acceptodds
Under review as a conference paper at ICLR 2027

GRAM: Correcting Frozen Time-Series Foundation Models via Graph-Retrieved Amplitude Memory

Abstract

Time-series foundation models (TSFMs) enable zero-shot forecasting through large-scale cross-domain pretraining, while retrieval augmentation further improves their performance by leveraging historical information. However, existing methods typically correct TSFM forecasts using the ground-truth futures of similar historical windows, which contain both predictive components already captured by the foundation model and sample-specific random fluctuation that is difficult to transfer. In contrast, recurring systematic model bias within prediction errors more directly characterizes the failure modes of a frozen TSFM and therefore provides more valuable correction signals. Effectively exploiting such model bias, however, poses two challenges: prediction errors at different numerical levels are difficult to compare due to scale differences, and the recurring bias must be extracted from prediction errors contaminated by random fluctuation. To address these challenges, we propose GRAM, a general retrieval-augmented framework for frozen TSFMs. GRAM first introduces an Amplitude Memory Module (AMM) that scales prediction errors by amplitude and aggregates them into retrievable prototypes. It then employs a Prototype Graph Module (PGM) to model relations among prototypes to aggregate consistent bias information while suppressing random fluctuation. During online forecasting, GRAM retrieves and expands prototypes relevant to the current query and generates per-horizon corrections to refine the original TSFM forecast. Experiments across multiple datasets and foundation models demonstrate consistent forecasting improvements.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZqsGOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a foundation model forecast correction framework based on prediction error retrieval. The framework comprises a core Amplitude Memory Module (AMM) and a Prototype Graph Module (PGM), which work together to generate per-horizon corrections aimed at refining the original predictions of the base foundation model.

Critical strengths

(1) Unlike conventional RAG methods that directly reuse raw historical ground-truth trajectories, the proposed method strategically targets prediction errors for retrieval. This design choice is insightful and well-motivated.

(2) The designed gating mechanism exhibits good robustness. When no useful bias evidence is retrieved, the model gracefully falls back to the original foundation model's prediction (a safe identity mapping), preventing harmful corrections.

(3) The authors extensively evaluated the proposed method across four distinct foundation models and seven datasets, demonstrating consistent overall performance improvements.

Critical weaknesses

(1) The framework explicitly models bias as an amplitude scaling error. Consequently, if the base model correctly captures the temporal pattern but suffers from temporal shifts (e.g., the predicted pattern is "early", "delayed", or "misaligned" along the time axis), the proposed amplitude-scaling mechanism seems fundamentally unequipped to directly handle or correct such phase offsets.

(2) As observed in the actual forecasting curves (e.g., Figure 3), although GRAM reduces the quantitative prediction error, the corrected curves largely retain the main trends and localized fluctuations of the original base model's predictions. Given this constrained behavior, have the authors considered applying GRAM as an enhancement plugin for supervised forecasting?

(3) The paper states that the Graph Fusion Module (GFM) is fitted once on the LOTSA corpus. Since the LOTSA dataset inherently contains noise, how do the authors ensure that the GFM learns a robust mapping rather than overfitting to the corpus noise? What criteria or validation strategies are used to determine that the GFM is robustly and sufficiently trained?

Minor Comments:

(1) The paper lacks evaluation on common, higher-dimensional multivariate datasets, such as the Traffic dataset, which are standard in the field.

(2) In the results tables, when the baseline and the proposed model achieve the exact same optimal result, only the proposed model is highlighted (marked in red). This formatting convention does not strictly align with academic norms; all best results should be equally highlighted.

Questions for authors

no additional question

5 comments
AuthorsOpenReview ↗

Response to Reviewer ZqsG (Part I)

Dear Reviewer ZqsG, we sincerely appreciate your attentive reading and your insightful suggestions, which have helped us broaden the evaluation.

W1: The framework explicitly models bias as an amplitude scaling error; can it handle temporal shifts of the forecast along the time axis?

Thank you for your valuable feedback on our work! Your suggestions have played a significant role in improving the quality of our paper.A frozen TSFM shows systematic errors of several kinds, and a temporal shift is one of them. GRAM records one amplitude scaling error at each forecast lead. Whether the predicted pattern is "early", "delayed" or "misaligned", the shift appears as a prediction error at every lead, and when such errors recur under similar patterns they are recorded and corrected lead by lead.

To verify this, we label as temporal-shift windows those test windows whose correlation with the ground truth increases markedly after the forecast is shifted in time, and evaluate them separately on each dataset. As Table 1 shows, averaged over H=96, 192, 336 and 720, GRAM is better than or equal to the frozen TSFM in all 28 pairs, reducing MSE by 2.71% and MAE by 1.39% on average, so the amplitude correction remains effective on temporal-shift windows. This experiment and the criterion for labelling such windows are added in Appendix B.5 in lines 1130–1165 of the revised paper‾\underline{\text{Appendix B.5 in lines 1130–1165 of the revised paper}} and Table 9 in lines 1097–1111 of the revised paper‾\underline{\text{Table 9 in lines 1097–1111 of the revised paper}}.

The prediction bias of a TSFM can also be represented by additive residuals instead of an amplitude scaling error. Compared with additive residuals, the amplitude scaling error is dimensionless, independent of the local numerical level of the series, and therefore transferable across windows. Table 2 compares GRAM with an additive residual correction under the same retrieval pipeline, with the frozen TSFM (Ori) as a reference, again averaged over the four horizons. Averaged over the 28 pairs, GRAM reduces MSE by 5.63% and MAE by 2.45% relative to the additive residual correction, and its MSE is no worse in all 28 pairs.

Table 1: Frozen TSFMs and GRAM on temporal-shift windows.

TSFMDatasetOriGRAM
MetricMSEMAEMSEMAE
SundialETTh10.4600.4620.4550.461
ETTh20.3520.3970.3510.393
ETTm10.5180.4040.5110.400
ETTm20.3620.3780.3450.367
ECL0.6050.4710.6010.469
Weather0.2540.3030.2450.295
Exchange0.4660.4230.3540.373
ChronosETTh10.5500.4770.5410.475
ETTh20.3900.4090.3880.409
ETTm10.5300.4030.5200.399
ETTm20.4230.3980.4170.396
ECL0.7290.4930.7280.493
Weather0.2410.2910.2370.288
Exchange0.3960.4300.3920.426
TimesFMETTh10.4310.4240.4260.423
ETTh20.3970.4200.3960.419
ETTm10.3960.3650.3930.363
ETTm20.4270.4070.4200.403
ECL0.7540.5140.7520.513
Weather0.2480.2940.2410.289
Exchange0.2530.3400.2530.339
MoiraiETTh10.5230.4650.5090.461
ETTh20.3630.3980.3420.386
ETTm10.6340.4510.6020.440
ETTm20.4090.3910.3810.382
ECL0.6760.4810.6660.479
Weather0.6620.5040.6410.497
Exchange0.1190.2390.1190.238
AuthorsOpenReview ↗

Response to Reviewer ZqsG (Part I)

Table 2: GRAM and the additive residual correction.

TSFMDatasetGRAMAdditive residualOri
MetricMSEMAEMSEMAEMSEMAE
SundialETTh10.3890.4080.395‾\underline{0.395}0.411‾\underline{0.411}0.3970.412
ETTh20.3390.3790.340‾\underline{0.340}0.382‾\underline{0.382}0.3430.383
ETTm10.3600.3670.365‾\underline{0.365}0.369‾\underline{0.369}0.3660.370
ETTm20.2640.3130.273‾\underline{0.273}0.319‾\underline{0.319}0.2750.321
ECL0.1470.2420.148‾\underline{0.148}0.2420.148‾\underline{0.148}0.243‾\underline{0.243}
Weather0.2360.2690.243‾\underline{0.243}0.274‾\underline{0.274}0.2440.275
Exchange0.4390.4400.554‾\underline{0.554}0.494‾\underline{0.494}0.5580.499
ChronosETTh10.4360.4140.446‾\underline{0.446}0.417‾\underline{0.417}0.4480.418
ETTh20.3620.3770.366‾\underline{0.366}0.379‾\underline{0.379}0.3670.380
ETTm10.4110.3790.419‾\underline{0.419}0.382‾\underline{0.382}0.4210.383
ETTm20.2870.3150.290‾\underline{0.290}0.317‾\underline{0.317}0.2910.317‾\underline{0.317}
ECL0.1590.2440.160‾\underline{0.160}0.245‾\underline{0.245}0.1590.244
Weather0.2570.2730.262‾\underline{0.262}0.275‾\underline{0.275}0.2630.276
Exchange0.3580.4020.368‾\underline{0.368}0.411‾\underline{0.411}0.3710.412
TimesFMETTh10.4170.4070.420‾\underline{0.420}0.408‾\underline{0.408}0.4210.409
ETTh20.3690.3850.371‾\underline{0.371}0.387‾\underline{0.387}0.3720.388
ETTm10.3780.3700.380‾\underline{0.380}0.3700.3810.372‾\underline{0.372}
ETTm20.2850.3150.290‾\underline{0.290}0.317‾\underline{0.317}0.2910.318
ECL0.1640.2490.1640.2490.1640.249
Weather0.2530.2640.261‾\underline{0.261}0.267‾\underline{0.267}0.2630.268
Exchange0.3790.4100.401‾\underline{0.401}0.418‾\underline{0.418}0.4040.420
MoiraiETTh10.5300.4730.600‾\underline{0.600}0.496‾\underline{0.496}0.6200.504
ETTh20.4050.4150.443‾\underline{0.443}0.433‾\underline{0.433}0.4510.438
ETTm10.6170.4850.727‾\underline{0.727}0.512‾\underline{0.512}0.7400.523
ETTm20.3440.3600.403‾\underline{0.403}0.381‾\underline{0.381}0.4040.388
ECL0.3910.4070.499‾\underline{0.499}0.442‾\underline{0.442}0.5200.446
Weather0.3150.3100.360‾\underline{0.360}0.327‾\underline{0.327}0.3740.333
Exchange0.4630.4430.573‾\underline{0.573}0.485‾\underline{0.485}0.6010.488
AuthorsOpenReview ↗

Response to Reviewer ZqsG (Part II)

W2: Have the authors considered applying GRAM as an enhancement plugin for supervised forecasting?

Thank you for your suggestions! GRAM only requires a frozen forecaster: it records that forecaster's prediction bias and adds the correction to its forecast, so it can be attached directly to supervised models. We attach GRAM as an enhancement plugin to MixLinear [1], xPatch [2], DUET [3], PDF [4], SparseTSF [5] and Pathformer [6], each trained with its official code and officially recommended hyperparameters on the training split of each dataset. The table reports H=96: over the 36 settings, GRAM reduces MSE by 0.95% and MAE by 0.80% on average, and improves the original forecaster in the large majority of settings. Even after supervised training, the backbone still shows recurring pattern-conditioned scaling biases on certain local patterns; GRAM captures and corrects these systematic residual errors without changing any parameter of the backbone. Only H=96 is reported here; the trend is consistent at longer horizons.

[1] MixLinear: Extreme Low Resource Multivariate Time Series Forecasting with 0.1K Parameters. ICLR 2026.

[2] xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend Decomposition. AAAI 2025.

[3] DUET: Dual Clustering Enhanced Multivariate Time Series Forecasting. KDD 2025.

[4] Periodicity Decoupling Framework for Long-term Series Forecasting. ICLR 2024.

[5] SparseTSF: Modeling Long-term Time Series Forecasting with 1k Parameters. ICML 2024.

[6] Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting. ICLR 2024.

ModelDatasetOri+GRAM
MetricMSEMAEMSEMAE
MixLinearETTh10.3510.3830.3510.383
ETTh20.2830.3400.2780.336
ETTm10.3330.3750.3270.370
ETTm20.1800.2710.1760.266
Weather0.1790.2370.1680.225
Exchange0.0840.2030.0830.202
xPatchETTh10.3660.3920.3700.393
ETTh20.2750.3340.2740.332
ETTm10.2890.3360.2890.336
ETTm20.1600.2460.1600.244
Weather0.1440.1830.1430.182
Exchange0.0840.2000.0820.199
DUETETTh10.3520.3840.3530.384
ETTh20.2730.3370.2710.336
ETTm10.2780.3330.2760.330
ETTm20.1610.2490.1610.247
Weather0.1480.1930.1470.192
Exchange0.0790.1970.0790.197
PDFETTh10.3600.3910.3600.390
ETTh20.2770.3410.2730.337
ETTm10.2860.3400.2860.339
ETTm20.1630.2510.1620.249
Weather0.1470.1970.1460.195
Exchange0.0840.2010.0830.200
SparseTSFETTh10.3600.3860.3600.386
ETTh20.2730.3340.2700.331
ETTm10.3050.3460.3020.344
ETTm20.1650.2550.1640.252
Weather0.1730.2250.1640.216
Exchange0.0840.2010.0830.200
PathformerETTh10.3720.3920.3720.392
ETTh20.2770.3360.2740.333
ETTm10.2900.3350.2890.334
ETTm20.1580.2460.1570.244
Weather0.1460.1900.1450.189
Exchange0.0880.2080.0870.207
Reviewer_NwVQOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the issue of zero-shot prediction correction for frozen TSFMs (Time-Series Foundation Models), this paper proposes a framework based on amplitude memory and prototype graph retrieval. The core idea is to shift the retrieval target from the "actual future trajectory" to "recurrent model biases within the prediction residuals," while employing amplitude-scaling errors to enable scale-invariant comparisons across different time windows.

Critical strengths

1、Existing retrieval-augmented methods directly reuse the actual future outcomes from similar windows, which introduces stochastic fluctuations specific to individual examples. This paper shifts the retrieval objective toward predicting the recurring bias within the residuals—a well-motivated approach.

2、Offers high efficiency; requires training only a lightweight GFM, entails minimal memory overhead, and holds potential for practical deployment.

Critical weaknesses

1、GFM is trained on LOTSA; however, the relationship between LOTSA and the evaluation datasets is not sufficiently clarified. Given that LOTSA serves as the pre-training corpus for Moirai, it remains unclear whether there is any overlap with the evaluation datasets, such as those for Weather, Electricity, and ETT.

2、In the graph fusion module, how exactly are Q, K, and V calculated within the MHA (Multi-Head Attention) operation? Given that each node contains five elements, which specific element is used to derive K and V?

3、The results from Table 1 and Table 2 can be consolidated. Fundamentally, this paper aims to demonstrate that the proposed plug-in is both efficient (low-overhead) and effective; therefore, comparing the "original base model + GRAM" against a fine-tuned version of the original base model is more meaningful.

4、The main text lacks a direct comparison with other RAG strategies found in related work. While Table 7 in the appendix provides a comparison with TS-RAG, this is far from sufficient. Furthermore, it is worth considering why the gains achieved by GRAM on the Exchange dataset are significantly smaller than those of TS-RAG. Given the non-stationary nature of the Exchange dataset, could it be that the GRAM method is constrained by such data, thereby revealing certain limitations?

5、This paper involves numerous hyperparameters—such as the number of prototypes CC, the number of retrieved items KK, and the number of neighbors LL—which significantly influence model performance. However, the absence of ablation studies regarding these components diminishes the completeness of the paper.

6、This text contains a large number of scattered symbols and formulas, which hinders readability—for instance, regarding the composition of node elements and the functions of various components within the gating networks. It is recommended to use pseudocode to improve clarity.

Questions for authors

See Weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer NwVQ (Part I)

Dear Reviewer NwVQ, we are deeply grateful for your careful and insightful review, which has helped us strengthen the paper.

W1: The relationship between LOTSA and the evaluation datasets is not sufficiently clarified, i.e., whether there is any overlap.

Thank you for your valuable feedback! The original LOTSA paper [1] states that the seven evaluation datasets used here are not part of the LOTSA dataset list, so the fitting corpus of the GFM does not overlap with our evaluation datasets. The amplitude memory and the prototype graph are built only from the historical observations of each target dataset, and the GFM is a generic read-out fitted once and then frozen; even if a boundary case exists in the pre-training corpus, no future information of the evaluation sets is leaked. The list of pretraining sub-corpora we select is given in Table 5 in lines 808–822 of the revised paper‾\underline{\text{Table 5 in lines 808–822 of the revised paper}} (Appendix A.3).

To further rule out any effect of the training corpus, we refit the GFM on the UTSD corpus released with Timer [2], using exactly the same configuration as in the paper; this subset contains none of the evaluation datasets. Averaged over H=96, 192, 336 and 720, the GFM trained on UTSD still lets GRAM reduce MSE by 6.24% and MAE by 2.77% relative to the frozen backbones, improving both metrics in 27 of the 28 dataset-backbone pairs (the remaining one, Sundial on ECL, is tied). The table below also lists the LOTSA-trained version of GRAM used in the paper as a reference. The two read-outs differ by 0.002 in average MSE and 0.002 in average MAE over the 28 pairs, with maximum differences of 0.006 and 0.007, so the results are essentially the same. Our method therefore does not depend on one particular training corpus.

[1] Unified Training of Universal Time Series Forecasting Transformers. ICML 2024.

[2] Timer: Generative Pre-trained Transformers Are Large Time Series Models. ICML 2024.

TSFMDatasetOriGRAM (UTSD)GRAM (LOTSA)
MetricMSEMAEMSEMAEMSEMAE
SundialETTh10.3970.4120.391‾\underline{0.391}0.410‾\underline{0.410}0.3890.408
ETTh20.3430.3830.340‾\underline{0.340}0.380‾\underline{0.380}0.3390.379
ETTm10.3660.3700.361‾\underline{0.361}0.368‾\underline{0.368}0.3600.367
ETTm20.2750.3210.2630.3120.264‾\underline{0.264}0.313‾\underline{0.313}
ECL0.148‾\underline{0.148}0.243‾\underline{0.243}0.148‾\underline{0.148}0.243‾\underline{0.243}0.1470.242
Weather0.2440.2750.2350.2680.236‾\underline{0.236}0.269‾\underline{0.269}
Exchange0.5580.4990.4380.4390.439‾\underline{0.439}0.440‾\underline{0.440}
ChronosETTh10.4480.4180.437‾\underline{0.437}0.415‾\underline{0.415}0.4360.414
ETTh20.3670.3800.363‾\underline{0.363}0.378‾\underline{0.378}0.3620.377
ETTm10.4210.3830.412‾\underline{0.412}0.380‾\underline{0.380}0.4110.379
ETTm20.2910.3170.288‾\underline{0.288}0.316‾\underline{0.316}0.2870.315
ECL0.159‾\underline{0.159}0.244‾\underline{0.244}0.1580.2430.159‾\underline{0.159}0.244‾\underline{0.244}
Weather0.2630.2760.258‾\underline{0.258}0.274‾\underline{0.274}0.2570.273
Exchange0.3710.4120.364‾\underline{0.364}0.408‾\underline{0.408}0.3580.402
TimesFMETTh10.4210.4090.4160.4040.417‾\underline{0.417}0.407‾\underline{0.407}
ETTh20.372‾\underline{0.372}0.3880.3690.3840.3690.385‾\underline{0.385}
ETTm10.3810.3720.3760.3680.378‾\underline{0.378}0.370‾\underline{0.370}
ETTm20.2910.3180.290‾\underline{0.290}0.317‾\underline{0.317}0.2850.315
ECL0.164‾\underline{0.164}0.249‾\underline{0.249}0.1630.2480.164‾\underline{0.164}0.249‾\underline{0.249}
Weather0.2630.2680.2510.2620.253‾\underline{0.253}0.264‾\underline{0.264}
Exchange0.4040.4200.384‾\underline{0.384}0.417‾\underline{0.417}0.3790.410
MoiraiETTh10.620‾\underline{0.620}0.5040.5300.475‾\underline{0.475}0.5300.473
ETTh20.4510.4380.409‾\underline{0.409}0.422‾\underline{0.422}0.4050.415
ETTm10.7400.5230.622‾\underline{0.622}0.489‾\underline{0.489}0.6170.485
ETTm20.4040.3880.3430.3570.344‾\underline{0.344}0.360‾\underline{0.360}
ECL0.5200.4460.396‾\underline{0.396}0.410‾\underline{0.410}0.3910.407
Weather0.3740.3330.320‾\underline{0.320}0.317‾\underline{0.317}0.3150.310
Exchange0.6010.4880.4610.4390.463‾\underline{0.463}0.443‾\underline{0.443}
AuthorsOpenReview ↗

Response to Reviewer NwVQ (Part II)

W2: How exactly are Q, K, and V calculated within the MHA operation, and which of the five elements of each node is used to derive K and V?

Thank you for your question! In the MHA, Q, K and V are obtained by linear projections of different inputs.

Q comes from the query window. The query feature is formed by concatenating the scale-normalized excursion excq=y^qtsfm−mq\mathrm{exc}_q=\widehat{\mathbf y}_q^{\mathrm{tsfm}}-m_q with its log scale, standard deviation and maximum, and is encoded by a two-layer MLP into the query vector cq\mathbf c_q (Section 3.3.2):

Q=cqWQ.{Q=\mathbf c_qW_Q.}

K and V come from the subgraph nodes. Each node corresponds to one five-element record; the features fed to the GNN are the amplitude curve α⃗c\vec\alpha_c, the spread curve δ⃗c\vec\delta_c, the coherence ρc\rho_c and the support ncn_c, together with the retrieval similarity simq,c\mathrm{sim}_{q,c}. The GNN block first encodes these node inputs with a shared two-layer MLP, then performs two layers of relation-aware message passing along the shape and bias edges (Section 3.3.2). Writing this block as GNN(⋅)\mathrm{GNN}(\cdot), K and V are projected from the resulting node representation hc\mathbf h_c:

K_c=\mathbf h_cW_K,\qquad V_c=\mathbf h_cW_V.$$ K and V are projected from the same node representation and differ only in the projection matrix. All notation is consistent with Section 3 of the paper; the node inputs and the computation of Q, K and V are now given in $\underline{\text{Section 3.3.2 in lines 286–291 of the revised paper}}$. **W3: The results from Table 1 and Table 2 can be consolidated, and comparing against a fine-tuned version of the original base model is more meaningful.** We thank you for this valuable suggestion. Table 1 and Table 2 concern two different ways of improving a frozen TSFM: retrieval augmentation keeps the backbone frozen and only corrects its forecast, whereas fine-tuning updates the model parameters. The table below places the four settings (Ori, Ori+GRAM, FT and FT+GRAM) side by side for a direct comparison. Reading Table 1 and Table 2 together, GRAM averages 0.348 / 0.361 MSE / MAE over the 28 pairs, better than the 0.356 / 0.363 of FT. Without using any target-domain labels, a read-out fitted once on an unrelated corpus already outperforms per-dataset fine-tuning on average. The comparison is not strictly symmetric: GRAM performs no gradient update on any target dataset, whereas FT is fine-tuned separately on each target dataset and uses target-domain labels; prior work also reports catastrophic forgetting after fine-tuning on a specialized dataset [1]. In terms of efficiency, the computation and memory overhead of GRAM is very low: the entire correction costs only 7.7-12.2 $\mu$s per window ($\underline{\text{Table 8 in lines 1026–1034 of the original paper}}$; see also $\underline{\text{Appendix A.3 in lines 789–806 of the original paper}}$), whereas FT requires a separate training run for every target dataset. GRAM can also be combined with fine-tuning. Attaching GRAM to a fine-tuned TSFM improves it further: Moirai at H=720 improves from 1.077 to 0.753 MSE, a 30.1% reduction. GRAM is thus both a low-cost alternative to fine-tuning and a plug-in that can be stacked on a fine-tuned model. [1] MEMTS: Internalizing Domain Knowledge via Parameterized Memory for Retrieval-Free Domain Adaptation of Time Series Foundation Models. KDD 2026.
AuthorsOpenReview ↗

Response to Reviewer NwVQ (Part II)

TSFMDatasetOriOri+GRAMFTFT+GRAM
MetricMSEMAEMSEMAEMSEMAEMSEMAE
SundialETTh10.397‾\underline{0.397}0.412‾\underline{0.412}0.3890.4080.4030.4150.4110.417
ETTh20.3430.3830.3390.3790.338‾\underline{0.338}0.376‾\underline{0.376}0.3350.374
ETTm10.3660.3700.3600.3670.351‾\underline{0.351}0.362‾\underline{0.362}0.3430.361
ETTm20.2750.3210.264‾\underline{0.264}0.313‾\underline{0.313}0.2740.3190.2620.311
ECL0.1480.243‾\underline{0.243}0.147‾\underline{0.147}0.2420.1460.2420.1460.242
Weather0.2440.275‾\underline{0.275}0.236‾\underline{0.236}0.2690.2490.2790.2350.269
Exchange0.5580.4990.439‾\underline{0.439}0.440‾\underline{0.440}0.4960.4620.4320.432
ChronosETTh10.448‾\underline{0.448}0.418‾\underline{0.418}0.4360.4140.4610.4210.4520.422
ETTh20.367‾\underline{0.367}0.380‾\underline{0.380}0.3620.3770.3750.3840.3700.383
ETTm10.4210.3830.4110.3790.349‾\underline{0.349}0.354‾\underline{0.354}0.3410.352
ETTm20.2910.3170.2870.3150.275‾\underline{0.275}0.308‾\underline{0.308}0.2740.307
ECL0.159‾\underline{0.159}0.244‾\underline{0.244}0.159‾\underline{0.159}0.244‾\underline{0.244}0.1570.2430.1570.243
Weather0.2630.2760.2570.2730.236‾\underline{0.236}0.247‾\underline{0.247}0.2290.245
Exchange0.371‾\underline{0.371}0.4120.3580.4020.3960.4200.3750.407‾\underline{0.407}
TimesFMETTh10.421‾\underline{0.421}0.409‾\underline{0.409}0.4170.4070.4890.4380.4890.438
ETTh20.3720.3880.3690.3850.366‾\underline{0.366}0.3880.3650.387‾\underline{0.387}
ETTm10.3810.3720.3780.3700.304‾\underline{0.304}0.355‾\underline{0.355}0.3030.354
ETTm20.2910.3180.2850.315‾\underline{0.315}0.269‾\underline{0.269}0.3050.2680.305
ECL0.164‾\underline{0.164}0.249‾\underline{0.249}0.164‾\underline{0.164}0.249‾\underline{0.249}0.1590.2460.1590.246
Weather0.2630.2680.2530.264‾\underline{0.264}0.229‾\underline{0.229}0.2500.2280.250
Exchange0.4040.420‾\underline{0.420}0.3790.4100.400‾\underline{0.400}0.4230.4260.435
MoiraiETTh10.6200.5040.530‾\underline{0.530}0.473‾\underline{0.473}0.5940.4930.5050.462
ETTh20.4510.4380.405‾\underline{0.405}0.415‾\underline{0.415}0.4320.4220.3940.404
ETTm10.7400.5230.6170.4850.566‾\underline{0.566}0.451‾\underline{0.451}0.4770.422
ETTm20.4040.3880.3440.3600.318‾\underline{0.318}0.342‾\underline{0.342}0.2930.329
ECL0.5200.4460.391‾\underline{0.391}0.407‾\underline{0.407}0.4670.4300.3760.402
Weather0.3740.3330.3150.3100.280‾\underline{0.280}0.288‾\underline{0.288}0.2560.275
Exchange0.6010.4880.4630.4430.5990.4870.470‾\underline{0.470}0.445‾\underline{0.445}
Avg0.3810.3740.348‾\underline{0.348}0.361‾\underline{0.361}0.3560.3620.3350.354
Reviewer_4ZDsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GRAM aims to improve the forecasting performance of frozen time-series foundation models by identifying recurring prediction errors in historical data and using them to correct future forecasts. It represents these errors through amplitude-based scaling, making them comparable across different time-series windows. The method is evaluated on seven benchmark datasets, four forecasting horizons, and four foundation models, showing improvements in both MSE and MAE across all datasets. The reported maximum reductions are 34.5% in MSE and 16.7% in MAE, with an additional computational cost of only 7.7–12.2 microseconds per window.

Critical strengths

1- The idea of retrieving model-specific forecasting errors rather than historical future trajectories is interesting and well motivated. Since the base TSFM already generates predictions, using historical residual errors to refine these forecasts provides a reasonable way to avoid reusing information that the model has already captured.

2- The formulation distinguishes the magnitude of the predicted change from the relative error associated with it. This allows the method to correct the frozen model’s forecast by adjusting its predicted deviation, rather than generating an entirely new prediction.

3- The approach is evaluated on four different TSFMs, demonstrating its applicability across multiple foundation models. The evaluation is further extended to LoRA-adapted models to examine its effectiveness in parameter-efficient fine-tuning settings.

Critical weaknesses

1- The prototype graph is one of the main contributions of the paper, based on the idea that similar historical patterns may share common forecasting biases. While this motivation is reasonable, the current results do not clearly show how much improvement comes specifically from the graph component rather than simply combining residual information from multiple prototypes. Although Table 3 and Section 4.2 provide relevant ablation studies, a more direct comparison that separates the contribution of graph propagation from the learned fusion mechanism would better justify the importance of the proposed graph structure.

2- The proposed amplitude-based correction mainly focuses on adjusting the magnitude of predicted changes. However, not all forecasting errors can be addressed through scaling, as some may result from systematic shifts or biases. For example, when the predicted change is very small or close to zero, the correction may not work effectively even if the model has a consistent prediction error. It would be helpful if the authors could clarify how the approach handles such cases and provide a comparison with a simple additive residual-correction method to better demonstrate the advantage of the proposed formulation.

3- The main experiments compare the original frozen TSFMs with their GRAM-enhanced versions, showing that the proposed approach can improve forecasting performance across different models. However, the paper also emphasizes its advantages over existing retrieval-based methods, such as TS-RAG and TimeRAF, which use historical future trajectories for forecasting. Although Appendix B.4 compares GRAM with TS-RAG, the methods are evaluated using different foundation models, making it difficult to judge whether the improvements come from the proposed retrieval strategy or the underlying model. The results in Table 7 are also mixed. GRAM achieves lower MSE on six datasets, while TS-RAG performs better in MAE on two datasets and achieves a considerably lower MSE on Exchange (0.332 vs. 0.439). I think a fairer comparison using the same frozen TSFM and experimental settings would strengthen the evaluation. In particular, comparing future-trajectory retrieval, direct residual retrieval, and the proposed amplitude-based graph retrieval would help better demonstrate the actual contribution of the proposed approach.

Questions for authors

Please address my all concerns raised in the weaknesses section.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 4ZDs (Part I)

Dear Reviewer 4ZDs, we are deeply grateful for your exceptionally thorough review and for the constructive suggestions that helped us sharpen the evaluation.

W1: The current results do not clearly show how much improvement comes specifically from the graph component rather than from the learned fusion mechanism.

Thank you for your valuable feedback! We add a direct comparison between the complete GRAM (+GRAM) and a variant without graph propagation (w/o prop.), so that the contribution of graph propagation is isolated. The two settings share the same prototype memory, test windows and labels, and have identical fusion, attention, gate and GFM parameters; the only difference is whether nodes exchange information along the prototype graph. In the complete method, node representations are propagated along the shape and bias edges, whereas each node of w/o prop. is read out from its own record alone. The table below reports four TSFMs on seven datasets, averaged over the four horizons, with the frozen backbone Ori listed as a reference; the best value in each row is bold and the second best underlined.

The complete GRAM outperforms w/o prop. throughout, confirming the effectiveness of graph propagation. Over all 28 pairs, it reduces MSE by 2.21% and MAE by 1.59% on average. The gain follows the error of the backbone: Moirai, the weakest frozen model, benefits most (4.77% MSE on average; on ECL, 0.430 / 0.422 improves to 0.391 / 0.407), while TimesFM, the strongest, still improves by 0.93% MSE. Bias evidence recorded on related prototypes is thus pooled along the graph, providing information beyond the fusion mechanism.

TSFMDatasetOriw/o prop.GRAM
MetricMSEMAEMSEMAEMSEMAE
SundialETTh10.3970.4120.394‾\underline{0.394}0.411‾\underline{0.411}0.3890.408
ETTh20.343‾\underline{0.343}0.3830.343‾\underline{0.343}0.382‾\underline{0.382}0.3390.379
ETTm10.3660.3700.362‾\underline{0.362}0.369‾\underline{0.369}0.3600.367
ETTm20.2750.3210.267‾\underline{0.267}0.316‾\underline{0.316}0.2640.313
ECL0.148‾\underline{0.148}0.243‾\underline{0.243}0.148‾\underline{0.148}0.2440.1470.242
Weather0.2440.2750.240‾\underline{0.240}0.273‾\underline{0.273}0.2360.269
Exchange0.5580.4990.476‾\underline{0.476}0.464‾\underline{0.464}0.4390.440
ChronosETTh10.4480.4180.442‾\underline{0.442}0.417‾\underline{0.417}0.4360.414
ETTh20.3670.3800.364‾\underline{0.364}0.379‾\underline{0.379}0.3620.377
ETTm10.4210.3830.417‾\underline{0.417}0.382‾\underline{0.382}0.4110.379
ETTm20.2910.317‾\underline{0.317}0.290‾\underline{0.290}0.317‾\underline{0.317}0.2870.315
ECL0.1590.2440.160‾\underline{0.160}0.246‾\underline{0.246}0.1590.244
Weather0.2630.276‾\underline{0.276}0.260‾\underline{0.260}0.276‾\underline{0.276}0.2570.273
Exchange0.3710.4120.364‾\underline{0.364}0.406‾\underline{0.406}0.3580.402
TimesFMETTh10.4210.409‾\underline{0.409}0.420‾\underline{0.420}0.4100.4170.407
ETTh20.3720.3880.371‾\underline{0.371}0.387‾\underline{0.387}0.3690.385
ETTm10.3810.372‾\underline{0.372}0.380‾\underline{0.380}0.372‾\underline{0.372}0.3780.370
ETTm20.2910.3180.289‾\underline{0.289}0.317‾\underline{0.317}0.2850.315
ECL0.1640.2490.165‾\underline{0.165}0.250‾\underline{0.250}0.1640.249
Weather0.2630.268‾\underline{0.268}0.256‾\underline{0.256}0.268‾\underline{0.268}0.2530.264
Exchange0.4040.4200.385‾\underline{0.385}0.415‾\underline{0.415}0.3790.410
MoiraiETTh10.6200.5040.565‾\underline{0.565}0.487‾\underline{0.487}0.5300.473
ETTh20.4510.4380.414‾\underline{0.414}0.424‾\underline{0.424}0.4050.415
ETTm10.7400.5230.638‾\underline{0.638}0.503‾\underline{0.503}0.6170.485
ETTm20.4040.3880.356‾\underline{0.356}0.373‾\underline{0.373}0.3440.360
ECL0.5200.4460.430‾\underline{0.430}0.422‾\underline{0.422}0.3910.407
Weather0.3740.3330.328‾\underline{0.328}0.320‾\underline{0.320}0.3150.310
Exchange0.6010.4880.489‾\underline{0.489}0.462‾\underline{0.462}0.4630.443
AuthorsOpenReview ↗

Response to Reviewer 4ZDs (Part II)

W2: Can the amplitude-based correction handle systematic shifts and predicted changes close to zero, and how does it compare with a simple additive residual-correction method?

We thank you for the question. The near-zero case is handled explicitly in the design. As stated in Section 3.2.2 in lines 230–231 of the original paper‾\underline{\text{Section 3.2.2 in lines 230–231 of the original paper}}, when the memory is built, the amplitudes of the member windows are aggregated by a weighted median whose weights are the absolute excursions. Windows whose excursion is close to zero receive very small weights, so their numerically unstable ratios do not enter the memory. When the predicted excursion is close to zero, the ratio is numerically unstable and the weighted median automatically down-weights those windows, keeping noisy corrections out of the memory.

Second, to test whether an additive correction is better suited to systematic shifts, we add a controlled additive residual-correction baseline. Under exactly the same retrieval pipeline, we only replace the quantity stored in each prototype with the mean of the member-window residuals and add the retrieved residual directly to the forecast. The frozen backbone Ori is listed as a reference. The table averages over horizons 96, 192, 336 and 720. Averaged over the 28 pairs, GRAM reduces MSE by 5.63% and MAE by 2.45% relative to the additive residual correction.

TSFMDatasetGRAMAdditive residualOri
MetricMSEMAEMSEMAEMSEMAE
SundialETTh10.3890.4080.395‾\underline{0.395}0.411‾\underline{0.411}0.3970.412
ETTh20.3390.3790.340‾\underline{0.340}0.382‾\underline{0.382}0.3430.383
ETTm10.3600.3670.365‾\underline{0.365}0.369‾\underline{0.369}0.3660.370
ETTm20.2640.3130.273‾\underline{0.273}0.319‾\underline{0.319}0.2750.321
ECL0.1470.2420.148‾\underline{0.148}0.2420.148‾\underline{0.148}0.243‾\underline{0.243}
Weather0.2360.2690.243‾\underline{0.243}0.274‾\underline{0.274}0.2440.275
Exchange0.4390.4400.554‾\underline{0.554}0.494‾\underline{0.494}0.5580.499
ChronosETTh10.4360.4140.446‾\underline{0.446}0.417‾\underline{0.417}0.4480.418
ETTh20.3620.3770.366‾\underline{0.366}0.379‾\underline{0.379}0.3670.380
ETTm10.4110.3790.419‾\underline{0.419}0.382‾\underline{0.382}0.4210.383
ETTm20.2870.3150.290‾\underline{0.290}0.317‾\underline{0.317}0.2910.317‾\underline{0.317}
ECL0.1590.2440.160‾\underline{0.160}0.245‾\underline{0.245}0.1590.244
Weather0.2570.2730.262‾\underline{0.262}0.275‾\underline{0.275}0.2630.276
Exchange0.3580.4020.368‾\underline{0.368}0.411‾\underline{0.411}0.3710.412
TimesFMETTh10.4170.4070.420‾\underline{0.420}0.408‾\underline{0.408}0.4210.409
ETTh20.3690.3850.371‾\underline{0.371}0.387‾\underline{0.387}0.3720.388
ETTm10.3780.3700.380‾\underline{0.380}0.3700.3810.372‾\underline{0.372}
ETTm20.2850.3150.290‾\underline{0.290}0.317‾\underline{0.317}0.2910.318
ECL0.1640.2490.1640.2490.1640.249
Weather0.2530.2640.261‾\underline{0.261}0.267‾\underline{0.267}0.2630.268
Exchange0.3790.4100.401‾\underline{0.401}0.418‾\underline{0.418}0.4040.420
MoiraiETTh10.5300.4730.600‾\underline{0.600}0.496‾\underline{0.496}0.6200.504
ETTh20.4050.4150.443‾\underline{0.443}0.433‾\underline{0.433}0.4510.438
ETTm10.6170.4850.727‾\underline{0.727}0.512‾\underline{0.512}0.7400.523
ETTm20.3440.3600.403‾\underline{0.403}0.381‾\underline{0.381}0.4040.388
ECL0.3910.4070.499‾\underline{0.499}0.442‾\underline{0.442}0.5200.446
Weather0.3150.3100.360‾\underline{0.360}0.327‾\underline{0.327}0.3740.333
Exchange0.4630.4430.573‾\underline{0.573}0.485‾\underline{0.485}0.6010.488
AuthorsOpenReview ↗

Response to Reviewer 4ZDs (Part III)

W3: A fairer comparison using the same frozen TSFM: future-trajectory retrieval, direct residual retrieval and the proposed amplitude-based graph retrieval, as well as existing RAG methods.

We thank you for this valuable suggestion, which has helped us improve the paper.We agree that a comparison under the same backbone is more convincing, and add two sets of experiments. All new results use the same test windows, labels and preprocessing, with the read-out and the TSFM frozen throughout.

(1) Comparison with existing RAG methods. Taking Sundial as the unified backbone, we compare GRAM with TS-RAG [1], TimeRAF [2] and SARAF [3]. Table 1 averages over the four horizons, Ori is the frozen Sundial, and the best and second best values in each row are bold and underlined. GRAM is better than all existing RAG methods, confirming the value of targeting model bias for retrieval. Averaged over the seven datasets, it reduces MSE by 2.4% and MAE by 1.3% relative to the three RAG baselines. Details are in Table 8 in lines 1080–1095 of the revised paper‾\underline{\text{Table 8 in lines 1080–1095 of the revised paper}} and Appendix B.4 in lines 1114–1126 of the revised paper‾\underline{\text{Appendix B.4 in lines 1114–1126 of the revised paper}}.

(2) Stored quantity under the same retrieval pipeline. The frozen TSFM, encoder, prototype partition and retrieval pipeline are all unchanged; we only replace what the memory stores: the dimensionless amplitude scaling error of GRAM, the direct residuals of the member windows, and the normalized ground-truth future trajectories. The frozen backbone Ori is listed as a reference. Table 2 averages over the four horizons, with the best value bold and the second best underlined. GRAM is no worse than future-trajectory retrieval in all 56 cells, reducing MSE by 5.73% and MAE by 2.58% on average over the 28 pairs; relative to direct residual retrieval, it reduces MSE by 5.63% and MAE by 2.45%. A dimensionless amplitude therefore transfers more readily between windows at different numerical levels.

[1] TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster. NeurIPS 2025.

[2] TimeRAF: Retrieval-Augmented Foundation Model for Zero-Shot Time Series Forecasting. IEEE TKDE 2025.

[3] Stationarity-Aware Retrieval-Augmented Time Series Forecasting. KDD 2026.

Table 1: Fair comparison with existing RAG methods.

DatasetGRAMTS-RAGTimeRAFSARAFOri
MetricMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
ETTh10.3890.4080.390‾\underline{0.390}0.4110.3940.4100.3950.409‾\underline{0.409}0.3970.412
ETTh20.3390.3790.340‾\underline{0.340}0.3820.3420.3810.3410.380‾\underline{0.380}0.3430.383
ETTm10.3600.3670.362‾\underline{0.362}0.368‾\underline{0.368}0.3640.3690.3650.3690.3660.370
ETTm20.2640.3130.265‾\underline{0.265}0.3190.2730.314‾\underline{0.314}0.2720.3150.2750.321
ECL0.1470.2420.148‾\underline{0.148}0.243‾\underline{0.243}0.1490.2450.1510.2490.148‾\underline{0.148}0.243‾\underline{0.243}
Weather0.2360.2690.2390.271‾\underline{0.271}0.237‾\underline{0.237}0.2740.2380.2720.2440.275
Exchange0.4390.4400.461‾\underline{0.461}0.450‾\underline{0.450}0.4790.4620.5180.4600.5580.499

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.