acceptodds
Under review as a conference paper at ICLR 2027

GNA: Granular Neighbor Assembly for Retrieval-Augmented Multivariate Time-Series Forecasting

Abstract

Deep forecasters predict from a fixed-length lookback window, and lengthening it gives diminishing returns at a growing cost. Retrieval augmentation instead shows the model how similar past situations continued. Existing retrieval forecasters treat a retrieved past window as one unit, so every variate receives the continuation of the same past moment. In multivariate series, however, the best past match differs from variate to variate. We present GNA (Granular Neighbor Assembly), a retrieval layer for forecasting backbones that assembles neighbors at two granularities: whole past windows, which keep the variates coherent, and per-variate neighbors, in which each variate takes its future from its own best-matching past. A learned gate decides, per forecast step and variate, how much to trust these futures against a persistence forecast, next to the backbone's own forecast. Candidates come from an embedding trained to predict each window's future, and retrieval is strictly causal: a past window is used only once its future has been observed. With the same lookback for every model and one retrieval configuration for all datasets, GNA improves two Transformer backbones in 85 of 96 dataset–horizon settings, gives the lowest MSE on 8 of 12 standard benchmarks and beats its backbone in every seed on 10 of them. Both granularities are needed, and neighbors of mismatched queries are worse than none. Retrieval helps most where the lookback says least: the gate shifts trust to retrieved futures further ahead, and on monitoring data of a particle-physics experiment the gain concentrates at the start of data-taking runs. Where it fails, on hourly non-stationary series at long horizons, we trace the loss to the drifting level of the retrieved futures.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series forecasting, retrieval augmentation, rag

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_z6ZLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Granular Neighbor Assembly (GNA) for retrieval-augmented multivariate time-series forecasting. Motivated by the limitations of the fixed-length lookback windows and different neighbors by different variables, the proposed GNA combines global retrieval with per-variate retrieval. It further introduces a future-aware retrieval embedding trained using forecasting supervision and a learned gating mechanism to adaptively combine retrieved future trajectories, persistence forecasts, and backbone predictions.

The experimental results are quite competitive. The proposed GNA improves GridTST and iTransformer in 43 and 42 out of 48 dataset–horizon settings, respectively, and achieves the lowest MSE on 8 out of 12 benchmark datasets.

Overall, the proposed framework seems combine multiple existing components together. In addition, the robustness of generality of the proposed method remains unclear to me as its performance decrease substantially with longer lookback windows, and the proposed method even degrades performance on ETTh1 and ETTh2.

Critical strengths
  1. It is a well-motivated paper. The combination of global and per-variate retrieval, which is similar to ensemble learning, sounds interesting to me.
  2. The experimental results are competitive, with substantial improvements on several benchmarks.
Critical weaknesses
  1. The novelty is somewhat limited. The proposed framework mainly combines existing techniques, including nearest-neighbor retrieval, forecasting-based representation earning, adaptive gating, and cross-attention.
  2. The effectiveness of retrieval is highly dependent on the datasets and the lookback length. For example, the improvements decrease substantially with longer lookback windows, and the proposed method even degrades performance on ETTh1 and ETTh2. Thus, the robustness of generality of the proposed method may remain unclear to me.
Questions for authors
  1. Can you justify the performance improvement for longer lookback windows empirically?
1 comment

We thank the reviewer for the positive assessment of the motivation and the results.

W2 / Q1 (longer lookbacks). We extended the lookback study to three datasets. Change in MSE of GNA vs. the plain backbone at the same L (H = 96):

L = 48: ETTm1 −33.4%, PEMS08 −86.6%, Weather −4.3% L = 96: ETTm1 −8.0%, PEMS08 −72.6%, Weather −1.7% L = 192: ETTm1 −1.2%, PEMS08 −67.8%, Weather −3.9% L = 336: ETTm1 +5.3%, PEMS08 −52.6%, Weather +2.2%

On all three datasets, the gain is largest at L = 48 and smallest at L = 336. On the sensor network it persists at L = 336 (−53%). On ETTm1 and Weather, where 336 steps span 3.5 and 2.3 days, the plain backbone is slightly better. We agree with the reviewer's reading and state this scope explicitly: GNA supplies context that a short window lacks, and on ETTm1 and Weather it does not add to a long window. Short windows are often imposed rather than chosen. After a restart or a change of operating conditions, only a short stretch of relevant history exists, and processes that stop and start face this repeatedly. At L = 48 the gains are largest, and they extend to foundation models: with a 48-step context, GNA lowers the MSE of a frozen zero-shot TimesFM-3 from 3.33 to 0.105 (PEMS08) and from 3.53 to 0.311 (ETTm1).

W2 (ETTh1/ETTh2). We report and analyze this failure (Sec. 7, App. I). On three-seed means, GNA's MSE is 1.1% higher on ETTh1 and 7.2% higher on ETTh2. The loss concentrates at the longest horizon. In the seed-2023 runs, GNA is 8.8% (ETTh1) and 22% (ETTh2) worse at H = 720, while it is better at H = 96 on both (−0.2%, −4.0%) and at H = 336 on ETTh1 (−6.0%). The same recordings sampled every 15 min (ETTm1, ETTm2) improve at every horizon, so this is not a failure on the domain as such. Hourly data at H = 720 forecast 30 days ahead from a 4-day window, with a 4× smaller bank. A hindsight oracle suggests that level-free retrieved futures would cut the ETTh error by 26–36%. A level correction of retrieved futures is our next step.

W1 (novelty). Nearest neighbors, learned embeddings, gates and attention are standard tools, and we do not claim them. The contribution is what is retrieved and how it is assembled. Among the retrieval forecasters in Tab. 17, RAFT and RATD retrieve whole past windows; TS-RAG, RAF and TimeRAF retrieve univariate series separately; and GTR uses a learned periodic memory instead of past windows. GNA builds composite slots in which each variate takes the future of its own best-matching past, combines them with coherent whole-window slots, and learns per forecast step and variate how much to trust each. Reviewer 548Z also notes that neighbor assembly has received little attention. The ablations show that the combination matters: global slots alone raise the MSE by 1.9–8.3%, and per-variate slots alone by 0.3–15.8%. The best per-variate slot is 17–45% closer to the true future than the best global slot. Your ensemble view fits well: the two slot types are complementary members, and the gate weights them per step and variate.

Generality. One retrieval configuration serves all datasets. GNA with GridTST beats the plain GridTST in every seed on 10 of 12 datasets and in 44 of 48 settings on seed means (Wilcoxon p < 1e-8). With PatchTST as a new third backbone, GNA improves 15 of 16 settings. Halving or doubling the fixed fusion weights changes the MSE by at most 3.1%.

Reviewer_548ZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a way to identify examples for retrieval-augmented forecasting. In multivariate time series, prior work retrieves complete neighbors in past i.e. all variables of all neighbors. However, the authors claim that the best examples / neighbors may be different for different variables. Therefore, they propose a neighbor assembly method combining global and local neighbors. Global neighbors are the same as prior work, whereas local neighbors are those which are the most similar at a variate level.

Critical strengths

Retrieval augmented forecasting is an interesting research direction, and bears resemblance to in-context learning. I believe that simple methods in this direction can be quite impactful. Neighbor assembly is a sub-problem, which has received little attention, to the best of my knowledge.

Critical weaknesses

The paper is quite hard to follow, and it took me multiple passes to understand what the authors were trying to do. A few things are also unintuitive or are adequetely motivated, for example, ρ,rj\rho, r_j on lines 194/195, and what teacher is ϕ\phi distilled from in Line 190. Similarly, the computational complexity in Line 260 is abrupt, with no accompanying description. It is clear that parts of the paper were written by AI (e.g., “With the same lookback … retrieved futures.”, "No practical window reaches a situation last seen weeks earlier. In some applications a short lookback is also a requirement rather than a choice:").

I am also concerned that the experimental settings do not inspire confidence. First, most datasets used by the authors have known issues. For example, in Exchange, repeating the last value can beat most forecasters. Recent studies use benchmarks such as fev-bench, GIFT-Eval etc. for a much better read on forecasting performance. Second, on the experimental protocol: without multiple independent runs, it is unclear if the differences are meaningful. Third, the authors use GridTST as a baseline, but it is not peer reviewed. While that is not a problem in and of itself, recent work [1] has catalogues a large number of multivariate time series forecasting methods, any of which can be used as backbones. For example, it is unclear why PatchTST, a simple patched transformer was not used as a backbone. I am also curious if these methods benefit natively multivariate forecasting foundation models such as Chronos 2.

References

  1. Potosnak, Willa, et al. "MICA: Multivariate Infini Compressive Attention for Time Series Forecasting." arXiv preprint arXiv:2604.06473 (2026).
Questions for authors
  • Line 190: “Retrieval uses a small Transformer ϕ distilled from the teacher” – which teacher?
  • Line 192: “A learned embedding can miss coarse context such as the operating level or the time of day, so we combine its ranking with three simple context signals j: the similarity of low-dimensional regime features, lags at which the training series is periodic, and calendar match.” – Does the forecaster use the same features? If not, why does the embedder need these features specifically?
  • What is line 194/195? What is ρ,rj,…\rho, r_j, …?
  • Line 199: “The search keeps C = 200 candidates, spread over different episodes so that one episode cannot fill the list.” – What if there are more than 200 candidates?
  • “Global slots preserve cross-variate structure, while per-variate slots find a closer match when variates decouple.”-- How does the model know the difference between global slots, and the slots with per-variate candidates?
  • Line 211: “The five global slots take the five top-ranked candidates for all variates (sk,d = sk).” – Why 5 top-ranked candidates? If this is an example, we should make that explicit.
  • Line 239: “The persistence forecast, i.e. the query’s last value repeated H times”: What is the intuition behind this persistence forecast? Why are the learned scalars α\alpha and β\beta initialized at 0.1? I would assume alpha to be much higher initially.
  • Line 213: “A second fused list, built as above but with candidates at least L steps apart, provides a pool of M = 50 candidates.” Where are these 50 candidates coming from? Why do they need to be atleast L steps apart?
  • Line 258: “forecaster is trained end to end with MSE, with the embedder and bank fixed”. Bank is the set of candidate windows, right? So how can you keep the bank fixed during training? Do you mean that the bank is computed ahead of time, and only valid candidates are ranked?
  • Line 260: What is dzd_z and drd_r?
  • “At our scales the bank is built in 3–7 minutes”: This depends on the compute infrastructure.
  • “Every run keeps the checkpoint with the lowest validation MSE over its 10 epochs.” How is the validation dataset selected?
  • How many repetitions per run? How do we know that differences in table 1 are significant?
  • Line 323: “Retrieved futures therefore help only when they are relevant, and the model cannot fully”-- What does relevant mean? Without a precise or grounded definition, this statement is a moot point. Of course, if relevant things are retrieved that would improve performance.
  • “gate puts substantial weight on persistence in the first forecast steps and moves it to retrieved futures further out” – why would retrieved neighbors, much farther along in the past, be more valuable than the lookback window? Is it explained by periodicity / cyclical trends? If that’s the case, then the lookback window should capture that.
  • “The error of GNA changes little with the lookback (0.297–0.310)” – Is that little, in comparison to differences in the Table 1? In Figure 4: What is rank fusion?
  • Line 42: “a hindsight oracle”: What is a hindsight oracle?
2 comments

We thank the reviewer for the detailed reading and the many concrete questions, which helped us improve the clarity.

Multiple runs and significance. All twelve datasets were run with three seeds (App. E, Tab. 11), for GNA with both backbones and for the plain GridTST. GNA with GridTST beats the plain GridTST in all three seeds on 10 of 12 datasets (one-sided Welch test, p ≤ 0.022 for each), and in 44 of 48 settings on seed means (Wilcoxon p < 1e-8). Its seed-to-seed variation is a median 0.3% of the MSE (4% for the plain backbone). On the eight datasets where GNA has the lowest MSE, its three-seed mean is also below every published baseline in App. D. The smallest margin is on ETTm1, 1.7 of GNA's seed standard deviations.

Datasets. Exchange is not one of our datasets. On the twelve we use, repeating the last value is far from competitive. Horizon-averaged MSE, persistence vs. GNA: PEMS03/04/07/08 0.57–0.64 vs. 0.065–0.082; ETTh1 1.321 vs. 0.463; ETTh2 0.539 vs. 0.420; ETTm1 1.271 vs. 0.359; ETTm2 0.385 vs. 0.272; ECL 1.612 vs. 0.153; Solar 1.451 vs. 0.198; Traffic 2.765 vs. 0.397; Weather 0.352 vs. 0.235. We use these benchmarks because the methods we compare with report results on them, which allows comparisons with published numbers at the same lookback. We agree that broader benchmarks are valuable. fev-bench and GIFT-Eval target general-purpose forecasters, in particular pretrained foundation models, across many heterogeneous series; GNA instead trains a retrieval layer on one multivariate series with a long history.

Backbones. GNA needs only a forecast and per-variate tokens. Besides GridTST, we use iTransformer (ICLR 2024). As suggested, we added PatchTST (ICLR 2023), with its official architecture and hyperparameters, trained with our common recipe; for PEMS, which PatchTST did not run, we use its large-dataset setting. GNA improves PatchTST in 15 of 16 dataset–horizon settings. Averaged over horizons, the MSE changes from 0.313 to 0.076 on PEMS08, 0.256 to 0.237 on Weather, 0.381 to 0.362 on ETTm1, and 0.466 to 0.470 on ETTh1. On ETTh1, ETTm1 and Weather, our plain PatchTST is on par with the L = 96 results published for PatchTST (0.469/0.387/0.259, Tab. 1), so the gains there are not over a weak baseline. On PEMS08 our plain PatchTST (0.313) is weaker than the published 0.280, but GNA with PatchTST (0.076) is 73% below that published number. Our plain GridTST is within 2.5% of GridTST's own L = 96 results on Weather/Traffic/ECL/Solar (0.245/0.412/0.174/0.228 vs. 0.242/0.402/0.175/0.229); the submission wrongly said that none existed.

Foundation models. GNA can wrap a frozen foundation model, and we tested this with a short context. With a 48-step context and H = 96, zero-shot TimesFM-3 is worse than persistence (MSE 3.33 vs. 1.48 on PEMS08; 3.53 vs. 1.21 on ETTm1). With only a learned output scale it reaches 1.41 and 1.20. GNA on top of the frozen model, training only the retrieval layer, reaches 0.105 and 0.311. With the 5-step window of PEMS04 (one-step forecasts), GNA (MSE 0.053) is also better than zero-shot TimesFM-3 with contexts of 5, 96 and 512 steps (0.070/0.062/0.056) and Timestral-209M with a 5-step context (0.065). We did not test Chronos-2.

Presentation. We will rewrite Sec. 3 so that every symbol is defined where it appears (answers below). As stated in our AI-use statement, an LLM assisted with drafting parts of the text, which we edited and checked. We will rewrite the passages you quote.

Answers (line numbers refer to the submitted PDF):

l.190, which teacher? The teacher φ_T of Eq. 2: an MLP on the flattened window, trained so that a head predicts the window's future. Retrieval uses a small Transformer φ trained to reproduce φ_T's embedding (App. A). φ_T's first layer alone has 512·L·D weights (42M on Traffic), and the small student keeps embedding the bank and online queries cheap.

l.192, context signals. They are not embedder inputs but three scores fused with the embedding ranking (Eq. 3). Regime features and periodic lags are used only for ranking. Calendar information also enters the forecaster where time stamps exist: the gate sees each slot's calendar match and the query's time features, and iTransformer takes time-stamp tokens as in its original design. Retrieval needs these signals because the embedding sees only the normalized window values, without time stamps, and can miss the time of day or the operating level. Without them, the MSE rises by 4.1/12.0/4.5% on ETTm1/ETTm2/Weather, but not on PEMS.

l.194–195, Eq. 3. Each shortlisted candidate (next answer) is ranked by embedding similarity, by regime-feature distance, by whether its time lag is a periodic lag of the series, and, where time stamps exist, by calendar mismatch. r̄ ∈ [0,1] is the rank scaled to [0,1] (0 = best). ρ(s) = r̄_φ(s) + Σ_j w_j r̄_j(s), with fixed weights w = (1, 0.25, 0.5), smallest first. This is the "rank fusion" of Fig. 4.

l.199, more than 200 candidates? Except for the earliest training queries, thousands of windows are eligible, and all are searched exactly. The embedding ranking and the regime ranking each keep 100 episode-diverse candidates. These, together with the windows at periodic lags, are fused into a shortlist of C = 200, and the 5 global slots are its top 5. Taking 400 or 1000 per ranking changes the MSE by at most 1.0%.

Global vs. per-variate slots. Each slot carries an explicit type indicator. It is a gate input, and a learned slot-type embedding is added to the slot's future token in the cross-attention (l.253).

l.211, why five? It is not an example: all main experiments use K = 10 slots, 5 global and 5 per-variate (Tab. 3). K = 5 and K = 20 are in Tab. 14.

l.239, persistence; α, β = 0.1. Persistence is a strong guess for the first steps and the gate's fallback when no slot fits. With GridTST at H = 96, the gate puts weight 0.29 (PEMS04) and 0.42 (Solar) on it over steps 1–12, and 0.03 and 0.08 over steps 49–96. α and β start small so that the freshly initialized branches do not dominate the initial forecast. Only α·f matters, because f's output layer is learned. The plain backbone uses the same initialization and is within 2.5% of GridTST's published L = 96 results. Learned values with GridTST: α 0.05–0.31, β 0.04–0.49.

l.213, M = 50. The pool comes from the same fused ranking, with candidates taken greedily at least L steps apart. Windows at stride 1 overlap almost completely; without spacing, per-variate slots could be near-copies of one episode.

l.258, "bank fixed". Yes: after the embedder is trained, the slots of every query are computed once from causally eligible windows (Eq. 1) and stay fixed while the forecaster trains. At test time, windows become eligible once their futures are observed (Fig. 7).

l.260, d_z, d_r. These are the embedding size (32 or 128, Tab. 5) and the size of the regime features. We will define both in the main text.

"3–7 minutes": agreed; we will state the hardware in the text (one RTX PRO 6000 GPU, 16-core CPU; App. G). Validation set. Standard chronological splits (Tab. 7): validation is the block between training and test, and its queries retrieve only from the past.

l.323, "relevant". Operationally: the slots retrieved for this query, compared with the same architecture fed the slots of a random other query (mismatched control, App. F). The mismatched control loses the entire gain. Tab. 12 measures how close the retrieved futures are to the truth. Across all settings, the gain grows with the weight the gate puts on neighbors (Spearman 0.71 with GridTST, App. I).

Why far-past neighbors further ahead? A 96-step window does not cover the cycle: it spans 8 h of the day on PEMS and 16 h on Weather. It is more than periodicity: GNA is 5–49% below the explicitly periodic CycleNet, TQNet and GTR on every PEMS dataset. Your intuition holds once the window covers the cycle: at L = 336 the plain backbone is slightly better than GNA on ETTm1 and Weather.

"Changes little (0.297–0.310)": a 4% spread vs. 62% for the plain backbone over the same lookbacks. This is not negligible next to the gaps in Tab. 1, so we will state this comparison.

l.420, hindsight oracle. Analysis only: among the same eligible past windows, it picks those whose futures are closest to the true future. Because it uses the target, it gives an optimistic estimate of what better retrieval could achieve (App. I).

Reviewer_2XkeOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GNA, a retrieval-augmented layer for multivariate time-series forecasting that assembles neighbors at two granularities: whole past windows (global slots) and per-variate neighbors, and learns a per-step, per-variate gate to weigh these retrieved futures against a persistence baseline and the backbone prediction. A future-aware retrieval embedding is trained with privileged information (predicting each window’s future via a teacher and distilling to a student used at test time), and retrieval is strictly causal with a bank that grows during deployment. Under a fixed-lookback protocol across 12 benchmarks, GNA improves two Transformer backbones in most settings, achieves the lowest MSE on 8 datasets, includes ablations showing both granularities are necessary, and provides analyses of when retrieval helps (e.g., further horizons, run starts) and where it fails (ETTh long horizons due to level drift).

Critical strengths
  1. Addresses a central limitation of deep forecasters (finite lookback) in a way that preserves causality and enhances far-horizon accuracy.

  2. Demonstrates substantial performance gains on sensor-network datasets and competitive results broadly, providing a general, backbone-agnostic layer with low deployment latency.

  3. Offers diagnostic insights (gate weights over horizon, run-start benefits) that can inform future retrieval-augmented designs for time series.

Critical weaknesses
  1. The offline bank construction is O(N^2) in time (exact search and ranking for all queries), which may become prohibitive on much longer datasets.

  2. The method relies on multiple engineered components (hybrid rank fusion with fixed weights, handcrafted per-variate matching features, and a gated MLP with handpicked inputs), which, while effective, reduce end-to-end learnability and may complicate transfer to novel domains.

Questions for authors

Beyond ETTm1, did the authors test additional lookback lengths on other datasets to more broadly characterize the substitution effect between retrieval and longer context windows?

1 comment

We thank the reviewer for the positive assessment.

Q1 (lookback beyond ETTm1). We ran the lookback study on PEMS08 and Weather with the same protocol: GridTST, H = 96, a fresh embedder and bank per L. Change in MSE of GNA vs. the plain backbone at the same L:

L = 48: ETTm1 −33.4%, PEMS08 −86.6%, Weather −4.3% L = 96: ETTm1 −8.0%, PEMS08 −72.6%, Weather −1.7% L = 192: ETTm1 −1.2%, PEMS08 −67.8%, Weather −3.9% L = 336: ETTm1 +5.3%, PEMS08 −52.6%, Weather +2.2%

On all three datasets, the gain is largest at L = 48 and smallest at L = 336, so retrieval substitutes for context that a short window lacks. On the sensor network, retrieval still halves the error at L = 336 (28 h of 5-min data). On ETTm1 and Weather, a 336-step window spans 3.5 and 2.3 days, and the plain backbone is slightly better. With the monitoring data removed, these public results replace the run-start analysis as short-window evidence. PEMS04 with a 5-step window (one-step forecasts) is another such case: there GNA lowers the MSE by about 10% vs. both backbones.

W1 (O(N²) offline bank). The quadratic term comes from precomputing the slots of every query offline, and its core is a dense matrix product. We timed exact causal top-200 search for all queries, on banks built by replicating our Weather embeddings (d = 128) with small noise. On one GPU it takes 0.3 s at 5·10^4 windows, 1.1 s at 2·10^5 and 21 s at 10^6, about 17 times our largest bank. Fusion and slot assembly are linear in N. In deployment, each forecast needs one retrieval step, with a median of 19–25 ms on a CPU (Tab. 16). Approximate search (HNSW) could replace exact search: on our real Weather and Solar embeddings it recovers the exact 200 nearest neighbors (recall 1.00), and 0.95 on the replicated 10^6-window bank.

W2 (engineered components). These components contribute to the gain. They are fixed and shared by all 12 datasets rather than tuned per dataset. At H = 96:

Rank fusion. Without it, the MSE rises by 4.1/12.0/4.5% on ETTm1/ETTm2/Weather and changes by −1.1/+0.4% on PEMS04/PEMS08, which have no time stamps (Tab. 13). Halving or doubling all fusion weights changes it by at most 3.1% (ETTm1 +3.1/0.0%, Weather +1.4/+2.1%). Per-variate matching. Replacing the multi-scale distance by a plain MSE changes the MSE by at most 0.6% (ETTm1 +0.3%, Weather +0.6%, PEMS08 0.0%). Candidate pool. Taking 400 or 1000 candidates per ranking instead of 100 changes it by at most 1.0%, slightly better (ETTm1, PEMS08). Gate inputs. These are generic statistics, not domain features: how well a slot's past fits the query's past (MSE, MAE, correlation, end-point gap, variance), the slot type, the system-wide fit, calendar information, and the change the slot proposes (App. A). The gate is learned end to end. Keeping the ranking fixed is a deliberate trade-off. The bank stays causal and cheap, and it grows at test time without re-indexing; a retriever trained end to end would require re-embedding the bank whenever it changes. We use the same configuration unchanged across five domains (traffic sensors, electricity, power transformers, solar, weather) and three backbones.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of new results

We thank all reviewers for their careful reading. Reviewers valued the problem's motivation (z6ZL), the causal, backbone-agnostic design and the far-horizon and sensor-network gains (2Xke), and found neighbor assembly an under-studied sub-problem (548Z). All new runs use seed 2023, 10 epochs and causality-audited banks, with L = 96 unless stated. Line numbers refer to the submitted PDF.

Lookback (2Xke, z6ZL). The lookback study now covers three datasets (L = 48, 96, 192, 336). On all three, GNA's gain over the plain backbone is largest at L = 48 (MSE: ETTm1 −33%, PEMS08 −87%, Weather −4%) and smallest at L = 336. On the sensor network it persists at L = 336 (PEMS08 −53%). On ETTm1 and Weather, the plain backbone is slightly better at L = 336 (GNA +5%, +2%). GNA is therefore most useful where the window has to stay short, e.g. after restarts, in start–stop operation, right after deployment or under latency limits. We state this scope explicitly. Third backbone: PatchTST (548Z). GNA improves PatchTST in 15 of 16 dataset–horizon settings. Averaged over horizons, the MSE changes by −76% on PEMS08, −8% on Weather, −5% on ETTm1 and +1% on ETTh1. Our plain PatchTST is on par with the L = 96 results published for it (ETTh1/ETTm1/Weather 0.466/0.381/0.256 vs. 0.469/0.387/0.259). Foundation model (548Z). With a 48-step context, zero-shot TimesFM-3 is worse than persistence at H = 96 on PEMS08 and ETTm1. GNA on top of the frozen model lowers its MSE from 3.33 to 0.105 (PEMS08) and from 3.53 to 0.311 (ETTm1). Significance (548Z). Three seeds on all 12 datasets were in App. E. GNA with GridTST beats the plain GridTST in all three seeds on 10 of 12 datasets (one-sided Welch test, p ≤ 0.022 for each), and in 44 of 48 settings on seed means (Wilcoxon p < 1e-8). Datasets (548Z). Exchange is not one of our datasets. On all twelve, GNA's MSE is 22–91% below persistence. Fixed components (2Xke). At H = 96, halving or doubling all rank-fusion weights changes the MSE by at most 3.1% (ETTm1, Weather). A plain-MSE per-variate match changes it by at most 0.6% (ETTm1, Weather, PEMS08). Taking 4× or 10× more candidates per ranking changes it by at most 1.0%, slightly better (ETTm1, PEMS08). Scalability (2Xke). Exact causal top-200 search for all queries of a 10^6-window bank takes 21 s on one GPU. Corrections. (a) GridTST does report L = 96 results for four of our datasets (its App. Table 6). Our plain GridTST is within 2.5% of them (Weather/Traffic/ECL/Solar 0.245/0.412/0.174/0.228 vs. 0.242/0.402/0.175/0.229). (b) "Untuned defaults" (l.286) was imprecise. The retrieval constants were fixed during development and are identical for all datasets; item 6 shows how much varying them changes the results. (c) Compared with the three-seed mean of the backbone, the mismatched-neighbor control loses the entire gain. It is worse than all three backbone seeds on PEMS08, ETTm1 and ETTm2, and within their range on PEMS04 and Weather; the submission said it was worse than no retrieval on all five datasets. (d) Shortlist (l.199): the embedding ranking and the regime ranking each keep 100 episode-diverse candidates. These, together with the windows at periodic lags, are fused into a list of C = 200. For data-governance reasons, the revised PDF removes the experiment-monitoring dataset and all results on it, including the run-start analysis. No benchmark result depends on it, and item 1 provides public short-lookback evidence instead. The final version will include these results and corrections, and Sec. 3 will define every symbol where it is used.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.