- 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
Response to Reviewer ZqsG (Part I)
Table 2: GRAM and the additive residual correction.
| TSFM | Dataset | GRAM | Additive residual | Ori | |||
|---|---|---|---|---|---|---|---|
| Metric | MSE | MAE | MSE | MAE | MSE | MAE | |
| Sundial | ETTh1 | 0.389 | 0.408 | 0.397 | 0.412 | ||
| ETTh2 | 0.339 | 0.379 | 0.343 | 0.383 | |||
| ETTm1 | 0.360 | 0.367 | 0.366 | 0.370 | |||
| ETTm2 | 0.264 | 0.313 | 0.275 | 0.321 | |||
| ECL | 0.147 | 0.242 | 0.242 | ||||
| Weather | 0.236 | 0.269 | 0.244 | 0.275 | |||
| Exchange | 0.439 | 0.440 | 0.558 | 0.499 | |||
| Chronos | ETTh1 | 0.436 | 0.414 | 0.448 | 0.418 | ||
| ETTh2 | 0.362 | 0.377 | 0.367 | 0.380 | |||
| ETTm1 | 0.411 | 0.379 | 0.421 | 0.383 | |||
| ETTm2 | 0.287 | 0.315 | 0.291 | ||||
| ECL | 0.159 | 0.244 | 0.159 | 0.244 | |||
| Weather | 0.257 | 0.273 | 0.263 | 0.276 | |||
| Exchange | 0.358 | 0.402 | 0.371 | 0.412 | |||
| TimesFM | ETTh1 | 0.417 | 0.407 | 0.421 | 0.409 | ||
| ETTh2 | 0.369 | 0.385 | 0.372 | 0.388 | |||
| ETTm1 | 0.378 | 0.370 | 0.370 | 0.381 | |||
| ETTm2 | 0.285 | 0.315 | 0.291 | 0.318 | |||
| ECL | 0.164 | 0.249 | 0.164 | 0.249 | 0.164 | 0.249 | |
| Weather | 0.253 | 0.264 | 0.263 | 0.268 | |||
| Exchange | 0.379 | 0.410 | 0.404 | 0.420 | |||
| Moirai | ETTh1 | 0.530 | 0.473 | 0.620 | 0.504 | ||
| ETTh2 | 0.405 | 0.415 | 0.451 | 0.438 | |||
| ETTm1 | 0.617 | 0.485 | 0.740 | 0.523 | |||
| ETTm2 | 0.344 | 0.360 | 0.404 | 0.388 | |||
| ECL | 0.391 | 0.407 | 0.520 | 0.446 | |||
| Weather | 0.315 | 0.310 | 0.374 | 0.333 | |||
| Exchange | 0.463 | 0.443 | 0.601 | 0.488 |
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.
| Model | Dataset | Ori | +GRAM | ||
|---|---|---|---|---|---|
| Metric | MSE | MAE | MSE | MAE | |
| MixLinear | ETTh1 | 0.351 | 0.383 | 0.351 | 0.383 |
| ETTh2 | 0.283 | 0.340 | 0.278 | 0.336 | |
| ETTm1 | 0.333 | 0.375 | 0.327 | 0.370 | |
| ETTm2 | 0.180 | 0.271 | 0.176 | 0.266 | |
| Weather | 0.179 | 0.237 | 0.168 | 0.225 | |
| Exchange | 0.084 | 0.203 | 0.083 | 0.202 | |
| xPatch | ETTh1 | 0.366 | 0.392 | 0.370 | 0.393 |
| ETTh2 | 0.275 | 0.334 | 0.274 | 0.332 | |
| ETTm1 | 0.289 | 0.336 | 0.289 | 0.336 | |
| ETTm2 | 0.160 | 0.246 | 0.160 | 0.244 | |
| Weather | 0.144 | 0.183 | 0.143 | 0.182 | |
| Exchange | 0.084 | 0.200 | 0.082 | 0.199 | |
| DUET | ETTh1 | 0.352 | 0.384 | 0.353 | 0.384 |
| ETTh2 | 0.273 | 0.337 | 0.271 | 0.336 | |
| ETTm1 | 0.278 | 0.333 | 0.276 | 0.330 | |
| ETTm2 | 0.161 | 0.249 | 0.161 | 0.247 | |
| Weather | 0.148 | 0.193 | 0.147 | 0.192 | |
| Exchange | 0.079 | 0.197 | 0.079 | 0.197 | |
| ETTh1 | 0.360 | 0.391 | 0.360 | 0.390 | |
| ETTh2 | 0.277 | 0.341 | 0.273 | 0.337 | |
| ETTm1 | 0.286 | 0.340 | 0.286 | 0.339 | |
| ETTm2 | 0.163 | 0.251 | 0.162 | 0.249 | |
| Weather | 0.147 | 0.197 | 0.146 | 0.195 | |
| Exchange | 0.084 | 0.201 | 0.083 | 0.200 | |
| SparseTSF | ETTh1 | 0.360 | 0.386 | 0.360 | 0.386 |
| ETTh2 | 0.273 | 0.334 | 0.270 | 0.331 | |
| ETTm1 | 0.305 | 0.346 | 0.302 | 0.344 | |
| ETTm2 | 0.165 | 0.255 | 0.164 | 0.252 | |
| Weather | 0.173 | 0.225 | 0.164 | 0.216 | |
| Exchange | 0.084 | 0.201 | 0.083 | 0.200 | |
| Pathformer | ETTh1 | 0.372 | 0.392 | 0.372 | 0.392 |
| ETTh2 | 0.277 | 0.336 | 0.274 | 0.333 | |
| ETTm1 | 0.290 | 0.335 | 0.289 | 0.334 | |
| ETTm2 | 0.158 | 0.246 | 0.157 | 0.244 | |
| Weather | 0.146 | 0.190 | 0.145 | 0.189 | |
| Exchange | 0.088 | 0.208 | 0.087 | 0.207 |
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 and 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.