acceptodds
Under review as a conference paper at ICLR 2027

PROJECTION-CALIBRATED RESIDUAL ADAPTER FOR FROZEN TIME-SERIES FORECASTERS

Abstract

Time-series forecasters can leave systematic prediction errors that admit correction without modifying the frozen backbone. Effective correction requires estimating both the temporal shape of the residual and the strength with which it should be corrected. We introduce the Projection-Calibrated Residual Adapter (PCRA), which supervises these factors separately. PCRA learns orthogonal temporal directions to represent complementary error patterns and uses analytical projection targets to supervise their signed strengths. For a fixed unit direction, the residual projection minimizes reconstruction error; orthogonality extends this result to decoupled coefficient targets. Implemented with lightweight MLPs, PCRA also supports test-time adaptation by updating the entire strength gate using fully observed historical outcomes. A conditional risk analysis characterizes when predictable residual energy captured by the directions exceeds coefficient prediction error. Static experiments on six supervised forecasters show a mean MSE reduction of 4.52% across five benchmarks; additional results demonstrate applicability to four frozen foundation models.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Residual Correction, Frozen Forecasters, Projection-based Supervision, 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_ELksOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PCRA, a plug-in adapter that corrects the predictions of a frozen forecaster by decomposing its forecasting residual into a direction and a strength. The proposed method consists of a Direction Head, which predicts unit correction vectors, and a Strength Gate, which produces the correction strength multiplied with each unit correction vector. Through the residual projection onto a unit vector, the authors show that the optimal correction strength is determined by the product of the residual magnitude and the directional alignment (s∗=∥r∥cos⁡(r,d)s^* = \|r\|\cos(r, d)), and based on this, they convert strength learning into a regression problem on the projection coefficients. In addition, the orthogonality of the direction vectors is structurally enforced via Gram–Schmidt orthogonalization, which prevents the target of each strength from becoming entangled due to overlap between directions. The Direction Head takes the normalized context and the frozen forecaster's prediction as input and outputs normalized orthogonal unit vectors. It is trained sequentially on the training set with a cosine loss against the normalized deflated residual. The Strength Gate takes 22 statistical and channel features as input and outputs a signed strength for each direction. It is trained on the projection coefficients sk∗=⟨r,dk⟩{s_k}^* = \langle r, d_k \rangle with a standardized L1 loss offline and with an MSE loss at test time. In the experiments, the proposed method improves forecasting performance by an average of 4.52% over frozen supervised forecasters (6 backbones, 5 datasets) and by an average of 25.90% over frozen TSFMs (MSE, 4 TSFMs, 5 datasets). PCRA-T, which updates the Strength Gate at test time using delayed feedback, achieves MSE reductions of 0.00–4.71% on supervised forecasters and 22.15–48.93% on TSFMs relative to existing TTA methods.

Critical strengths
  • The paper systematically organizes the direction–strength decomposition, the projection-based strength targets, and the orthogonal decomposition, and connects them to a conditional risk decomposition, thereby clearly stating the conditions under which a correction and the addition of a direction are beneficial.
  • The method is compared against a number of baselines in both static adaptation and online TTA, and is shown to be applicable to 6 supervised backbones and multiple TSFMs without modifying the backbone weights.
  • The paper provides a broad set of analyses, including component ablations, the number of directions K, and online hyperparameters.
Critical weaknesses
  • W1 - Reproducibility: The Reproducibility Statement claims that "generation scripts reproduce every table and figure," but no source code (including an anonymous link) or supplementary material is provided, so implementation details such as normalization, window boundaries, and delayed-feedback handling cannot be verified. Moreover, (1) only the categories of the 22 input features of the Strength Gate are given (Appendix G.1), without a concrete list, (2) the lookback length L is not specified, and (3) the joint reconstruction refinement, described in Appendices A.4 and G.1 as the final stage of offline PCRA, is missing from Algorithm 1 and Sec. 3.3 and is described as "optional" in Appendix E.3, making the descriptions inconsistent. Therefore, it appears difficult to reproduce the reported results from the paper alone.

  • W2 - Statistical significance: The main results (Tables 1–4) and the ablations (Tables 5 and 14) report no confidence intervals, standard deviations across seeds, or repeated-run results, so the significance of small differences relative to the baselines cannot be assessed. Repeated runs are provided only for the K analysis at H=192 (Appendix D.1, Figure 10), and the number of seeds is not specified. In addition, the number of repeated-run tasks is described inconsistently between D.1 (15 tasks) and D.4 (12 per dataset, 11 for ETTm2).

  • W3 - Dataset coverage and lack of justification: The paper states that it follows the standard protocol (Zhou et al., 2021, Wu et al., 2021), but gives no reason for excluding Exchange Rate and Illness, which are included in that benchmark. Furthermore, the datasets used differ across experiments, mostly without explanation: Table 2 excludes Electricity and Traffic. Table 3 excludes Traffic, although Traffic results exist in Table 11, and the ablation and K analyses use only ETT and Weather.

  • W4 - Fairness of comparison: (1) For the Traffic results in Table 1, a validation-selected correction-norm cap is applied only to PCRA. Since there are no results without the cap, no results applying it to other datasets, and no results applying the same cap to the baselines, neither the effect of the cap nor the fairness of the comparison can be assessed. Furthermore, when the cap is active, the final correction is no longer Ds^D\hat{s}, so the equivalence in Eq. (4), which underpins PCRA-T, does not hold exactly on Traffic. (2) In the online comparison of Table 3, Ada-T is the only baseline with matched feedback conditions, the others retain different native protocols (partially observed targets, warm-up, optimizer, update/no-update selection, etc.) (Table 3 caption, Appendix E.2). PCRA-T ties with Ada-T, the feedback-matched baseline, on 2 of the 3 backbones. (3) Tables 3 and 4 do not include a static PCRA column, so the pure contribution of TTA cannot be identified. Recomputing by cross-referencing the tables, the additional gain over static PCRA is only about 0.6% on supervised forecasters. (4) The online hyperparameter sensitivity analysis (Figure 6b–d) was conducted on the test stream (Appendix E.4), while the validation sweep (Appendix E.4.1) was conducted with a different variant (output-layer, coefficient-MAE). Hence, the basis for selecting the formal PCRA-T settings is unclear.

  • W5 - Ambiguity of the zero-shot setting: In Table 2, the backbone TSFM is frozen, but the proposed method is trained with the training-split labels of each dataset. Therefore, the system as a whole can hardly be regarded as zero-shot. A rigorous zero-shot evaluation would require evaluating on data not used for training, as in GIFT-Eval, or conducting cross-dataset experiments, but no such experiments are provided, even in the appendix. Moreover, in a setting that likewise uses in-domain labels, the appropriate comparison groups should be fine-tuned TSFMs and other post-hoc methods (PIR, Ada-Y, UEC-STD), yet Table 2 reports only improvements over the frozen base.

  • W6 - Lack of interpretation of the ablation results: The values of each variant in Table 5 fluctuate substantially across datasets without a consistent trend. For example, w/o factorization yields 6.32% on Weather (the best) but −9.12% on ETTm2, and w/o cosine yields −9.82% on ETTh2 but 2.58% on ETTm1. The main text describes the wins and losses of each variant and adds only generic interpretations such as "regularization," "flexibility," and "reliability," without explaining the causes of the variation in relation to dataset characteristics. In addition, the large negative averages stem largely from a small number of long-horizon tasks. For example, w/o cosine on ETTh2 yields −34% to −39% at H=720, and excluding these gives about −1.2%. For w/o decoupled supervision, no task-level raw results (Table 14) are provided, so this cannot be verified.

  • W7 - Insufficient evidence for efficiency: The paper claims that the method is "lightweight," "without the computational burden … of parameter retraining," and that "updating backbone parameters is computationally prohibitive," but the main text contains no quantitative efficiency results, which are presented only to a limited extent in Appendix E.5. Table 16 and Figure 17 measure a 66-parameter output-layer control, and the runtime of the formal PCRA-T (802 parameters) used for the main results is not reported. Even the measured control is slower than Ada-Y (2.14 s vs. 1.67 s per 1,000 windows). Peak memory figures, runtimes of the other TTA baselines, and offline training costs are also not reported.

  • W8 - Choice of data for residual learning: The Direction Head and the Strength Gate are trained on residuals from the training split, on which the frozen forecaster has already been trained. These residuals may differ from the distribution of generalization errors at test time. Indeed, the baseline UEC (Nguyen et al., 2026) trains its corrector on a held-out validation set for this reason, whereas PCRA provides neither a discussion of the validity of using training-split residuals nor a comparative experiment.

  • W9 - Accuracy of presentation (minor): (1) Errors in reference descriptions: CycleNet is described as an RNN, TQNet as temporal-quantization-based, and TimesFM as a model supporting multivariate forecasting (Appendix A.2). The description of COSA conflicts between §2.2 and Appendix A.3, and UEC-STD is described as "Uncertainty-guided Error Correction," whereas the original paper names it the "Universal Error Corrector". The expansions of the acronyms PIR and TAFAS also differ from the original papers (Appendix A.3). (2) Notational inconsistencies: Ada-Y and Ada-T are used interchangeably in the online tables (Tables 4, 13, and 16). The legend of Figure 17 ("PCRA-T") is inconsistent with its caption (output-layer control). The set of ablation variants is inconsistent across Table 5, Table 14, Figure 7, and Appendix C.

Questions for authors
  • Q1: Could you release the source code via supplementary material or an anonymous GitHub repository? Could you also specify the concrete list of the 22 input features of the Strength Gate, the lookback length L, and whether the joint reconstruction refinement is included and what it updates (i.e., whether it includes the direction head)?
  • Q2: Could you provide the mean and standard deviation over multiple seeds, or the results of statistical significance tests, for the main results and the ablations? Please also clarify the number of seeds and the task composition of the repeated runs.
  • Q3: Why were Exchange Rate and Illness excluded? Could you provide results on these datasets, as well as the reasons for excluding some datasets in Tables 2 and 3?
  • Q4: Could you provide (1) the results of PCRA without the cap on Traffic and of the baselines with the same cap applied, (2) a direct comparison between static PCRA and PCRA-T, (3) an online comparison under unified feedback conditions, and (4) the selection procedure for the formal PCRA-T hyperparameters? Please also specify the initialization of TAFAS, PETSA, and COSA (the state of the adapter at the start of testing) in the baseline settings.
  • Q5: We understand that the proposed method does not assume a setting without target-dataset labels, such as cross-dataset evaluation or GIFT-Eval. Accordingly, would you be willing to revise the related descriptions from "zero-shot" to "without (instead of) fine-tuning"? In addition, could you provide comparisons with fine-tuned TSFMs and with PIR, Ada-Y, and UEC-STD?
  • Q6: Could you provide an analysis of the causes of the dataset-dependent variation in the ablation results, as well as the task-level results for w/o decoupled supervision?
  • Q7: Could you report the runtime and peak memory of the formal PCRA-T, a cost comparison with the other TTA baselines, and the offline training cost?
  • Q8: Could you provide a comparison with results obtained by training on validation-split residuals, as in UEC-STD, instead of training-split residuals?
  • Q9: PCRA-T uses only windows whose full horizon has been observed as feedback (Appendices E.1 and E.3), which incurs a delay of H steps and makes the feedback substantially stale for long horizons (e.g., H=720). Have you considered leveraging partially observed ground truth, as done by the baselines TAFAS and PETSA? Table 15 in the paper also describes this protocol as stronger than PCRA's matured-only boundary. In particular, the projection target s∗=D⊤rs^* = D^\top r requires the full residual, and DD restricted to the observed time steps is no longer orthonormal, so the decomposition in Eq. (4) no longer holds. Could you explain how the strength targets could be defined under partial observation (e.g., least-squares coefficients over the observed time steps), and provide results when such a scheme is applied?
Reviewer_mJwTOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper is concerned with what one could call "adapter fine-tuning" of a deep pretrained forecasting model, whose weights remain frozen. This is done by fitting a rather simple, MLP-based predictor to remaining residuals.

The paper recalls some very well-known facts about orthogonality and least-squares estimation, motivating a particular decomposition into orthogonal directions and scales for these correction adapters. Orthogonality is enforced posthoc, by running Gram-Schmidt on the outputs. I am not convinced this would make a principled difference over just not doing it.

What I find quite odd is that MLPs are being used as predictors. Sure, this is simple, but we are talking about forecasting here. Both context length L and horizon length H are different for every call. Also, there is some recursive structure to forecasts. None of this is captured by MLPs. In fact, if L gets really big (or also H), why would an MLP with some 32 or 64 hidden units be a sensible choice? Given the whole approach is rather simplistic, I'd at least expect some default parameterized simple forecasting model being employed here.

There is enough open source forecasting libraries out there the authors can use (such as Gluon-TS), so they won't even have to spend much extra work here.

Critical strengths

This is a simple approach to a problem that affects many users of pre-trained deep forecasting models. As long as they use it with smallish L, H, and few different values for these, the suggested method may work well enough. Given that, it is a simple baseline.

I am not an expert on this domain (i.e., adapter fine-tuning of neural forecasting models). The empirical evaluation looks solid to me, in the sense that quite some methods are compared against. I have reservations on the tasks and datasets, see below.

Critical weaknesses

The proposal is just very simple. The paper spends a lot of space recalling very well known facts about least-squares prediction and orthogonality, this would be linear algebra 101 really. Consider skipping this. Better spend more effort showing that this kind of decomposition actually pays off in practice.

I doubt the usefulness of MLPs in order to predict a residual. This is also a forecasting problem, which has temporal structure, and L, H vary over a large range. Maybe I am missing something (but then, the paper is not clear about it), but your input and output layer scales as L+H, but L and H are different every time. Or are you suggesting to train different MLPs for every (L, H)? In any case, MLPs are not useful, especially if L (and/or H) get large. A very obvious thing to do is to just pick a simple forecasting model instead and train its parameters.

The tasks and datasets look pretty toyish to me, but I think this is the same for other recent forecasting work. It is a bit of a shame, because there are really advanced forecasting problems out there, plus good neural approaches to them (e.g., DeepMind's work on weather forecasting; or other work on massively multi-variate problems in the energy domain), so I am wondering why work on forecasting is still being accepted despite authors not making much effort to work on really challenging problems.

Questions for authors

Most importantly, explain how MLPs apply given that L, H are different all the time. Explain why MLPs are sensible choices if the problem is forecasting. Why not just use a simple forecasting method?

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

Strengths

  • Simple and well-motivated method. Separating correction directions from their strengths is intuitive, and projection-based supervision gives each component a clear learning target. The resulting adapter can be added to an existing forecaster without retraining the backbone.
  • Relevant empirical validation. Experiments cover multiple forecasting architectures and datasets. Comparisons with a parameter-matched direct residual predictor, fixed directions, and joint training help assess the contribution of the proposed decomposition and supervision.
Critical strengths
  • Simple and well-motivated method. Separating correction directions from their strengths is intuitive, and projection-based supervision gives each component a clear learning target. The resulting adapter can be added to an existing forecaster without retraining the backbone.
  • Relevant empirical validation. Experiments cover multiple forecasting architectures and datasets. Comparisons with a parameter-matched direct residual predictor, fixed directions, and joint training help assess the contribution of the proposed decomposition and supervision.
Critical weaknesses
  • Limited methodological novelty. The approach builds on established residual adaptation and orthogonal projection techniques. The theoretical analysis primarily provides an algebraic error decomposition, without establishing why the proposed staged objectives should yield better predictions. Improvements over direct residual prediction and joint training also vary across datasets.
  • Adaptation protocols complicate interpretation. The foundation-model results involve adapters trained with target-dataset labels, so the complete system should be clearly distinguished from zero-shot forecasting. Moreover, online comparisons use different update protocols and budgets. Clearer accounting of offline supervision and matched adaptation budgets would make the reported advantages easier to assess.
Questions for authors

please see weakness above

Comments on the paper

2 comments

Dear Reviewers,

I'm writing to gently remind you that the rebuttal discussion phase deadline is approaching. Please take a moment to read the authors' rebuttal, and feel free to respond if any concerns remain that you'd like them to address. Thank you very much for your time and effort.

AuthorsOpenReview ↗

General Response

We thank the reviewers for their careful evaluations and constructive discussion. Your feedback helped us clarify the contribution of PCRA, examine how reliably it improves existing forecasts, and distinguish the gains from offline fitting and online adaptation.

1. The practical value of staged residual supervision (Reviewers wHKM, mJwT and ELks)

We clarified why PCRA learns correction directions before estimating their signed strengths. When both are learned jointly, small initial strength estimates can weaken the gradient used to learn the directions. PCRA therefore supervises the directions directly using forecast residuals, then uses cross-fitted directions to construct the Strength Gate’s training targets. We will make this procedure explicit in the revised algorithm. Across matched tasks, PCRA increased MSE relative to the frozen backbone less often than either direct residual prediction or joint training of the same adapter. The ablations also helped explain why the gains vary across datasets. On some datasets, the learned directions capture more of the residual, while on others, the gate predicts their strengths more accurately.

2. Simpler alternatives and broader forecasting benchmarks (Reviewers mJwT and ELks)

We tested whether a linear correction would suffice and extended the evaluation to higher-dimensional forecasting problems. PCRA achieved lower mean test MSE than the linear residual predictor on each dataset in the direct comparison. The results on electricity demand, traffic and solar power show that a corrector shared across channels can improve existing multivariate forecasts. Across different adapter seeds, both PCRA and PCRA-T retained an average improvement over the frozen backbone on each dataset in the repeated evaluation.

3. Target-domain supervision and backbone fine-tuning (Reviewers wHKM and ELks)

We clarified the role of target-domain labels and added a direct comparison with backbone fine-tuning. PCRA uses these labels to fit its adapter, so we will describe the setting as target-domain supervised adaptation with a frozen backbone. Across the evaluated datasets, PCRA consistently improved the frozen forecasts from TimesFM and Chronos. The fine-tuning comparison showed that adapting the forecast output can be competitive with updating the backbone on the full target training set, with the relative advantage varying by dataset.

4. The contribution and cost of online adaptation (Reviewers wHKM and ELks)

We distinguished the gains from offline fitting and test-time adaptation by reporting each method’s starting point and comparing online updates under matched feedback conditions and update budgets. Updating only the Strength Gate further reduced MSE relative to static PCRA for most backbone–dataset pairs. Following ELks’s suggestion, we also showed that updating from a partially observed forecast window can mitigate degradation when long horizons delay feedback. In our end-to-end evaluation on PatchTST, which included backbone inference for every method, PCRA-T had the lowest mean runtime and mean peak GPU memory usage among the five online adaptation methods.

Thank you so much for the time and care you have devoted to this work. We will incorporate the follow-up analyses and clarifications into the revised manuscript and appendix to make the paper clearer and more complete.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.