- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper presents TS-CorM, a memory-based framework for adapting frozen time-series foundation models through forecast correction. The method represents forecast residuals using compact coefficients over a structured dictionary and constructs memory keys from joint history–forecast representations. Slow Memory learns offline correction mappings through ridge regression, while Fast Memory incorporates delayed feedback through recursive least squares without backpropagation. Experiments across eight datasets and five TSFM backbones report improvements across all 40 backbone–dataset combinations. Compared with the strongest adaptation baseline, the offline method reduces average MSE by 3.78% with 98.32% fewer adaptation parameters. Online updates provide an additional average MSE reduction of 0.565% in the evaluated settings.
Critical strengths
The paper focuses on how to achieve target domain adaptation of TSFM without modifying pre trained backbone parameters, or it can be understood as post training, which has practical application significance.
From the experimental results in Table 1, it does indeed exceed the fine-tuning method of the compared TSFM, although only ChronosBolt and we have some doubts about the subsequent experimental results.
Critical weaknesses
We have concerns about the transparency and reproducibility of the experimental evaluation. While we respect the authors' decision not to release supplementary materials or source code, the anonymous repository linked in the abstract contains only an essentially empty README file. This is particularly concerning given the existing issues with the fairness and comprehensiveness of the experimental comparisons, the insufficiently documented experimental settings, and the mismatch between the claimed contributions of Slow and Fast Memory and their empirical validation. The lack of detailed experimental configurations and implementation information in the appendix further undermines our confidence in the reliability and reproducibility of the reported results.
W1. Limited benchmark diversity and generalization evidence. Although the paper evaluates eight datasets, four belong to the ETT family, leaving relatively limited diversity in domains and temporal characteristics. TiRex-2 is also evaluated only at horizons 96 and 192. Considering the paper's claim of general TSFM adaptation, I recommend extending the evaluation to broader benchmarks such as GIFT-Eval and fev-bench, using appropriate target-domain adaptation protocols.
W2. Insufficient and potentially unfair comparisons under supervised adaptation. The paper compares supervised TS-CorM adaptation with the original zero-shot TSFMs, but these results do not establish its advantages under comparable training conditions. Since TS-CorM requires labeled target-domain data, it would be important to evaluate how the original TSFMs perform after fine-tuning with the same supervision budget, beyond the reported LoRA baseline. Moreover, once downstream training data are available, conventional long-term forecasting models, such as DLinear, PatchTST, and iTransformer, become relevant alternatives. These models may achieve competitive or better performance without requiring a pretrained TSFM or an additional memory module. Although TS-CorM introduces relatively few adaptation parameters, this alone does not establish its efficiency or practical advantages over lightweight forecasting models trained directly on the target domain. The paper also does not sufficiently clarify whether all adaptation baselines use identical supervised data budgets and training protocols.
W3. Limited contribution and validation of Fast Memory. Although the paper emphasizes a fast–slow memory framework, the main performance comparisons and ablations rely exclusively on the offline Slow Memory. Fast Memory is evaluated separately on only six datasets with one backbone, providing an additional average MSE reduction of just 0.565% and a negligible MAE improvement of 0.061%. Since both memories are essentially linear correction mappings learned through ridge regression or its recursive counterpart, the experiments do not sufficiently establish the necessity of the fast–slow decomposition or its advantages over simpler offline and online residual regression baselines.
W4. Numerical inconsistencies across experimental results. Several results appear inconsistent across the main experiments and appendix. For example, the average Weather MSE in Tables 1 and 2 is approximately 0.228, whereas Table 6 reports 0.174 for the full model. Similarly, the offline MSE for Exchange at horizon 336 differs between Tables 1 and 7 (0.307 vs. 0.332). Without clarification of the underlying configurations, the reliability of these experimental conclusions remains uncertain.
Questions for authors
Q1. Could the authors clarify the numerical inconsistencies across Tables 1, 2, 6, and 7, particularly the different offline results reported for identical dataset–horizon settings?
Q2. What distinguishes the proposed memory mechanism from conventional ridge regression and RLS? What are the specific advantages of separating Slow Memory and Fast Memory?
Q3. Figure 2 and Eq. (8) appear to use different residual definitions for Fast Memory. Which definition is used in the actual implementation?
Q4. Can the author provide more experimental settings and code, as many of the experiments in this paper require detailed verification, and providing an empty repository is irresponsible behavior.
4 comments
Author Response to Reviewer UXFN (2/3)
W2 Insufficient and potentially unfair comparisons under supervised adaptation.
Response: We thank the reviewer for raising this point. To evaluate the benefit of TS-CorM when labeled target-domain data are available, we added full fine-tuning of Chronos-Bolt alongside the existing LoRA baseline. We also evaluated DLinear, PatchTST, and iTransformer trained directly on each target dataset. The table below reports results on six datasets at four forecast horizons (96, 192, 336, and 720 steps). It also includes FW-FC, EC-FC, and MEMTS, the additional baselines requested by Reviewer VUFa in W1 & Q1.
Across the 24 settings, offline TS-CorM achieved the lowest average MSE/MAE (0.326/0.350), compared with 0.338/0.358 for fully fine-tuned Chronos-Bolt and 0.359/0.391 for DLinear, the strongest of the three directly trained models. TS-CorM achieved the lowest MSE in 20/24 settings and the lowest MAE in 22/24 settings. We will include these additional experiments in the revised manuscript.
| Dataset | Horizon | TS-CorM(offline) | FW-FC | EC-FC | MEMTS | Chronos-Bolt(full FT) | DLinear | PatchTST | iTransformer |
|---|---|---|---|---|---|---|---|---|---|
| ETTh1 | 96 | 0.359/0.372 | 0.364/0.382 | 0.362/0.381 | 0.394/0.392 | 0.372/0.375 | 0.372/0.398 | 0.390/0.422 | 0.397/0.420 |
| ETTh1 | 192 | 0.405/0.403 | 0.414/0.416 | 0.413/0.414 | 0.460/0.440 | 0.425/0.406 | 0.407/0.421 | 0.464/0.470 | 0.430/0.443 |
| ETTh1 | 336 | 0.440/0.422 | 0.456/0.443 | 0.465/0.447 | 0.479/0.439 | 0.461/0.428 | 0.443/0.447 | 0.511/0.496 | 0.454/0.458 |
| ETTh1 | 720 | 0.445/0.442 | 0.479/0.479 | 0.484/0.479 | 0.610/0.541 | 0.475/0.453 | 0.474/0.496 | 0.587/0.529 | 0.465/0.482 |
| ETTh2 | 96 | 0.279/0.323 | 0.381/0.385 | 0.370/0.379 | 0.289/0.327 | 0.275/0.318 | 0.272/0.341 | 0.328/0.374 | 0.326/0.374 |
| ETTh2 | 192 | 0.335/0.363 | 0.604/0.465 | 0.604/0.465 | 0.372/0.403 | 0.360/0.372 | 0.352/0.398 | 0.413/0.430 | 0.384/0.412 |
| ETTh2 | 336 | 0.369/0.393 | 0.706/0.509 | 0.522/0.473 | 0.393/0.405 | 0.396/0.401 | 0.406/0.442 | 0.436/0.449 | 0.413/0.443 |
| ETTh2 | 720 | 0.381/0.411 | 0.551/0.507 | 0.555/0.515 | 0.423/0.455 | 0.398/0.419 | 0.716/0.600 | 0.429/0.460 | 0.437/0.465 |
| ETTm1 | 96 | 0.293/0.321 | 0.301/0.334 | 0.294/0.335 | 0.332/0.339 | 0.296/0.327 | 0.307/0.351 | 0.333/0.368 | 0.330/0.371 |
| ETTm1 | 192 | 0.336/0.349 | 0.343/0.363 | 0.339/0.361 | 0.355/0.373 | 0.341/0.350 | 0.339/0.372 | 0.377/0.393 | 0.370/0.394 |
| ETTm1 | 336 | 0.385/0.376 | 0.381/0.388 | 0.370/0.386 | 0.412/0.410 | 0.393/0.394 | 0.387/0.406 | 0.372/0.400 | 0.381/0.406 |
| ETTm1 | 720 | 0.450/0.413 | 0.445/0.428 | 0.429/0.428 | 0.478/0.449 | 0.444/0.420 | 0.451/0.451 | 0.430/0.433 | 0.447/0.444 |
| ETTm2 | 96 | 0.166/0.239 | 0.175/0.255 | 0.252/0.321 | 0.174/0.247 | 0.169/0.244 | 0.169/0.259 | 0.180/0.270 | 0.189/0.273 |
| ETTm2 | 192 | 0.229/0.284 | 0.249/0.310 | 0.245/0.307 | 0.259/0.317 | 0.232/0.290 | 0.230/0.311 | 0.234/0.304 | 0.266/0.323 |
| ETTm2 | 336 | 0.283/0.321 | 0.312/0.356 | 0.307/0.352 | 0.317/0.336 | 0.289/0.338 | 0.284/0.347 | 0.295/0.350 | 0.336/0.367 |
| ETTm2 | 720 | 0.362/0.374 | 0.404/0.419 | 0.396/0.415 | 0.423/0.439 | 0.380/0.377 | 0.407/0.427 | 0.375/0.398 | 0.400/0.410 |
| Exchange | 96 | 0.090/0.209 | 1.091/0.527 | 1.102/0.528 | 0.092/0.215 | 0.090/0.210 | 0.083/0.207 | 0.205/0.326 | 0.139/0.275 |
| Exchange | 192 | 0.174/0.297 | 1.510/0.667 | 1.537/0.665 | 0.179/0.299 | 0.177/0.301 | 0.178/0.306 | 0.196/0.322 | 0.254/0.371 |
| Exchange | 336 | 0.307/0.401 | 2.962/0.874 | 3.026/0.892 | 0.342/0.424 | 0.352/0.422 | 0.529/0.536 | 0.771/0.701 | 0.388/0.454 |
| Exchange | 720 | 0.821/0.680 | 2.898/1.026 | 1.402/0.852 | 1.001/0.760 | 0.828/0.697 | 0.834/0.690 | 1.240/0.872 | 1.200/0.854 |
| Weather | 96 | 0.153/0.192 | 0.168/0.220 | 0.166/0.219 | 0.183/0.212 | 0.163/0.192 | 0.177/0.243 | 0.158/0.210 | 0.163/0.218 |
| Weather | 192 | 0.195/0.233 | 0.212/0.262 | 0.212/0.262 | 0.225/0.273 | 0.201/0.249 | 0.213/0.270 | 0.207/0.254 | 0.213/0.264 |
| Weather | 336 | 0.246/0.272 | 0.270/0.304 | 0.269/0.303 | 0.277/0.296 | 0.251/0.275 | 0.257/0.307 | 0.265/0.299 | 0.261/0.300 |
| Weather | 720 | 0.316/0.321 | 0.349/0.356 | 0.348/0.355 | 0.367/0.398 | 0.344/0.333 | 0.327/0.368 | 0.339/0.350 | 0.329/0.341 |
| Average | — | 0.326/0.350 | 0.668/0.445 | 0.603/0.439 | 0.368/0.383 | 0.338/0.358 | 0.359/0.391 | 0.397/0.412 | 0.374/0.398 |
Author Response to Reviewer UXFN (3/3)
Q2 & W3 Limited contribution and validation of Fast Memory.
Response:
1. Extending the test sequences. The modest 0.061% improvement in average MAE for Fast Memory in the paper is mainly due to the fact that Fast Memory often has a negative effect at longer forecast horizons. This is because, for a forecast horizon of , the model must wait steps to receive complete feedback. Therefore, for a test sequence of fixed length, the longer the horizon, the less feedback is available for updates. We adjusted the proportions of the training, validation, and test sets to extend the test sequences. Under the new split, enabling Fast Memory reduced both MSE and MAE on all six datasets compared with using frozen Slow Memory alone. Average MSE and MAE decreased by 3.23% and 2.31%, respectively, compared with reductions of 0.565% and 0.061% under the original split. This result shows that the online gains persist and are more pronounced on longer test sequences.
| Dataset | TS-CorM (offline) | TS-CorM (online) | ΔMSE | ΔMAE |
|---|---|---|---|---|
| ETTm1 | 0.537722 / 0.448633 | 0.529754 / 0.447623 | −1.482% | −0.225% |
| ETTm2 | 0.382896 / 0.387774 | 0.379025 / 0.386001 | −1.011% | −0.457% |
| Weather | 0.392253 / 0.402631 | 0.390255 / 0.401152 | −0.509% | −0.367% |
| ETTh1 | 0.662415 / 0.692347 | 0.601727 / 0.643529 | −9.162% | −7.051% |
| ETTh2 | 0.432258 / 0.427357 | 0.402891 / 0.409825 | −6.794% | −4.102% |
| Exchange | 0.857721 / 0.700241 | 0.856021 / 0.700184 | −0.198% | −0.008% |
| Average | 0.544211 / 0.509831 | 0.526612 / 0.498052 | −3.23% | −2.31% |
2. Contribution of separating Slow Memory and Fast Memory and its relationship to ridge regression and RLS. Slow Memory and Fast Memory both use established regression methods. They differ in what they learn and how they correct the forecast. We represent TSFM forecast residuals as correction coefficients. Slow Memory learns an offline mapping from the forecast state to these coefficients and uses them to correct the backbone forecast. Once the ground truth for the full forecast horizon is available, Fast Memory uses RLS to learn an online correction for the remaining residual. This design preserves the offline mapping and allows us to adjust the strength of the online correction independently.
To test the benefit of separating Slow and Fast Memory, we added a direct-update baseline that updates Slow Memory as feedback becomes available. The table reports average results across six datasets and four forecast horizons using Chronos-Bolt.
| Variant | Mean MSE ↓ | Mean MAE ↓ |
|---|---|---|
| A. Frozen Slow Memory | 0.330871 | 0.353754 |
| B. Direct online updates to Slow Memory | 0.329564 | 0.353695 |
| C. Slow Memory + Fast Memory | 0.329002 | 0.353538 |
Direct updates improved performance over Slow Memory. Separating Slow and Fast Memory yielded further improvement over direct updates, with lower MSE in 20/24 settings, lower MAE in 20/24. The relative reductions averaged 0.171% and 0.044% for MSE and MAE. We will include this comparison in the revised manuscript.
Q3 Figure 2 and Eq. (8) appear to use different residual definitions for Fast Memory. Which definition is used in the actual implementation?
Response: Thank you for pointing this out. The residual formula in Figure 2(c) is incorrect and should match Eq. (8). We will correct it to , where is the forecast corrected by Slow Memory.
Q4. Experimental settings and code availability
Response: Thank you for raising this concern. We apologize for providing an empty anonymous repository. We originally intended to release the code at submission. Since this work was conducted during a company internship, public release requires the company’s confidentiality approval. We should have made this clear when we shared the link.
We are committed to releasing the code for all experiments as soon as approval is granted. In the meantime, we will expand the appendix with the data splits, fitting procedures, online update protocol, and aggregation rules used for each table to make the experimental protocol explicit and support reproducibility.
To Reviewer UXFN
Dear Reviewer UXFN,
Thank you for your careful review and constructive feedback. We appreciate your recognition of the practical value of adapting time series foundation models while keeping their backbones frozen.
In our rebuttal, we expanded the evaluation to include GIFT-Eval and added supervised comparisons with full fine-tuning of Chronos-Bolt and with DLinear, PatchTST, and iTransformer trained directly on the target datasets. We also evaluated Fast Memory on longer test sequences, explained why we separate Slow and Fast Memory, and clarified the residual definition used for online updates. Finally, we addressed the numerical discrepancies across tables, provided more details on the experimental settings, and explained the current status of the code release.
We would appreciate your feedback on whether our responses address your concerns. If you find them satisfactory, we would be grateful for your support. Please let us know if you have any further questions. Thank you again for your time and consideration.
Author Response to Reviewer UXFN (1/3)
Response: We thank the reviewer for pointing out the limited dataset diversity in our original evaluation. We extended our experiments to five datasets from GIFT-Eval. US Births records daily birth counts, Saugeen measures river flow, and M_DENSE contains urban traffic data. BizITObs Application and BizITObs-L2C cover application monitoring and business-process monitoring, respectively.
Using Chronos-Bolt as the backbone, we evaluated offline target-domain adaptation at horizons of 96, 192, 336, and 720 steps. TS-CorM reduced both MSE and MAE relative to unadapted Chronos-Bolt in all 20 settings. Compared with TS-Memory, TS-CorM achieved lower MSE in 18/20 settings and lower MAE in 18/20, with both metrics lower in 17/20. These results show that the gains from TS-CorM extend beyond the ETT family to all five additional datasets.
Each cell reports MSE / MAE. Bold indicates the best.
Response: Thank you for pointing this out. We sincerely apologize for this oversight. The discrepancy between the offline MSE for Weather in Table 2 (0.228) and Table 6 (0.174) arose because we averaged only the results for H=96 and H=192 in Table 6, inadvertently omitting H=336 and H=720. We have recalculated the results across all four horizons. The corrected offline MSE is 0.228, consistent with Table 2. The corrected Weather results for all configurations are shown below.
The discrepancy between Tables 1 and 7 has a different source. Table 1 presents our main experimental results and reports offline performance averaged across multiple random seeds. For the Fast Memory experiment in Table 7, we used a saved checkpoint from a single seed for both the offline and online runs. Consequently, some offline results in Table 7 differ slightly from those in Table 1. Using the same checkpoint for both runs in Table 7 allows us to isolate the effect of enabling Fast Memory. We will clarify how the results were aggregated in the captions of both tables. We apologize again for this oversight.