE changes by at most 0.003 on all eight datasets.

Regarding the role of the scan direction, both the recent anchor and the recent-to-old scan introduce the same recency prior. Appendix I shows that this prior concentrates the model's input sensitivity near the forecast boundary, gives the largest gains on the first horizon steps in most settings, and helps on synthetic series when the predictive information lies near the forecast boundary.

Within MARO, the two contributions can partly overlap, and the anchor accounts for most of the gain (Appendix I.2). The revised text therefore presents the recent anchor as the central component of MARO. Appendix K further adds the anchor and the recent-to-old scan step by step to RNN and causal-attention updates, and both steps lower the average MSE. This shows that the recent-to-old scan adds to the benefit of the anchor, so combining the two is useful.

We hope this clarifies the open points in our discussion. Thank you again for your careful review.

Best regards,
Authors

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

The authors introduce a framework that avoids the obfuscation of recent patches by anchoring, or basically inducing an ordering into the method. The method scans the patches from the most recent to the oldest with one shared recurrent module, so the recurrent parameters do not grow with the look-back length, and it fuses intermediate states from different depths for the forecast. They test it on general TSLib-style forecasting benchmarks plus PEMS for supervised methods, and the method beats the majority of the selected baselines in long- and short-term forecasting. Besides the main benchmarks, the authors provide ablations, a comparison of scan directions, and an efficiency study.

Critical strengths

1. Efficiency on high-dimensional data. Because one shared recurrent module is reused for every patch, the recurrent parameters do not grow with the look-back length, and the cost study supports this: on Electricity. It also has the best MSE at every horizon on Traffic and Electricity. The authors are open about where it is slower (ETTh1, few channels). This trade-off is the inverse of xLSTM-Mixer (Kraus et al., 2025): xLSTM-Mixer runs its recurrence over the variates, so its sequential cost grows with the number of channels, whereas MARO recurs over time patches and processes the channels in parallel. MARO therefore covers the high-dimensional case, where a variate-wise recurrence is most expensive.

2. Clear structure and transparency. I like how the related work is structured, and the contributions are laid out in a clear way. The paper is also transparent: it gives the full selected configurations and three-seed standard deviations, reports negative results (e.g., a learned channel graph does not consistently beat fixed ones, Table 14), and scopes Proposition 1 carefully instead of overclaiming.

Critical weaknesses

1. Robustness of the hyperparameter setup. I have the feeling that the results are not really robust. Looking at the parameters selected for the different runs (Tables 8–10), they are very diverse and change drastically between runs: every dataset and horizon length has a completely different selection of parameters (e.g., patch length, token width, refinement steps, and varying channel-neighbourhood sizes and gate initializations). I would definitely like to see a more unified prediction setup, where at most the look-back length, batch size, or learning rate is changed. Without it, it is hard to tell how much of the improvement comes from the architecture and how much from the per-setting search and how viable it is on another dataset.

2. Missing recurrent baselines and limited evaluation of the ordering. I think it is fine to only compare against a few methods. However, showcasing a recurrent method without comparing against any recurrent method is a bit odd to me. The paper even mentions SegRNN, xLSTMTime, and xLSTM-Mixer, but I don't understand why those are not included in the evaluation. In particular, I don't see why the anchoring cannot be a generic framework. Figure 2 and Appendix D.1 take a first step with a plain RNN and causal attention, but the paper would be much stronger if the authors evaluated their order anchoring in a more general fashion.

Questions for authors

Clarification requests

  1. How do the results change with one unified configuration across all datasets and horizons, where at most the look-back length, batch size, or learning rate is tuned?
  2. Why are SegRNN, xLSTMTime, and xLSTM-Mixer, which the paper discusses, not included as baselines?
  3. Can the recent-to-old anchoring be applied as a generic component to other forecasting architectures?
  4. The individual ingredients of MARO are established: reversing the input order so that the most relevant inputs end up close to the final state (Sutskever et al., 2014), anchoring the forecast on the last observation (NLinear in Zeng et al., 2023; DeepAR, Salinas et al., 2020), and recency-weighted aggregation of past states (e.g., exponential smoothing). Their combination seems new to me, but the paper does not discuss this prior work. How does MARO relate to it, and how much of the gain comes from the scan order itself rather than from anchoring on the newest patch.

Also if a unified configuration retains most of the gains and the recurrent baselines are included, I would raise my score.

Suggestions for future work and presentation

  • The paper has its niche audience, but what would really increase the impact is showing that the same holds for foundation models as well, especially with the current rise of new methods.
  • The related work reads very dated; really recent methods are excluded.
  • Figure 5: the legend is not centered, TimeKAN is not visible, and the colors do not align.
12 comments
AuthorsOpenReview ↗

Unified configuration, recurrent baselines, and anchor vs. order

Thank you for the detailed reviews. All references below follow the revised PDF.

W1 / Q1. Following TimeMixer (Wang et al., 2024), Table 13 gives MARO and nine baselines one unified architecture configuration on ETTh1 and tunes only the look-back length (96, 336, 512), batch size, and learning rate. MARO gives the lowest MSE at H=192 and H=720, more horizons than any other model. We will add the remaining datasets to the paper as the runs complete. In the main comparison, every model including MARO receives the same budget of 50 Optuna trials (Akiba et al., 2019) per setting (Appendix B.1).

W2 / Q2. SegRNN (Lin et al., 2023), xLSTMTime (Alharthi & Mahmood, 2024), and xLSTM-Mixer (Kraus et al., 2024) are now included on all eight long-term datasets (Table 1, Table 12). MARO gives the lowest MSE in 30 of 32 settings and the lowest MAE in 31, and its average MSE is lower than that of the strongest recurrent model on every dataset, by up to 17.8% on ETTh2.

Q3. Figure 2 applies recent-to-old scanning to causal attention and to a plain RNN, and it lowers MSE in all four comparisons. The benefit of the scan order carries over across update rules.

Q4. Section 2 discusses input reversal (Sutskever et al., 2014), last-value anchoring in NLinear (Zeng et al., 2023) and DeepAR (Salinas et al., 2020), and exponential smoothing (Hyndman et al., 2008). Proposition 1 shows that checkpoint fusion gives older increments non-increasing coefficients, as in exponential smoothing, but on learned increments. Even with nearly uniform learned checkpoint weights, the most recent increment receives 3.26 times the coefficient of the oldest (Appendix E.2). A factorial study separates anchor and order (Table 18). The recent anchor gives most of the gain (0.0153 MSE on average), and recent-to-old order adds a smaller gain (0.0020) that is non-negative in every setting. We will add the remaining datasets to the paper as the runs complete.

In addition, We revised Figure 5 so that the legend is centered, TimeKAN is visible, and the colors are consistent across the three panels. Section 2 covers recent forecasters such as TimeFilter (Hu et al., 2025), WPMixer (Murad et al., 2025), and S-Mamba (Wang et al., 2025) as well as time-series foundation models such as TimesFM (Das et al., 2024) and Chronos (Ansari et al., 2024). Recent-to-old anchoring for foundation models is listed as future work in Section 5.

Reviewer_bWGBOpenReview ↗

Thank You

Thank you for the response,

Before engaging, it would really be helpful if you would color code the revisions you made per reviewer in the manuscript. Unfortunately, its not feasible in the rebuttal period to read all of the paper again to filter only for the changed passages.

Best bWGB

AuthorsOpenReview ↗

Update the paper with the highlights

Thank you for your feedback. We have updated the revised manuscript to highlight the modifications and additions.

Additional experiments are still ongoing, and we will keep you updated as soon as further results become available.

Thank you.

Best regards

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

The paper proposes MARO, a recurrent-depth, patch-based forecasting architecture that explicitly encodes recency by scanning the look-back window from the most recent patch to the oldest. The model anchors the latent state at the most recent patch, progressively integrates older patches via a shared recurrent update block, fuses exponentially spaced checkpoints to capture multiple effective depths, and applies post-scan latent refinement using the same shared update. Across standard long-term (ETT family, Weather, Electricity, Traffic, Solar) and short-term (PEMS) benchmarks, MARO achieves strong performance, with consistent gains over competitive baselines and ablations that identify the contributions of anchoring, progressive integration, and checkpoint fusion.

Critical strengths
  1. Architecture and algorithmic details are carefully described, with diagrams and pseudocode for the full forward pass.

  2. Broad evaluation on widely used long-term and short-term datasets; consistent improvements on Electricity and Traffic and competitive or best results elsewhere, including many short PEMS horizons.

Critical weaknesses
  1. The learned sparse channel graph is static (input-agnostic) and identity-based; this may limit adaptivity to time-varying inter-variable relationships.

  2. Many baseline numbers are quoted from prior work. Although common in TSF, This may introduce tuning and implementation discrepancies.

Questions for authors

The sparse channel graph is identity-conditioned and static across samples. Did you evaluate input-adaptive routing (e.g., light attention or conditioning scores on current window statistics)? How does it compare in accuracy and cost?

1 comment
AuthorsOpenReview ↗

Input-adaptive channel routing and rerun baselines

We appreciate the positive assessment. For more detailed results, please refer to the revised pdf.

W1 / Q. We evaluated an input-adaptive neighbourhood that conditions the channel scores on the mean and standard deviation of each window and recomputes the Top-κ neighbours for every sample (Appendix D.2). It gives higher MSE than the learned static graph in five of six settings and needs more parameters and memory in every setting, since it forms a per-sample C × C score tensor (Tables 14 and 15). Therefore, we keep the static graph. We will add the remaining datasets to the paper as the runs complete.

W2 All long-term baselines are now rerun through TFB with the same look-back candidates (96, 336, 512) and the same budget of 50 Optuna trials per setting as MARO (Appendix B.1, Appendix C). We also added the recurrent forecasters SegRNN, xLSTMTime, and xLSTM-Mixer on all eight datasets. MARO gives the best MSE and MAE on all five dataset groups in Table 1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.