acceptodds
Under review as a conference paper at ICLR 2027

Test-time State-augmented Adaptive Forecasting for Time Series Foundation Models

Abstract

Time series foundation models (TSFMs) have demonstrated zero-shot forecasting capability on unseen time series data from various domains. When forecasting with a TSFM, a fixed-length input window is fed into the TSFM. Since typical TSFMs are transformer-based, they incur high computational costs when the input window length is large, due to the transformers' quadratic computational complexity with respect to the input length. Although shortening the input window length can save computational costs, there is a trade-off between the computational costs and forecast performance because transformers are stateless and cannot capture temporal dependencies beyond an input window, and short input windows make a distributional gap relative to the TSFM's pre-training. In this paper, we address the trade-off between the computational cost and performance when forecasting with a TSFM. To mitigate accuracy drops with short input windows, we propose State-augmented Adaptive Forecasting (SAAF), which enables off-the-shelf pre-trained TSFMs to capture temporal dependencies beyond the input window under restricted window lengths. During inference, SAAF incorporates a stateful module that captures temporal dependencies with short input windows. In addition, SAAF transforms a TSFM's outputs to mitigate the TSFM's performance degradation caused by short input windows. Experimental results show that SAAF had lower forecast errors than baselines, especially with short input windows, across various forecasting benchmarks, at lower computational costs.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series foundation models, test-time adaptation

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_kEVXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes SAAF, a training-free post-hoc module that lets a frozen TSFM use information beyond a short input window. A HiPPO state accumulates the whole past and feeds an online linear forecaster (CAFM), a second online linear map corrects the TSFM output (FTM), and the two are combined with an error-based weight; all updates are closed-form Bayesian linear regression at inference. Experiments on ETT and 12 Monash datasets with TimesFM (other backbones in the appendix) show clear MAE gains at short context lengths but not at long ones.

Critical strengths
  1. The method needs no training, no backpropagation, and no access to TSFM internals, so it attaches to any off-the-shelf TSFM at negligible cost, and it still gives clear MAE gains at short context lengths across several backbones. This simplicity is the main reason I'd consider the paper.

  2. The error-based weighting makes the adaptation respond to the current situation online, relying on the corrected TSFM output early and shifting to the accumulated state once it has warmed up, rather than applying a fixed correction.

Critical weaknesses
  1. I'm not convinced the method actually improves the TSFM. In Fig. 3, SAAF is worse than No adapt (and than its own FTM component) for L≥256L \geq 256 on ETT-h1, h2, and m2, and the claim in Sec. 5.5 that it also wins at large LL on m2 doesn't match the figure. Since the paper's own motivation is that TSFMs can't use information beyond the window, which holds at any LL, I'd expect the gains to carry over to longer windows as well; if they only appear at short LL, it looks like SAAF is compensating for short-window degradation (which FTM alone does) rather than exploiting beyond-window context. The gate doesn't help here either, since with γ=10\gamma = 10 CAFM keeps roughly 10-20% weight even when clearly worse.

  2. The evaluation is narrower than it should be for a plug-in method whose core claim is that it helps any TSFM: one backbone and two baselines in the main text, and only ETT plus 12 Monash datasets rather than a broad benchmark. The Monash subset was selected for "sufficient records and timesteps," which is exactly the long-TT regime where CAFM can warm up (Sec. 5.5, Appendix A), so the dataset choice favors the method by construction and needs justification.

  3. The method has many hyperparameters (α\alpha, β\beta, mm, γ\gamma, DD), all set once on ETTh1, which is also a test dataset (Sec. 5.3). The paper's own Fig. 14 shows the setting doesn't transfer to Monash, and in an online setting there is no validation data to tune on, so the reported numbers are open to cherry picking.

  4. Fonts in Tables and Figs are far too small; most per-dataset numbers can't be read without heavy zooming.

  5. Section 4 is a sequence of notation and equations with little explanation of what each component does, and most of the notation isn't needed to understand the method. It's hard to follow.

Questions for authors
  1. (Clarification) Appendix E.2 states that the Monash datasets were not used to pre-train Chronos or TimesFM. Could you verify this per dataset? As far as I know, several of the selected datasets (e.g., London Smart Meters, KDD Cup 2018, Pedestrian Counts, Wind Farms) are in Chronos's training corpus, and TimesFM 2.0 was trained with LOTSA, which includes Monash. If there is overlap, please report results restricted to the non-overlapping datasets.

  2. (Suggestion) A compact summary across all backbones and baselines (e.g., mean MAE and average rank per backbone) in the main text would make the generality claim much easier to assess.

For other points, please refer to the weaknesses.

3 comments

Dear reviewer kEVX,

We appreciate your constructive comments. We updated the manuscript based on your comments (highlighted in blue).

Weakness 1: I'm not convinced the method actually improves the TSFM. In Fig. 3, SAAF is worse than No adapt (and than its own FTM component) for L≥256L \geq 256 on ETT-h1, h2, and m2, and the claim in Sec. 5.5 that it also wins at large LL on m2 doesn't match the figure. Since the paper's own motivation is that TSFMs can't use information beyond the window, which holds at any LL, I'd expect the gains to carry over to longer windows as well; if they only appear at short LL, it looks like SAAF is compensating for short-window degradation (which FTM alone does) rather than exploiting beyond-window context. The gate doesn't help here either, since with γ=10\gamma=10 CAFM keeps roughly 10-20% weight even when clearly worse.

Thank you for pointing it out. We revised the statement about ETT-m2 in Sec. 5.5 accordingly.

When the input window length LL is large, the TSFM backbone generally produces more accurate forecasts, thereby improving the performance of FTM. However, we observed occasional spikes in FTM's forecasting error. During these periods, DWM assigns a larger weight to CAFM. Because DWM estimates each module's reliability using an exponential moving average of past errors, the increased weight assigned to CAFM can persist. This delayed adjustment can lead to higher forecasting errors for SAAF, particularly when CAFM is less accurate than FTM. In our experiments, we used the same Bayesian linear regression hyperparameters (α\alpha and β\beta) for both CAFM and FTM. We hypothesize that tuning these hyperparameters separately for each module could improve the stability of online regression updates, potentially mitigating the resulting performance degradation.

Weakness 2: The evaluation is narrower than it should be for a plug-in method whose core claim is that it helps any TSFM: one backbone and two baselines in the main text, and only ETT plus 12 Monash datasets rather than a broad benchmark. The Monash subset was selected for "sufficient records and timesteps," which is exactly the long-TT regime where CAFM can warm up (Sec. 5.5, Appendix A), so the dataset choice favors the method by construction and needs justification.

We will report a compact summary that includes the other TSFMs in the main paper (Question 2).

Although SAAF can increase forecasting errors when TT is small, the performance degradation by TT is limited owing to DWM. To investigate the effect of the total number of time steps, TT, on forecasting performance, we conducted an additional analysis using the Monash datasets. Specifically, we varied TT and computed the accumulated mean absolute error (MAE) over the forecasting period truncated at each TT. We added the results to Appendix F.6.

For (a) London Smart Meters and (b) Australian Electricity Demand, SAAF exhibits increased forecasting errors when TT is small. The error peaks of SAAF and CAFM occur at approximately the same TT values, suggesting that this behavior is attributable to CAFM. Nevertheless, SAAF achieves the lowest MAE among the baselines for T≥1000T \geq 1000 across both datasets, demonstrating that its performance advantage is maintained even as TT decreases. For (c) Wind Farms, SAAF exhibits higher forecasting errors than the baselines, mainly due to the relatively poor performance of CAFM. However, the additional degradation associated with small TT is limited.

Weakness 3: The method has many hyperparameters (α\alpha, β\beta, mm, γ\gamma, DD), all set once on ETTh1, which is also a test dataset (Sec. 5.3). The paper's own Fig. 14 shows the setting doesn't transfer to Monash, and in an online setting there is no validation data to tune on, so the reported numbers are open to cherry picking.

We would like to clarify that our results were not obtained through dataset-specific hyperparameter cherry-picking. We used the same hyperparameter configuration for the remaining ETT datasets and the Monash datasets, without any additional tuning. Therefore, the reported results were not obtained through dataset-specific hyperparameter selection.

As shown in the sensitivity analysis on the Monash datasets (Fig. 14), the effects of α\alpha and β\beta differ notably from those observed on the ETT datasets. However, the other hyperparameters exhibit similar trends across the two benchmarks, suggesting that their effects generalize across datasets.

Additionally, we evaluated SAAF on GIFT-Eval using the same hyperparameter configuration. We confirmed its effectiveness (Appendix F.10), providing additional evidence that SAAF can generalize to other benchmarks without dataset-specific tuning.

Weakness 4: Fonts in Tables and Figs are far too small; most per-dataset numbers can't be read without heavy zooming.

We apologize for the inconvenience. We adjusted figures and tables in the revised paper.

Weakness 5: Section 4 is a sequence of notation and equations with little explanation of what each component does, and most of the notation isn't needed to understand the method. It's hard to follow.

We apologize for the inconvenience. We will update the explanation in Sec. 4 to clarify what each component does during the rebuttal period.

Question 1: (Clarification) Appendix E.2 states that the Monash datasets were not used to pre-train Chronos or TimesFM. Could you verify this per dataset? As far as I know, several of the selected datasets (e.g., London Smart Meters, KDD Cup 2018, Pedestrian Counts, Wind Farms) are in Chronos's training corpus, and TimesFM 2.0 was trained with LOTSA, which includes Monash. If there is overlap, please report results restricted to the non-overlapping datasets.

Thank you for pointing it out. We acknowledge that we used TimesFM 2.0 in our experiments but mistakenly referred to the paper describing TimesFM 1.0. We also confirm that, as the reviewer correctly noted, some of the Monash datasets used in our evaluation were included in Chronos's pre-training corpus. We apologize for these inaccuracies and corrected the description in Appendix E.2. We also explicitly acknowledged the potential overlap between the TSFMs' pre-training datasets and our evaluation datasets in Sec. 5.1.

The main purpose of our experiments is not to compare the forecasting accuracy of different TSFMs, but to evaluate adaptation methods applied to the same backbone TSFM. Even on Monash, which is also used to pre-train TimesFM, LL largely affects the performance, as we observed in Tab. 2. Therefore, we believe that potential overlap with pre-training data does not necessarily invalidate the relative comparisons between adaptation methods, as they are evaluated using the same backbone and observations.

To further address this concern, we conducted additional experiments on GIFT-Eval. We added the results to Appendix F.10. We observed similar trends to those on ETT and Monash, with SAAF providing particularly pronounced improvements when the input window length LL is short.

Question 2: (Suggestion) A compact summary across all backbones and baselines (e.g., mean MAE and average rank per backbone) in the main text would make the generality claim much easier to assess.

Thank you for your suggestion. We will add a compact summary to the main paper during the rebuttal period.

Thanks for the response and the revision. I read the updated paper, but my main concerns remain.

The degradation at long context lengths is still there, and it also shows up on Monash and the new GIFT-Eval results, so short series alone can’t explain it. The response explains this by a delayed adjustment of the weights after error spikes in FTM, but this is only a hypothesis with no supporting experiment, and the paper still gives a different reason (CAFM not being warmed up). My point about the gate keeping weight on CAFM even when it is clearly worse was not addressed. It is still unclear why a method motivated by beyond-window information hurts once the window gets longer.

The Monash selection is still not justified, and the new analysis on total timesteps actually confirms that the method needs long series to win. The GIFT-Eval experiment covers only a few configurations from traffic datasets with one backbone, so it is too small to count as a broad benchmark evaluation, and SAAF doesn’t beat No adapt on the shorter series there either.

On hyperparameters, my concern was not per-dataset tuning. The configuration was chosen by looking at ETTh1, which is one of the test sets, and that is not a valid way to select it. It should be chosen on data that is not used for evaluation. The response also agrees that alpha and beta behave differently across benchmarks.

For these reasons I’m keeping my score.

Reviewer_iPPMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes State-augmented Adaptive Forecasting (SAAF) to improve the forecasting performance of time series foundation models (TSFMs) with short input windows. SAAF consists of three modules: CAFM, which accumulates historical information beyond the current input window, FTM, which adapts the TSFM's predictions, and DWM, which dynamically combines their outputs. The method works at test time without modifying the pre-trained TSFM or requiring backpropagation. Experiments on multiple benchmark datasets show that SAAF can improve forecasting accuracy, especially with short input windows, while keeping computational costs low.

Critical strengths
  1. The paper evaluates SAAF on multiple benchmark datasets and several pre-trained TSFMs, showing consistent improvements across different settings, especially with short input windows. The authors also provide ablation studies and sensitivity analyses to better understand the contribution of each component.
Critical weaknesses
  1. The authors attribute the poor performance of TSFMs with short input windows to the stateless nature of Transformers. However, this limitation seems to come more from the fixed-window forecasting setup rather than the Transformer architecture itself. The paper could provide a clearer explanation of this distinction.
  2. While SAAF shows promising results, the paper does not compare it with parameter-efficient fine-tuning methods, or a wider range of lightweight forecasting models designed for short input windows. These comparisons would help show whether SAAF offers clear advantages in both forecasting accuracy and computational cost.
Questions for authors
  1. The authors argue that Transformers cannot access information beyond the current input window, while RNNs can retain past information through hidden states. However, RNNs face the same limitation if their hidden states are reset between windows. Therefore, this distinction seems to depend more on whether states are carried across windows than on the model architecture itself. Could the authors clarify this point?
  2. The proposed CAFM accumulates historical information beyond the current input window into a fixed-dimensional state vector. While this may not be information leakage, it allows SAAF to use more historical information than the original TSFM and some baselines. Could the authors discuss whether this makes the experimental comparisons less fair?
  3. How does SAAF compare with directly fine-tuning a small number of parameters in the TSFM? Which approach performs better? Have the authors explored this comparison?
  4. One of the main advantages of TSFMs is their strong zero-shot forecasting ability. However, SAAF introduces additional trainable modules on top of a pre-trained TSFM. In this case, why not simply train a new model specifically designed for short input windows, without relying on a TSFM? Could the authors discuss the advantages of SAAF over this alternative?
3 comments

Dear reviewer iPPM,

We appreciate your insightful comments. We updated the manuscript based on your comments (highlighted in blue).

Weakness 1: The authors attribute the poor performance of TSFMs with short input windows to the stateless nature of Transformers. However, this limitation seems to come more from the fixed-window forecasting setup rather than the Transformer architecture itself. The paper could provide a clearer explanation of this distinction.

Question 1: The authors argue that Transformers cannot access information beyond the current input window, while RNNs can retain past information through hidden states. However, RNNs face the same limitation if their hidden states are reset between windows. Therefore, this distinction seems to depend more on whether states are carried across windows than on the model architecture itself. Could the authors clarify this point?

Thank you for pointing it out. We agree that the fundamental limitation arises not from the model architecture itself, but from whether historical information can be retained across input windows. Even stateful architectures, such as RNNs, face the same limitation if their hidden states are reset at each inference step. This limitation is especially problematic for Transformer-based architectures. Accordingly, we have revised the Abstract and Sec. 1 to clarify that this limitation stems from the rolling window forecasting setting in which each input window is processed independently, rather than from an inherent property of the Transformer architecture.

Nevertheless, our contributions remain unchanged: SAAF enables pre-trained TSFMs to leverage historical information beyond the current input window via a lightweight, stateful module that can be added post-hoc without modifying or retraining the backbone.

Weakness 2: While SAAF shows promising results, the paper does not compare it with parameter-efficient fine-tuning methods, or a wider range of lightweight forecasting models designed for short input windows. These comparisons would help show whether SAAF offers clear advantages in both forecasting accuracy and computational cost.

Question 3: How does SAAF compare with directly fine-tuning a small number of parameters in the TSFM? Which approach performs better? Have the authors explored this comparison?

Thank you for your suggestion. We conducted additional experiments in which we attached LoRA adapters to TimesFM and fine-tuned them online under the rolling window setting on the ETT dataset. We added the results in Appendix F.8.

The results show that LoRA is unstable in our setting and performs worse than No adapt in most cases. We hypothesize that this degradation is due to noisy gradient estimates arising from the limited amount of data available at each update step. Although larger batch sizes may improve the stability of fine-tuning, they would require accumulating more observations before each update, which is inconsistent with our goal of lightweight and immediate online adaptation.

Furthermore, fine-tuning the TSFM alone does not directly address the inability to access historical observations beyond the current input window. In contrast, SAAF explicitly incorporates such information through CAFM.

Finally, gradient-based fine-tuning is not universally applicable to TSFMs due to differentiability constraints (e.g., sampling-based inference) or limited access to model parameters and gradients through APIs. In contrast, SAAF treats the backbone TSFM as a black box, requiring neither gradients nor access to its internal parameters. This enables SAAF to be applied to a broader range of pre-trained TSFMs.

Question 2: The proposed CAFM accumulates historical information beyond the current input window into a fixed-dimensional state vector. While this may not be information leakage, it allows SAAF to use more historical information than the original TSFM and some baselines. Could the authors discuss whether this makes the experimental comparisons less fair?

Thank you for your clarification. We believe the comparisons are fair because all methods have access to the same observations under the rolling window forecasting setting, where input windows are provided sequentially over time. The difference lies in how each method utilizes the available information, rather than in the information provided to it.

A key motivation of our work is that existing TSFMs and adaptation methods cannot effectively exploit historical information beyond the current input window under this setting, particularly when the input window is short. Feeding only short input windows to a TSFM can be one practical solution to reduce computation under resource-constrained environments, as mentioned in Sec. 1. Our main contribution is to address this limitation by introducing a post-hoc mechanism that enables pre-trained TSFMs to retain and use historical information across input windows without modifying the backbone TSFM.

Question 4: One of the main advantages of TSFMs is their strong zero-shot forecasting ability. However, SAAF introduces additional trainable modules on top of a pre-trained TSFM. In this case, why not simply train a new model specifically designed for short input windows, without relying on a TSFM? Could the authors discuss the advantages of SAAF over this alternative?

The main advantage of SAAF is that it can be applied online without requiring training a new forecasting model in advance, enabling adaptation to incoming time-series data while retaining the benefits of pre-trained TSFMs. SAAF achieves its best forecasting performance by combining the zero-shot forecasting capabilities of a pre-trained TSFM with the complementary information from the additional modules, as demonstrated by the CAFM ablation study. This suggests that leveraging the pre-trained TSFM and adapting its predictions online is more effective than relying solely on the additional forecasting module.

Also, we conducted an additional experiment to introduce a new module designed for short input windows. We trained a Linear Module (LM) as a baseline that predicts from short input windows. LM directly forecasts future values from raw input windows using a linear model, whose parameters are updated online via Bayesian linear regression, similar to CAFM and FTM.

We added the results for LM and SAAF-LM, a variant of SAAF in which CAFM is replaced with LM, to Fig. 3. The results show that LM incurs large forecasting errors when the input window is short, because it lacks sufficient historical context. Even with longer input windows, its performance deteriorates, potentially because the increased input dimensionality yields a larger regression parameter matrix, making reliable online estimation more difficult given limited observations. Furthermore, SAAF-LM consistently exhibits higher forecasting errors than SAAF, with particularly pronounced differences at short input lengths.

Dear reviewer iPPM,

Thank you again for your constructive feedback. As the discussion period is ending soon, we would like to kindly ask whether our response has addressed your concerns. Please let us know if you have any further questions or concerns.

Best regards,

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

This paper proposes SAAF, a test-time adaptation framework designed to improve the trade-off between the computational cost and performance of TSFMs under restricted input context lengths. The proposed framework can be applied to pretrained TSFMs without modifying their parameters or requiring backpropagation. Experiments on ETT and Monash datasets with several TSFM backbones demonstrate improved forecasting accuracy. The paper further includes component ablations, sensitivity analyses, and computational efficiency evaluations.

Critical strengths
  1. The method exhibits excellent model-agnostic properties. SAAF does not rely on the internal representations or gradients of the backbone, so it can be applied to various pre-trained prediction architectures, offering a certain degree of plug-and-play flexibility.
  2. The actual improvement in short-window prediction is quite significant. The experimental results in Table 2 demonstrate that SAAF is indeed valuable under short-window conditions, while also highlighting its applicability limits.
Critical weaknesses
  1. The methodological innovation is limited, and the advantages offered by the HiPPO-based memory mechanism are not fully highlighted. The three main components of SAAF—the HiPPO-based state representation, online Bayesian linear regression, and error-driven ensemble weighting—are all existing techniques. Although integrating them into a model-agnostic, test-time adaptive framework has practical value, the method’s methodological contribution is relatively limited beyond this integration. Although the paper includes component ablation experiments and a comparison with ESN in Appendix F.7, these results do not sufficiently isolate the advantages of the proposed HiPPO representation over simpler historical features in online regression processes where all other conditions are held constant. The observed improvements may stem in part from the introduction of an additional online predictor rather than from a specific long-term memory representation.
  2. The paper provides limited validation of the universality of the claimed accuracy-efficiency trade-off. All key experiments used a fixed prediction length H = 64 and a rolling step size s = 8. In time-series forecasting, the prediction length significantly affects the difficulty of the online regression task and determines when historical observations can be used for updates. The rolling step size, in turn, affects the update frequency per unit time and the computational burden. Therefore, it remains unclear whether these conclusions can be generalized to other prediction horizons and feedback frequencies. Furthermore, the efficiency results depend on the base model. As noted in Section 5.8 and Appendix F.5, although more significant improvements were observed on Moirai and Lag-LLaMA, SAAF did not demonstrate notable improvements in the accuracy-efficiency tradeoff on TimesFM and Chronos-bolt. This limits the generalizability of the paper’s core efficiency claim. This is because the primary motivation of this paper is to improve the accuracy-efficiency tradeoff, rather than merely increasing prediction accuracy.
Questions for authors
  1. Could the authors conduct a controlled comparison between HiPPO-based memory and simpler historical representation methods using the same online Bayesian regression approach and a comparable memory budget? This would help determine whether these advantages truly stem from HiPPO-based historical encoding.​
  2. How does SAAF perform under different prediction horizons and rolling window sizes? Does the proposed method remain effective when the prediction horizon is significantly extended?
  3. Could the authors further clarify under what conditions SAAF can improve the trade-off between accuracy and efficiency? Given that the gains obtained vary across different TSFM backbone networks, it would be of significant value to clarify how update costs—such as different numbers of channels (C) and different prediction lengths (H)—determine whether SAAF can provide a significant computational advantage.
3 comments

Dear reviewer ifYB,

We appreciate your thoughtful comments. We updated the manuscript based on your comments (highlighted in blue).

Weakness 1-1: The methodological innovation is limited, and the advantages offered by the HiPPO-based memory mechanism are not fully highlighted. The three main components of SAAF—the HiPPO-based state representation, online Bayesian linear regression, and error-driven ensemble weighting—are all existing techniques. Although integrating them into a model-agnostic, test-time adaptive framework has practical value, the method’s methodological contribution is relatively limited beyond this integration.

Thank you for pointing it out. Our main contribution is to address the limitation of TSFMs with restricted input windows, rather than to integrate the components. As discussed in Sec. 2, existing studies generally assume that sufficiently long input contexts are available and have not explicitly addressed the performance degradation caused by the inability to access historical information beyond a restricted input window. Addressing this limitation is an important contribution of our work.

Weakness 1-2: Although the paper includes component ablation experiments and a comparison with ESN in Appendix F.7, these results do not sufficiently isolate the advantages of the proposed HiPPO representation over simpler historical features in online regression processes where all other conditions are held constant. The observed improvements may stem in part from the introduction of an additional online predictor rather than from a specific long-term memory representation.

Question 1: Could the authors conduct a controlled comparison between HiPPO-based memory and simpler historical representation methods using the same online Bayesian regression approach and a comparable memory budget? This would help determine whether these advantages truly stem from HiPPO-based historical encoding.

The improvements achieved by SAAF are not solely due to the addition of an online predictor but also stem from the HiPPO-based historical encoding employed in CAFM. To verify this, we conducted additional experiments using a Linear Module (LM) as a baseline online predictor. LM directly forecasts future values from raw input windows using a linear model, whose parameters are updated online via Bayesian linear regression, similar to CAFM and FTM.

We added the results for LM and SAAF-LM, a variant of SAAF in which CAFM is replaced with LM, to Fig. 3. The results show that LM incurs large forecasting errors when the input window is short, because it lacks sufficient historical context. Even with longer input windows, its performance deteriorates, potentially because the increased input dimensionality yields a larger regression parameter matrix, making reliable online estimation more difficult given limited observations. Furthermore, SAAF-LM consistently exhibits higher forecasting errors than SAAF, with particularly pronounced differences at short input lengths.

These findings suggest that simply adding an online predictor operating on the same restricted input window yields limited benefits, whereas incorporating HiPPO-based historical encoding enables SAAF to leverage information beyond the input window. Thus, the improvements of SAAF cannot be explained solely by the addition of an online predictor, supporting the contribution of HiPPO-based historical encoding.

Weakness 2: The paper provides limited validation of the universality of the claimed accuracy-efficiency trade-off. All key experiments used a fixed prediction length H = 64 and a rolling step size s = 8. In time-series forecasting, the prediction length significantly affects the difficulty of the online regression task and determines when historical observations can be used for updates. The rolling step size, in turn, affects the update frequency per unit time and the computational burden. Therefore, it remains unclear whether these conclusions can be generalized to other prediction horizons and feedback frequencies. Furthermore, the efficiency results depend on the base model. As noted in Section 5.8 and Appendix F.5, although more significant improvements were observed on Moirai and Lag-LLaMA, SAAF did not demonstrate notable improvements in the accuracy-efficiency tradeoff on TimesFM and Chronos-bolt. This limits the generalizability of the paper’s core efficiency claim. This is because the primary motivation of this paper is to improve the accuracy-efficiency tradeoff, rather than merely increasing prediction accuracy.

Question 2: How does SAAF perform under different prediction horizons and rolling window sizes? Does the proposed method remain effective when the prediction horizon is significantly extended?

Question 3: Could the authors further clarify under what conditions SAAF can improve the trade-off between accuracy and efficiency? Given that the gains obtained vary across different TSFM backbone networks, it would be of significant value to clarify how update costs—such as different numbers of channels (C) and different prediction lengths (H)—determine whether SAAF can provide a significant computational advantage.

Thank you for the clarification. We conducted an additional sensitivity analysis of the forecasting horizon HH and rolling step size ss on the ETT datasets, with the results reported in Appendix F.9. As for HH, forecasting errors generally increase with HH, as longer-horizon predictions are more challenging. Nevertheless, SAAF consistently maintains its advantage at short input window lengths LL, regardless of HH. Furthermore, when HH exceeds the backbone TSFM's supported prediction horizon, additional forecasts must be generated autoregressively, increasing inference costs. Therefore, we expect SAAF to offer greater benefits in the accuracy-efficiency trade-off as HH increases.

As for ss, on h1 and h2, SAAF improves forecasting accuracy across the evaluated values of ss. However, on m1 and m2, which have longer total timesteps TT, we observed substantial increases in error at s=1s=1 and s=2s=2. We hypothesize that smaller step sizes lead to more frequent online updates, potentially destabilizing Bayesian linear regression under our current hyperparameter settings. This instability may be mitigated by adjusting α\alpha and β\beta to reduce the magnitude of individual updates.

The revised experimental section addresses my concerns and strengthens the empirical evaluation. However, there is one issue: Figure 21 shows that performance drops significantly when the rolling step size is small, which is puzzling and remains unexplained. Given this unresolved concern, I maintain my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.