acceptodds
Under review as a conference paper at ICLR 2027

Decomposed Local-to-Global Retrieval for Time Series Forecasting

Abstract

Time series forecasting fundamentally requires capturing both short-term dynamics and long-term structural patterns. While parametric models often struggle to internalize these diverse dependencies from limited data, retrieval-based approaches offer a promising alternative by explicitly reusing informative historical patterns. However, existing methods typically perform retrieval in a monolithic feature space, overlooking that varying temporal structures (\eg, trend vs. seasonality) require distinct matching mechanisms. In this paper, we propose DLoG (Decomposed Local-to-Global retrieval), a framework that augments forecasting by leveraging decomposed historical patterns as inductive biases. DLoG specifically addresses the multi-scale nature of time series through two key designs: (1) Component-wise Retrieval, which decomposes series into complementary components to enable targeted pattern alignment; and (2) Dual-stage Retrieval, which first retrieves from a local window to preserve short-term consistency, followed by a global search to incorporate long-range structural regularities. By integrating these component-specific future references, DLoG reduces the burden on purely parametric forecasting by explicitly reusing informative historical patterns. Extensive experiments on nine benchmark datasets show that DLoG achieves best or second-best average MSE and MAE across all evaluated datasets, demonstrating strong and consistent performance against competitive forecasting baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, Retrieval augmented model, Deep learning

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

DLoG is a retrieval-augmented framework for multivariate long-term forecasting. It adaptively decomposes each series into trend and seasonal components, performs local patch retrieval, and then searches a global training-memory repository separately for the two components. Seasonal matching uses shift-tolerant correlation, whereas trend matching combines correlations of the raw component and its first differences. Retrieved component futures, local references, and lightweight parametric forecasts are fused to form the prediction. Across nine benchmarks, DLoG reports consistently strong accuracy relative to forecasting and retrieval-based baselines. The paper also provides component ablations, three-seed main results, leakage-aware memory construction, and a detailed scalability analysis.

Critical strengths
  1. The proposed inductive bias is plausible and practically useful. A single similarity score can be dominated by one component of a mixed signal, whereas separate searches can find a suitable trend and seasonal analogue even when no historical window jointly matches both. The matched holistic-retrieval ablation improves consistently on five datasets, with especially substantial gains on ECL, Traffic, and Weather. This is meaningful initial evidence for decomposed retrieval.

  2. The empirical evaluation is broad and unusually attentive to retrieval mechanics. The study covers nine datasets, including high-dimensional ECL and Traffic, and compares against several retrieval and long-context approaches. The explicit overlap mask and training-only test repository address leakage concerns. Repository refresh cost, online latency, scaling with repository size, and the accuracy–efficiency trade-off are also reported clearly.

Critical weaknesses
  1. The holistic control changes several factors simultaneously. It changes the representation, similarity functions, neighbor identities, future aggregation, and component-specific post-processing. Table 2 therefore shows that the complete DLoG system outperforms this holistic variant, but does not isolate which design choice causes the gain.

  2. Separate retrieval introduces a compositional advantage beyond decomposition. The trend and seasonal branches may select different historical samples, increasing the chance of finding individually suitable matches. The paper does not separate this benefit from that of decomposed representations or component-specific similarity rules.

  3. Recombining independently retrieved components requires an untested assumption. DLoG implicitly assumes that a trend future from one historical episode and a seasonal future from another can be combined into a valid forecast. This may hold for approximately additive, weakly coupled processes, but its behavior under strong trend–seasonality coupling is not examined.

  4. Seasonal shift-tolerant matching may leave the retrieved future misaligned. The manuscript does not state that the maximizing temporal shift is applied inversely to the retrieved seasonal future. Without this step, a phase-shifted key can achieve high similarity while still supplying a future with the wrong phase; amplitude and bias calibration cannot correct that error.

  5. The trend similarity does not measure level similarity as claimed. Pearson correlation is invariant to additive level shifts. Equation (8) therefore measures centered shape agreement and first-difference correlation, rather than similarity in absolute trend level.

  6. The need for distinct component-specific metrics is not directly validated. Table 9 compares a common correlation metric with generic distances, but does not ablate seasonal shifting, trend differencing, or applying the same metric to both components. The paper therefore provides limited direct evidence that different components require different matching rules.

Questions for authors
  1. What exactly is held constant in DLoG w/ Holistic Retrieval, including the number of searches and candidates, parameters, retrieved values, calibration/refinement modules, and fusion inputs?

  2. Is the maximizing shift in Equation (7) recorded and applied to the associated seasonal future before aggregation? If not, how is phase consistency between the query and retrieved future maintained?

  3. Can the authors report a factorial comparison between (i) decomposed representations with a combined score and shared neighbors, (ii) decomposed representations with separate neighbors but a common metric, and (iii) the full component-specific design? A capacity- and retrieval-budget-matched two-head holistic control would also be informative.

  4. Can the authors evaluate retrieval quality before the final forecasting fusion, such as retrieved-future error for holistic versus decomposed retrieval, and test controlled signals where trend/seasonal coupling and decomposition quality are varied? This would distinguish improved retrieval from additional forecast capacity and identify when recombining component futures is valid.

2 comments
AuthorsOpenReview ↗

Response to Reviewer TDHt Part1

W1 & Q1: Holistic Retrieval Control

We thank the reviewer for highlighting the importance of isolating the contributions of different retrieval components. DLoG is designed as an integrated component-aware retrieval framework, where decomposition, component-specific matching, and learned forecasting jointly contribute to prediction performance.

In Table 2, the holistic variant follows the same overall retrieval architecture and training configuration, while changing the retrieval representation and matching strategy. These changes can also affect neighbor selection, retrieved references, and their subsequent processing. Therefore, Table 2 evaluates the effectiveness of the complete framework rather than independently isolating every design factor.

The results demonstrate that DLoG consistently outperforms holistic retrieval across all five datasets. For example, MSE decreases from 0.155 to 0.134 on ECL and from 0.495 to 0.425 on Traffic. Furthermore, using both trend and seasonal retrieval branches outperforms either branch alone, supporting the effectiveness of integrating complementary component information.

To further distinguish the contributions of individual retrieval choices, we conducted additional controlled ablations, as reported in our response to W2, W6, and Q3. These experiments explicitly match the decomposition architecture, forecasting heads, trainable parameter count, and training protocol.

W2, W6 & Q3: Separate Retrieval and Component-Specific Matching

We appreciate the reviewer's suggestion to distinguish the benefits of independent neighbor selection from those of component-specific similarity rules. DLoG allows trend and seasonal components to retrieve historical patterns independently, while employing matching rules tailored to their temporal characteristics.

To evaluate these design choices, we conducted additional experiments on ETTh1 with three random seeds. All configurations use the same decomposition architecture, forecasting heads, trainable parameter count, and training protocol.

ConfigurationTest MSE ↓\downarrowTest MAE ↓\downarrowDistinct candidates
Shared neighbors, common Pearson0.3569 ±\pm 0.0030.3901 ±\pm 0.00260.3
Separate neighbors, common Pearson0.3610 ±\pm 0.0070.3935 ±\pm 0.005101.0
Separate, full matching0.3531 ±\pm 0.0040.3852 ±\pm 0.00498.3
Without seasonal shift0.3576 ±\pm 0.0050.3910 ±\pm 0.00484.3
Without trend differencing0.3608 ±\pm 0.0070.3932 ±\pm 0.005108.8

The full component-specific design achieves the lowest mean MSE (0.3531), outperforming both shared-neighbor retrieval with a common Pearson metric (0.3569) and separate-neighbor retrieval with a common Pearson metric (0.3610). Notably, separate neighbors with a common metric do not improve performance over shared neighbors. In contrast, introducing component-specific matching reduces MSE from 0.3610 to 0.3531 (approximately 2.2%). Removing seasonal shift or trend differencing further increases MSE to 0.3576 and 0.3608, respectively.

These results support the importance of component-specific matching beyond independent neighbor selection alone. In particular, trend differencing and seasonal shift tolerance both contribute to the full design's performance on ETTh1.

AuthorsOpenReview ↗

Response to Reviewer TDHt Part2

W3 & Q4: Component Recombination and Pre-Fusion Retrieval Quality

We thank the reviewer for highlighting component recombination. DLoG uses trend and seasonal references as complementary forecasting information, rather than requiring their combination to correspond to a single observed historical episode. These references are processed together with query-derived local forecasts through learned calibration and fusion modules to generate the final prediction.

To investigate retrieval quality before forecasting fusion, we conducted additional experiments under both synthetic and real-data settings. Reference MSE measures the error between directly aggregated retrieved futures and the ground truth, without learned calibration or forecasting fusion. The synthetic experiment introduces strong trend-seasonality coupling through a trend-slope/seasonal-amplitude correlation of 0.95, together with decomposition leakage of 0.2. For real ETTh1, all retrieval strategies use the same last-input level anchor to ensure a consistent evaluation.

Table 1. Pre-fusion retrieval quality (Reference MSE ↓\downarrow).

SettingRetrieval strategyReference MSE ↓\downarrowDistinct episodes
SyntheticShared Top-100.0082010.0
SyntheticSeparate Top-100.0058217.1
ETTh1Shared / Common0.6693—
ETTh1Separate / Common0.6557—
ETTh1Full component-specific0.6433—

In the synthetic experiment, Separate Retrieval reduces reference MSE by approximately 29% compared with Shared Retrieval, even under strong component coupling and imperfect decomposition. This demonstrates that independent component retrieval can produce more accurate point references beyond the weakly coupled setting. The two strategies use the same top-KK per component, although Separate Retrieval accesses more distinct historical episodes.

On real ETTh1, Separate Retrieval with a common metric improves reference MSE from 0.6693 to 0.6557, while full component-specific matching further reduces it to 0.6433, achieving the best retrieval quality among the evaluated strategies. This corresponds to an overall improvement of approximately 3.9% over Shared/Common Retrieval. These results provide direct evidence that both independent component retrieval and component-specific matching contribute to improved pre-fusion reference accuracy, complementing the superior final forecasting performance of the full DLoG framework.

W4 & Q2: Seasonal Shift and Phase Consistency

We thank the reviewer for pointing out the distinction between shift-tolerant matching and explicit future phase alignment.

In Eq. (7), seasonal similarity is computed by maximizing correlation over a bounded range of circular temporal shifts from −3-3 to 33. This enables the retrieval mechanism to recognize similar periodic patterns despite small phase differences.

In the current implementation, the maximizing shift determines the similarity score and neighbor selection, but is not explicitly retained or inversely applied to the retrieved seasonal future. Therefore, DLoG performs shift-tolerant neighbor selection rather than explicit phase alignment of retrieved future values.

The retrieved seasonal reference is subsequently processed through query-conditioned forecasting and amplitude/bias calibration. However, phase consistency is not explicitly enforced for each retrieved future, and these operations do not guarantee correction of an arbitrary phase error.

Our ETTh1 ablation shows that removing seasonal shift increases MSE from 0.3531 to 0.3576 and MAE from 0.3852 to 0.3910. These results support the contribution of shift-tolerant matching to forecasting performance, although they do not independently establish future phase alignment.

W5: Interpretation of Trend Similarity

We appreciate the reviewer's observation regarding the interpretation of trend similarity in Eq. (8).

The trend similarity function is designed to capture centered trend-shape agreement and local temporal changes, rather than directly comparing absolute trend levels. Specifically, the Pearson correlation term measures similarity between centered trend trajectories, while the first-difference term captures local increment patterns, reflecting growth and decline dynamics.

This formulation allows DLoG to identify similar trend evolution patterns across historical segments with different absolute levels. The current input's absolute level is handled separately through last-value normalization and restoration in the forecasting pathway.

The ETTh1 ablation further supports the contribution of first-difference matching: removing trend differencing increases MSE from 0.3531 to 0.3608. We will revise the terminology from “level similarity” to “centered trend-shape and local-change similarity” to accurately describe the implemented mechanism.

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

This paper focuses on retrieval-based approaches for long-term Time-series forecasting. Unlike previous works performing retrieval in a monolithic feature space, the proposed DLoG (Decomposed Local-to-Global retrieval) performs retrieval in the decomposed seasonal and trend spaces. Moreover, DLoG performs dual-stage retrieval, which first retrieves from a local window for short-term consistency, followed by a global search to incorporate long-range structural regularities. Extensive experiments on nine benchmark datasets show that DLoG achieves the best or second-best average MSE and MAE across all evaluated datasets.

Critical strengths
  1. The proposed method achieves good empirical performance.
  2. The ablation study and model analysis are thorough.
Critical weaknesses
  1. The writing need to be improved.

    • Several terms are not clearly defined:
      • the term Local Window WlocalW_\text{local} appears in Sec. 4.4 and is not defined beforehand.
      • Δt\Delta_t in Eq. (8) is actually not defined.
      • The term shift-tolerant normalized correlation for Eq. (7) is not well defined either. Does it denote correlation with part of q\mathbf{q}, or $\mathbf {q} $ is shifted beyond a single sample?
      • The Corr\text{Corr} operator in Eq. (7) is not defined as well. Is it the same as the channel-averaged Pearson correlation introduced in Eq. (6).
    • The writing organization is not good.
      • The retrieval database is only used for global retrieval. However, this paper first introduces the retrieval database, then talks about local retrieval, and afterwards comes back to global retrieval. This organization makes the reading hard.
  2. I have some concerns about the fairness of the comparison.

The time-series datasets are usually constructed by cutting a long time series into multiple samples. However, this is only for dataset construction, and most baselines treat the datasets as isolated samples: the model takes an input of a lookback window LL and predicts the next HH points.

However, the local retrieval in DLoG uses Wlocal>LW_\text{local} > L. This means that the local retrieval goes beyond independent samples and requires the ordinal relationship among samples.

I don't think this setting is invalid for time-series forecasting, which is more like a streaming setting. However, it is indeed different from the common setting followed by baselines. Thus, this introduces unfairness in empirical comparison.

I believe that the authors should have more discussion regarding it. Potentially, add an experiment with Wlocal≤LW_\text{local} \leq L for a fair comparison and justify its applicability to the isolated-sample setting.

Questions for authors

Please check the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer gEFL Part1

Response to Reviewer gEFL

W1. Clarification of Definitions and Organization

Thank you for pointing out these issues. We agree that the terminology, mathematical definitions, and organization of the retrieval section need to be improved. We will revise the manuscript accordingly, providing clearer definitions of WlocalW_{\mathrm{local}} and Δt\Delta_t.

Regarding Eq. (7), the shift-tolerant normalized correlation refers to circularly shifting the query qq along the temporal dimension within the range [−Δ,Δ][-\Delta,\Delta] before computing its correlation with the key kk. Specifically, each shift moves the temporal positions of qq while preserving its original sequence length. Values shifted beyond one boundary wrap around to the opposite boundary. Therefore, this operation does not use only a partial segment of qq, nor does it extend qq beyond the original sample. Instead, it allows the retrieval process to identify historical seasonal patterns with small temporal phase differences.

Furthermore, the Corr\mathrm{Corr} operator in Eq. (7) is exactly the same channel-averaged Pearson correlation defined in Eq. (6), applied to the temporally shifted query and the corresponding key. The correlation is computed for each channel and then averaged across channels. The only difference from Eq. (6) is that Eq. (7) additionally considers a range of circular temporal shifts and selects the maximum correlation.

We will explicitly clarify these definitions and operations in the revised manuscript. We will also improve the organization of the retrieval section by introducing Local Retrieval before the retrieval database and Global Retrieval, making the overall presentation more logical and easier to follow.

AuthorsOpenReview ↗

Response to Reviewer gEFL Part2

W2. Fairness of Historical Context Comparison

We thank the reviewers for raising this important concern regarding the use of historical information. We would like to clarify the motivation behind our local retrieval setting and provide additional experimental evidence under a strictly controlled historical context.

1. The use of extended historical windows follows an established forecasting setting.

Our design is motivated in part by IRPA (Efficiently Enhancing Long-term Series Forecasting via Ultra-long Lookback Windows, AAAI 2025), which explicitly investigates leveraging ultra-long historical sequences to improve long-term forecasting.

In IRPA, the input window length is selected from {480,1920,3600}\{480,1920,3600\}, while the refined representation has a fixed length of 96. Meanwhile, the compared baseline models use an input length of 96 (IRPA, Section 4, Table 2).

Following this established research direction, DLoG also explores how to efficiently exploit extended historical information. However, unlike IRPA, which refines ultra-long input sequences, DLoG maintains a primary input length of L=96L=96 and leverages additional historical observations through local patch retrieval.

Importantly, our original comparison already includes IRPA as a strong long-context baseline. As reported in Table 1, DLoG achieves lower MSE than IRPA on 6 out of 9 datasets and lower MAE on 7 out of 9 datasets. These results suggest that the effectiveness of DLoG cannot be attributed solely to the availability of longer historical contexts.

2. Additional Experiments with Restricted Local Historical Context

To further address the reviewers' concern regarding the local historical context, we conduct two complementary comparisons.

(1) Comparison under the original local retrieval settings.

DLoG (Original) and IRPA use exactly the same dataset-specific local retrieval window, as shown in Table 1.

Table 1. Comparison between DLoG and IRPA under the same local retrieval windows at H = 96 (MSE ↓).

DatasetShared W_localDLoG (Original)IRPA
ETTh14800.3530.361
ETTh24800.2650.270
ETTm119200.2940.307
ETTm219200.1600.164
Exchange4800.0820.089
Solar-Energy19200.1790.196
ECL4800.1340.143
Traffic4800.4250.429
Weather36000.1530.156

(2) Comparison under a restricted local historical context.

We further restrict DLoG's local retrieval window to W_local = 96, matching the standard input look-back length L = 96. The model architecture, retrieval mechanisms, and global retrieval memory remain unchanged. We compare it with the best non-IRPA baseline using L = 96.

Table 2. Forecasting performance under a restricted local retrieval window at H = 96 (MSE ↓).

DatasetDLoG (W_local = 96)Best Non-IRPA Baseline (L = 96)
ETTh10.3590.384
ETTh20.2870.295
ETTm10.3080.329
ETTm20.1680.175
Exchange0.0850.084
Solar-Energy0.1940.201
ECL0.1490.159
Traffic0.4690.534
Weather0.1780.172

The original results at H = 96 are taken from Table 14 (Appendix E) of our submission. The best non-IRPA baseline is selected from the methods reported in Table 14.

Under the restricted local window of W_local = 96, DLoG outperforms the best non-IRPA baseline on 7 out of 9 datasets. In particular, it achieves relative MSE improvements of 6.5% on ETTh1 and 12.2% on Traffic. Compared with the original configuration, restricting the local window results in an average relative MSE increase of approximately 7.7%.

These results demonstrate that DLoG maintains competitive forecasting performance under a restricted 96-step local retrieval window, while the original comparison confirms that DLoG outperforms IRPA under identical local retrieval windows.

Thanks for the detailed rebuttal and additional experimental results. I appreciate the authors' efforts in addressing my concerns.

However, I still have two concerns:

  1. Fairness of comparison. Although the proposed setting follows IRPA, I still believe accessing a longer historical context introduces an unfair comparison against standard forecasting baselines under the current benchmark protocols. The method essentially leverages longer histories efficiently. I would evaluate the work more favorably if it were positioned as efficient long-context forecasting.

  2. Local patch retrieval. I acknowledge the additional experiments with Wlocal=96W_{\mathrm{local}}=96. However, I have concerns about it. As mentioned in the response to Reviewer azs9, the default setting of patch length PP is 9696. When Wlocal=96W_{\mathrm{local}}=96, there appears to be at most one complete patch available. I am concerned about the degeneration of the local retrieval part in the additional experiments.

Overall, I appreciate the additional evidence, but my main concern regarding the fairness of the empirical evaluation remains. Therefore, I will maintain my rating of 2.

Reviewer_azs9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DLoG (Decomposed Local-to-Global retrieval), a retrieval-augmented framework for multivariate long-term time series forecasting. The authors argue that existing retrieval-augmented forecasting methods match look-back windows in a monolithic space where trend and seasonal signals are entangled, which can retrieve historical patterns with mismatched underlying semantics.

DLoG consists of four main stages:

Last-Value Normalization and Adaptive Multi-Scale Decomposition: Given an input window x ∈ ℝC × L, DLoG subtracts the last value x0 = x - x:, L (Eq. 1) and decomposes x0 = x(se) + x(tr) using channel-wise learnable 1D smoothing kernels (softmax over kernel sizes 𝒦 = {12, 25, 50}) combined via a gating network conditioned on channel mean and standard deviation (Eq. 2–5). Epoch-Wise Decomposed Memory Bank: For every training window (ki, vi), both key and value are decomposed into {ki(se), ki(tr)} and {vi(se), vi(tr)}. The repository is detached from gradients (sg) and refreshed at the start of each training epoch, with temporal-overlap masking during training to prevent future leakage. Local-to-Global Component Retrieval: First, Local Retrieval selects the top-R (R=3) historical patches from a local candidate window Wlocal that have the highest channel-averaged Pearson correlation with the anchor patch plast(m), aggregating them into a local reference p̄loc(m) and forming an expanded query set of Q = R + 1 patches. Second, Global Retrieval searches the training repository using shift-tolerant correlation sse(q, k) = maxδ ∈ [-Δ, Δ] Corr(Shiftδ(q), k) for the seasonal component (Eq. 7) and a convex combination of level and first-order difference correlation str(q, k) = (1-λ)Corr(q, k) + λ Corr(Δt q, Δt k) for the trend component (Eq. 8). Component Reference Calibration and Linear Fusion: A linear fusion head h maps the concatenation of length 6F—comprising the two parametric patch projections r(se), r(tr), two local-retrieval projections l(se), l(tr), amplitude/bias-calibrated global seasonal reference ṽ(se), and increment-refined global trend reference ṽ(tr)—to the horizon F, before adding back x:, L (Eq. 10–14).

Across nine multivariate benchmarks (ETTh1, ETTh2, ETTm1, ETTm2, Exchange, Traffic, Solar-Energy, Weather, ECL) at H ∈ {96, 192, 336, 720}, DLoG achieves the best or second-best average MSE and MAE on all nine datasets against seven baselines (RAFT, IRPA, PatchTST, TimesNet, DLinear, SCINet, Autoformer), supported by ablations (Table 2), TS-RAG/RAF comparisons (Tables 4 and 8), and repository-stride scaling analyses (Appendix D).

Critical strengths

Well-Motivated Component-Specific Similarity Design for Time-Series Retrieval: Separating retrieval into trend and seasonal subspaces with tailored similarity metrics—shift-tolerant correlation (maxδ ∈ [-Δ, Δ] Corr(Shiftδ(q), k)) to handle minor phase offsets in periodic components, and combined level + first-order difference correlation to align both trajectory and slope in trend components—is intuitive and directly addresses a real failure mode of monolithic time-series retrieval. The controlled ablation in Table 2 (DLoG vs. DLoG w/ Holistic Retrieval, Only Trend, Only Seasonal, and w/o Adaptive Decomposition) confirms that component-wise matching improves over holistic retrieval under the same architecture. Transparent Computational Breakdown and Repository Subsampling Analysis: Appendix D candidly reports the exact wall-clock overhead of epoch-wise memory refreshes (42.5% of per-epoch training time on ETTm1) and online retrieval (76.8% of batch inference latency), and demonstrates how strided repository subsampling (s ∈ {1, 2, 4, 8}) trades off accuracy against latency.

Critical weaknesses

Unfair Historical Context Comparison (W_local up to 3600 vs. L = 96 for Baselines): Section 4.1 (Line 362) and Appendix B (Line 728) claim that "the input look-back length is fixed to L = 96 for DLoG and all compared baselines... Using identical input lengths and prediction horizons across all methods ensures a controlled comparison under the same amount of historical context." However, Table 3 (Page 9) and Appendix B Table 7 (Page 15) reveal that DLoG's Local Retrieval searches a dataset-specific local candidate window Wlocal ∈ {480, 1920, 3600} (480 on ETTh1/ETTh2/Exchange/ECL/Traffic, 1920 on ETTm1/ETTm2/Solar-Energy, and 3600 on Weather). Thus, at inference time, while baselines such as PatchTST, DLinear, TimesNet, and RAFT only observe the most recent L = 96 steps, DLoG accesses 5$\times$ to 37.5$\times$ longer recent history (Wlocal up to 3600 steps)! Indeed, Table 2 shows that removing Local Retrieval (w/o Local Retrieval, which restricts DLoG to L = 96 plus the training memory bank) degrades DLoG's MSE by 23.1% on ECL (0.134 -> 0.165), 27.1% on Traffic (0.425 -> 0.540), and 13.7% on Weather (0.153 -> 0.174), where w/o Local Retrieval falls behind IRPA on all five datasets. To claim a fair comparison under L = 96, Local Retrieval must either be restricted to patches strictly within the L = 96 look-back window, or all baselines must be given the same Wlocal look-back horizon. Undefined Architectural Quantities, Dimension Mismatches in Section 3, and Questionable TS-RAG / RAF Comparisons: Omitted Definitions and Dimension Mismatches: Section 3.3 divides components into patches of length P with stride r, and then matches expanded patch queries q ∈ ℝC × P against global memory keys ki ∈ ℝC × L (Eq. 7–8). However, the paper never specifies the numerical values of P or r (even in Table 6), never explains how Pearson correlation is computed when patch length P ≠ L or how boundary steps are handled under shift Shiftδ(q) (Appendix D.1 Line 915 refers to an undefined "effective temporal dimension D ≤ L"), and never specifies how target windows vi ∈ ℝC × H are normalized prior to decomposition or how a(se), b(se) in Eq. (12) are parameterized. Parameter Count Explosion from the 6F → F Fusion Head: Although Section 3.4 describes the projection heads as "lightweight", the linear fusion head h maps 6F → F (and ftr maps 3F → F), resulting in 3.23M parameters at H = 720 (Table 3)—an order of magnitude larger than DLinear or PatchTST. Inconsistent / Unexplained Comparisons with TS-RAG (Table 4) and RAF (Table 8): In Table 4, DLoG is compared against TS-RAG without specifying the horizon H (and the reported DLoG MSEs 0.350, 0.238, 0.285, 0.139, 0.072 do not match any horizon or average in Table 1 or Table 14). Moreover, TS-RAG is a zero-shot retrieval-augmented foundation model, whereas DLoG is trained supervised in-domain on each dataset. In Table 8, DLoG is evaluated using WQL (Weighted Quantile Loss) against RAF. However, DLoG is a deterministic point forecaster trained with MSE loss (Eq. 13–14) that does not output probabilistic quantiles; the paper never explains how WQL was computed for DLoG or why WQL drops by 62.2% on Weather at H=10 (0.166 -> 0.056) while MASE simultaneously worsens (1.265 -> 1.284). Missing Modern Baselines (2024–2026) and High-Dimensional Scalability: The main table omits standard 2024–2025 baselines such as iTransformer, TimeMixer, SparseTSF (cited in references but omitted from experiments), and pretrained time-series foundation models (Chronos, TimesFM, Moirai). Additionally, wall-clock complexity in Appendix D is only measured on ETTm1 (C = 7), omitting high-dimensional datasets like Traffic (C = 862) where 𝒪(Q N C D (2Δ+ 1)) online shift-correlation is over 100× more expensive.

Questions for authors

Clarification / Core Assessment — Role of Wlocal ∈ {480, 1920, 3600} vs. L = 96: During validation and testing, does Local Retrieval search Wlocal time steps preceding the forecast origin t (i.e., looking back up to 3600 steps into the recent validation/test stream)? What is DLoG's performance across all 9 datasets when Local Retrieval is strictly confined to patches inside the L = 96 look-back window so that it observes the exact same historical context as the baselines? Clarification / Core Assessment — Missing Patch Hyperparameters (P, r, D) and Value Normalization: What are the exact values of patch length P, stride r, and global matching dimension D? If P < L, how are patch queries q ∈ ℝC × P correlated against full-window global keys ki ∈ ℝC × L in Eq. (7)–(8)? Furthermore, how is each future value window vi ∈ ℝC × H normalized before decomposition in Section 3.3 (subtracting ki,:,L, subtracting vi,:,1, or unnormalized)? Clarification / Core Assessment — Protocols in Table 4 (TS-RAG) and Table 8 (RAF): What prediction horizon H is reported in Table 4, and why do DLoG's MSEs in Table 4 differ from Table 14? In Table 8, how was the probabilistic metric WQL computed from DLoG's deterministic point predictions? Clarification / Scalability — Wall-Clock Runtime on High-Dimensional Datasets (Traffic, C = 862): What are the per-epoch repository refresh time and per-batch online retrieval latency on Traffic (C = 862) and ECL (C = 321) corresponding to Table 11?

2 comments
AuthorsOpenReview ↗

Response to Reviewer azs9 Part1

Q1. Local Retrieval Scope

During validation and testing, Local Retrieval searches historical observations within the preceding WlocalW_{\mathrm{local}} time steps relative to each forecast origin tt. This may include previously observed data from the validation/test stream.

Importantly, all retrieved local patches are restricted to observations available before the forecast origin tt. No future observations or ground-truth values from the prediction horizon are accessed. Furthermore, the global retrieval memory is constructed exclusively from the training split.

Regarding DLoG's performance across all nine datasets when restricted to the same historical context as the baselines, please refer to our response to Reviewer gEFL under W2: Fairness of Historical Context Comparison.

Q2. Implementation Details and Parameter Accounting

Thank you for identifying these omissions. In the current implementation, the default patch length and stride are P=96P=96 and r=48r=48, respectively. The effective temporal dimension for global matching is D=P=96D=P=96. When the input look-back length LL exceeds PP, each memory key is truncated to its last PP observations before decomposition. Thus, each length-PP query is compared with a length-PP key representation, rather than directly with a length-LL key. For global matching, the channel and temporal dimensions are flattened, mean-centered, and ℓ2\ell_2-normalized before computing the dot product. For seasonal matching, shifts ranging from −3-3 to 33 are implemented as circular shifts, such that values crossing one boundary wrap around to the other.

Each future value window is extracted from data standardized using per-variable statistics fitted on the training split. Before decomposition, neither the last value of its associated key nor the first value of the future window is subtracted. After decomposition, the implementation constructs multiscale component representations and subtracts the last value of each resulting component representation at each scale. For Eq. (12), the seasonal amplitude and bias are computed using two independent affine mappings from the seasonal input context to the forecast horizon. Specifically,

a^{(\mathrm{se})} &= 1 + 0.5\\tanh\left(W_a x^{(\mathrm{se})}+b_a\right),\\ b^{(\mathrm{se})} &= W_b x^{(\mathrm{se})}+b_b. \end{aligned}$$ The amplitude mapping applies a bounded nonlinear transformation, while the bias mapping is applied directly. Both mappings are shared across channels. The fusion module contains two horizon-dependent linear projection layers. The first layer projects a 6H-dimensional input into an H-dimensional output, using an H × 6H weight matrix and an H-dimensional bias vector, resulting in 6H² + H trainable parameters. The second layer projects a 3H-dimensional input into an H-dimensional output, using an H × 3H weight matrix and an H-dimensional bias vector, resulting in 3H² + H trainable parameters. Therefore, the total number of additional trainable parameters introduced by the fusion module is 9H² + 2H. These parameters are exclusively associated with the fusion module; the backbone architecture and its parameters remain unchanged.
AuthorsOpenReview ↗

Response to Reviewer azs9 Part2

Q3. Clarification of TS-RAG and RAF Evaluation Settings

We thank the reviewer for pointing out the insufficient explanation of the experimental settings and evaluation metrics. We provide the following clarifications.

1. Comparison with TS-RAG (Table 4).

Following the experimental settings of TS-RAG, we additionally evaluated DLoG using a look-back length of L=512L=512 and a prediction horizon of H=64H=64, corresponding to the TS-RAG Chronos-based evaluation setting. Therefore, the DLoG results reported in Table 4 (0.350, 0.238, 0.285, 0.139, and 0.072) differ from those in Tables 1 and 14, which use L=96L=96 and H∈{96,192,336,720}H\in\{96,192,336,720\}.

We also recognize that TS-RAG adopts a zero-shot forecasting paradigm, whereas DLoG is trained in-domain. Thus, this comparison is intended to provide supplementary empirical context under the same input and prediction lengths, rather than claim a strictly equivalent training protocol.

2. Comparison with RAF (Table 8).

DLoG is a deterministic forecaster and produces a single point prediction rather than separate probabilistic quantiles. To calculate the value reported as WQL, we use the same point prediction y^j\hat{y}_j at each of the nine quantile levels q∈Q={0.1,0.2,…,0.9}q\in\mathcal{Q}=\{0.1,0.2,\ldots,0.9\}. We calculate the pinball loss for each quantile level and average it over all levels and forecast elements:

ej=yj−y^je_j=y_j-\hat{y}_j

WQLDLoG=19M∑q∈{0.1,0.2,…,0.9}∑j=1Mmax⁡(qej,(q−1)ej)WQL_{DLoG}=\frac{1}{9M}\sum_{q\in\{0.1,0.2,\ldots,0.9\}}\sum_{j=1}^{M}\max(qe_j,(q-1)e_j)

Here, MM is the number of forecast elements. The computation is performed on the standardized data scale and does not include an additional normalization by the sum of absolute target values. Because the same point forecast is used at every equally weighted quantile level, this quantity is equal to half the MAE:

WQLDLoG=0.5×MAE.\mathrm{WQL}_{\mathrm{DLoG}}=0.5\times\mathrm{MAE}.

Q4. Runtime Evaluation on High-Dimensional Datasets

Thank you for highlighting the scalability question. We evaluated the exact-retrieval implementation on the full training repositories of ECL (321 channels, 18,221 windows) and Traffic (862 channels, 12,089 windows), with a look-back length and prediction horizon of L=H=96L=H=96 and a batch size of 32. All experiments were conducted on an H200 NVL GPU with a 16-thread AMD EPYC 9654 CPU. The measured wall-clock runtimes are summarized below.

DatasetChannelsRepository Refresh / EpochOnline Retrieval / Batch
ECL32118.50 s0.87 s
Traffic86237.80 s1.92 s

Despite the increase in dimensionality from 321 to 862 channels (2.69×2.69\times), repository refresh time increased by approximately 2.04×2.04\times, while online retrieval latency increased by approximately 2.21×2.21\times. Notably, online retrieval remained below 2 seconds per batch on both datasets, and repository refresh completed within 40 seconds per epoch.

These results demonstrate that the current implementation can perform exact retrieval on high-dimensional datasets with manageable runtime overhead. They complement the ETTm1 results in Table 11 by providing additional wall-clock measurements at substantially higher dimensionalities. Since the repository sizes also differ, these measurements should be interpreted as empirical scalability evidence rather than a controlled complexity comparison.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.