acceptodds
Withdrawn from ICLR 2027

DBCM: Dual-Branch Calibration Module for Test-Time Adaptation in Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Non-stationarity induces test-time distribution shifts in time series forecasting (TSF), resulting in forecast residuals between predictions and the corresponding ground truth. Recent works on test-time adaptation for TSF address this by introducing lightweight calibration modules around frozen forecasters, updated using observations that become incrementally available at test time. Existing approaches typically employ channel-wise adapters, implicitly assuming that forecast residuals are inherently channel-specific. However, we observe that forecast residuals often exhibit substantial shared structure across channels, suggesting that fully independent channel-wise parameterization may be redundant. Motivated by this observation, we propose the Dual-Branch Calibration Module (DBCM), which shares adaptation capacity across channels through a shared adapter while retaining limited channel-specific flexibility through a low-rank channel-wise adapter, further enhanced by channel mixing and nonlinear gated modulation. Extensive experiments show that DBCM consistently outperforms prior methods across diverse datasets, horizons, and backbones. Code will be released.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Time Series Forecasting, 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_MF2EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DBCM, which constructs a dual-branch calibration module for time-series forecasting by incorporating channel mixing and non-linear gating to enhance forecasting performance. DBCM also combines shared adapters with low-rank channel-wise adapters after channel mixing to produce the calibrated mixture sequence.

Critical strengths

The overall framework is reasonable, and the paper is well-structured. Each block in the DBCM module is explicitly explained. The methodology is evaluated and supported through extensive experiments and ablations.

Critical weaknesses
  1. Limited methodological novelty. The proposed framework is primarily built upon a combination of existing techniques, including low-rank, parameter sharing, and non-linear gating, lacking sufficiently distinct technical innovation. The paper does not explicitly distinguish the methodological contribution from alternative low-rank calibration architectures.
  2. The connection between the residual structure and DBCM requires further justification. Although the ablation in Section 5.4 shows the effectiveness of individual DBCM, it remains unclear why the low-rank structure of cross-channel residuals identified in Section 4 motivates sharing temporal transformation parameters in Eq. 3 rather than explicitly modeling the low-dimensional channel subspace. The authors should clarify this point.
Questions for authors

See weakness.

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

DBCM combines shared and low-rank channel-specific adapters to calibrate a frozen forecaster under test-time distribution shifts. Residual analysis and experiments across multiple backbones support sharing adaptation capacity while retaining limited channel-specific flexibility.

Critical strengths

Shared adaptation substantially reduces parameter count while preserving accuracy, and a small channel-specific branch provides further gains. Experiments across multiple backbones support its practical value.

Critical weaknesses
  • The reported spectral concentration could arise from unequal residual variances, even without cross-channel dependence. The analysis does not rule out this alternative, weakening the evidence for shared residual structure.
  • The method largely combines established adapter components, with limited additional insight into when shared versus channel-specific adaptation is beneficial.
Questions for authors
  • What evidence distinguishes shared cross-channel structure from unequal residual variances? Does the spectral concentration persist after standardizing residual variances or against a shuffled control that preserves each channel’s variance?
  • Across datasets, does stronger shared residual structure correspond to smaller gains from the channel-specific branch? Which results support this relationship?
Reviewer_4nqiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper works in the TAFAS test-time adaptation setting. Calibration modules around a frozen forecaster are updated online as ground truth is partially revealed. The authors measure the channel-space spectrum of the forecaster's residuals and find most energy in a few shared directions (EVR@30% of roughly 93% on ETT). They conclude that fully channel-wise adapters are redundant. DBCM therefore combines a channel mixer, a temporal map shared by all channels, a per-channel rank-8 LoRA correction and a sigmoid gate, and it uses the same block on both sides of the forecaster. Over 144 settings, DBCM reduces MSE by 6.20% on average, against 4.07% for TAFAS and 2.61% for PETSA.

Critical strengths

S1. The paper clearly identifies the limitation of earlier adapters, grounds its design in a measurement of residual structure, and delivers a module that is both more accurate and much smaller than TAFAS. The paper measures shared error structure with explained-variance ratios and subspace alignment, and it applies the same diagnostics to the baselines. A single shared adapter matches TAFAS (0.3530 vs 0.3531, Fig 3a) with 0.18M instead of 1.70M parameters. The diagnostic is valuable for this setting well beyond this particular method.

S2. The main evaluation is broad, leakage-free, and consistent across backbone families. DBCM has the higher average improvement on every backbone in Table 2 and is the best method in all but four of the 144 cells. The released code confirms that full-horizon targets are used only after the horizon has elapsed, and that the partial loss is masked to the steps already observed. No forecast is adapted on values from its own future, so the gains are credible as genuine test-time results.

Critical weaknesses

W1. The share of the margin over TAFAS and PETSA that comes from the architecture, rather than from the objective and tuning, is not yet isolated. The headline gap is about two points (6.20% vs 4.07%). DBCM's partial loss averages the revealed steps of every window, about 300 step pairs, while the released baseline code supervises only the first window's 24 pairs (tta/tafas.py, line 253). This is not a protocol violation, but the difference is not yet disclosed or ablated. In the released scripts, PETSA's gating initialization, rank and alpha are fixed across all 144 settings, while DBCM's learning rate differs from TAFAS's in 88. The selection split is not specified, and the LoRA rank sweep (Fig 3b) is reported on test MSE. Settling this point is the main step toward a clear contribution.

W2. The shared branch is shown to buy parameter efficiency; its contribution to accuracy, and theoretical backing for the low-rank premise, are not yet shown. In Fig 3a the shared adapter gives parity with TAFAS (0.3531 to 0.3530), and the accuracy gains come from the LoRA branch, the mixer and the gate. The ablation is incremental only, is run on iTransformer alone, and includes no variant without the shared branch. App G treats residual low-rankness as an empirical hypothesis, and no result yet links the degree of concentration to the size of the gain. Traffic (C=862), which would stress both the premise and the C x C mixer, is not evaluated. Because the components are individually standard and follow the TAFAS template, the conceptual step beyond TAFAS and PETSA is modest until this link is shown.

W3. The comparison with the most recent methods is not yet brought into the main protocol or the cost analysis. The 144-cell comparison uses only TAFAS and PETSA, both from 2025. DynaTTA, COSA and DynaME appear only in Appendix C on three or fewer backbones, and DynaTTA as run is worse than no adaptation (0.4977 vs 0.3687 on iTransformer). COSA is not run under the paper's protocol; instead, DBCM is moved into COSA's protocol (Table C.4). The efficiency analysis (Tables C.7 to C.9) includes only TAFAS, PETSA and DBCM. ADAPT-Z (Huang et al., ICLR 2026) is not yet cited or compared, and ELF (Lee et al., ICML 2025) is cited but not compared, although Table C.2 is exactly its setting.

W4. Reproduction from the paper alone is not yet possible, and the released material covers the main table but not the appendix. No anonymous repository is provided, and code is promised upon publication. The paper omits the selected per-setting hyperparameters and the initialization of the module's weights. Both matter here, because in the supplement the module does not start at the identity and the learning rate and gate initialization are tuned per setting. The supplementary code reproduces the 144-cell main table, which is valuable, but not the appendix experiments, including the comparisons with recent methods and the efficiency measurements.

Minor points, listed separately from the weaknesses above and not weighed in the assessment:

  1. The OneNet reference should be cited as Zhang et al., 2023, not with Qingsong Wen as first author.
  2. Sec 5.2 reports 4.07% for TAFAS while the per-backbone values in Table 2 average to 4.08%; please harmonize the rounding.

Further requests, not decisive for the assessment:

  • Please explain why DBCM falls below the frozen forecaster at ETTh2-192 and ETTm2-720 on all three backbones in Table C.4, and whether it was re-tuned for COSA's protocol.
  • Please explain the zero-shot ETTm1 numbers for Chronos (1.5167) and Moirai (1.0543) in Table C.2, and extend that experiment beyond H=96.
  • Please report paired significance tests and additional seeds on Exchange at long horizons, where DBCM's seed standard deviation reaches 0.1460 at Exchange-720.
Questions for authors

Q1 (W1). A loss-matched comparison would settle the main question. Please report it on a representative subset of Table 2, training TAFAS and PETSA with DBCM's all-window loss or DBCM with the first-window loss, and state which loss each row of Fig 3a uses. If DBCM keeps most of its two-point margin, the gain belongs to the architecture. If the margin largely disappears, the contribution lies in the training objective rather than the dual-branch design.

Q2 (W1). On which split, and by which criterion, were hyperparameters, including the LoRA rank of Fig 3b, selected? Please report the selected values for each method and re-tune PETSA under the grid described in App E. Selection on a held-out split and a lead over PETSA that holds after re-tuning would resolve this part of W1.

Q3 (W2). A leave-one-out ablation of each component from the full model, on a backbone other than iTransformer, would show what drives accuracy. Please also relate residual concentration to DBCM's gain over TAFAS across datasets. A clear drop without the shared branch, and gains that track concentration, would show that the shared low-rank branch drives accuracy. Otherwise, its value lies in parameter efficiency, which is itself a useful result.

Q4 (W3). Can COSA's adapter be run under the main protocol on the six main backbones, and can DynaTTA, COSA and DynaME be added to Tables C.7 to C.9? A retained lead over COSA and a better accuracy-efficiency trade-off against these methods would resolve W3.

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

Related papers

Discussion (0)

Sign in to comment.