acceptodds
Under review as a conference paper at ICLR 2027

Learning Where Frozen Forecasters Fail: Residual-Based Retrieval Augmentation for Time-Series Forecasting

Abstract

Retrieval-augmented generation (RAG) supports time-series forecasting by re- trieving historical trajectories and representations to inform predictions. However, the potential of historical evidence to refine a forecaster’s existing predictions by targeting its remaining errors remains underexplored. Such refinement builds on the temporal patterns already captured by the forecaster, while its histori- cal residuals provide model-specific evidence about the direction and magnitude of past errors. Using RAG to augment forecasting via residuals presents two challenges: identifying relevant historical errors when the current residual is un- observed, and determining their correction strength when historical and current prediction conditions differ. To address these challenges, we propose Residual- based Retrieval-Augmented Forecasting of Time-series (ResRAF), which organizes a frozen forecaster’s historical errors into a residual-based knowledge base. Its Residual-Aware Retriever learns from historical residual supervision to predict residual pattern representations from the current input and base forecast, then combines residual proximity with forecast-shape similarity to retrieve complete historical residual sequences. State-Adaptive Retrieval Calibration adjusts the con- tributions of these sequences according to the prediction states of both the current and historical instances, derived from observable input and forecast characteristics. The calibrated residuals are aggregated and added to the base forecast, with the forecaster remaining frozen and all additional components fitted on training data and fixed during evaluation. Experiments across multiple datasets, forecasting architectures, and prediction horizons show that ResRAF outperforms existing methods in forecasting accuracy.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series Forecasting, Retrieval Augmented Generation, Residual Calibration, Forecast Correction, Residual Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xHwHOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose ResRAF, a framework that uses historical forecast errors in a retrieval-augmented setting to correct the predictions of existing forecasting models. The authors empirically demonstrate that applying the framework to several supervised models and a time-series foundation model improves forecasting accuracy.

Critical strengths

1. Easy to use. The framework is applied post hoc. It is evaluated on a diverse set of backbones (linear, Transformer-based, retrieval-based, multiscale), including a TSFM.

2. Sensible ablations. The ablations are informative: random retrieval removes almost all of the gain, dropping the residual similarity term removes most of it, and state-adaptive calibration mainly reduces the number of degraded tasks. I also appreciate that the authors report the median improvement, the number of degraded tasks, and worst-case performance rather than only averages, and that they specify which data each fitting stage uses.

Critical weaknesses

1. Limited methodological novelty for a complex pipeline. Unfortunately, the central idea seems to have substantial overlap with existing literature. Retrieving historical forecasts that resemble the current forecast and using their corresponding observations or errors for post-processing is well established in analog-based weather forecasting (Delle Monache et al., 2011; 2013). Nearest-neighbor methods for updating existing forecasts have also been studied in hydrology (Brath et al., 2002), while regression-based and regime-dependent forecast calibration have a long history (Glahn & Lowry, 1972; Allen et al., 2019). More directly, ResMem (Yang et al., 2023) already proposes fitting a nearest-neighbor regressor to the residuals of a base predictor and adding the estimated residual to its output. Although ResMem was primarily evaluated outside time-series forecasting, the underlying residual-correction formulation is particularly close to ResRAF.

I expected a more detailed discussion of this literature. The predicted residual representation as an additional retrieval key and the state-dependent correction coefficients may be useful contributions, but relative to these established ideas, the methodological novelty appears rather incremental. At the same time, the framework combines randomized SVD, a CNN-based residual predictor, K-means clustering over 78 handcrafted features, a clipped ridge regression fit, and several manually chosen constants. While each component may be motivated individually, I am not convinced that the reported performance gains justify the complexity of the complete pipeline. Moreover, there is no analysis of fitting time, memory consumption, or additional inference latency. This is especially important for a retrieval-based post-processing method: demonstrating low computational overhead would substantially strengthen its practical contribution and help justify the added complexity.

2. Limited experimental evaluation. The evaluation is restricted to six relatively low-dimensional datasets (7–21 channels), without experiments on higher-dimensional benchmarks such as Electricity or Traffic. I am also missing stronger comparisons that isolate the contributions of the proposed retrieval mechanism. The paper mainly compares its retrieval score against random retrieval and forecast-shape similarity alone. Random retrieval is a useful sanity check, but it is not a particularly competitive alternative.

There are several informative alternatives that are not evaluated: retrieval based on input-window similarity, as in RAFT (Han et al., 2025); retrieval in the frozen forecaster's representation space; simpler nearest-neighbor residual correction, as in ResMem; and stationarity-aware retrieval strategies, as in SARAF (Zhou et al., 2026). These methods are not all directly interchangeable with ResRAF, but comparisons or controlled adaptations would help establish which parts of the proposed retrieval mechanism are necessary.

In particular, there is no baseline that directly decodes the predicted residual representation into a forecast correction without retrieval. Such a comparison would clarify whether the retrieval stage provides additional value beyond the learned residual predictor itself.

I am also concerned by the performance of the two included correction baselines. UEC-STD reportedly increases MSE by approximately 4%–18% on ETTh1, while PIR increases MSE by up to 21.7% on Exchange. These degradations raise questions about the comparability of their configurations and tuning budgets.

Finally, ResRAF fits additional components using target-domain data, so the resulting Chronos-based system is not strictly zero-shot, even though the underlying Chronos model remains frozen. A comparison with lightweight target-domain adaptation or parameter-efficient fine-tuning would therefore be helpful. A simpler alternative would also be to retrieve similar historical samples and oversample them during fine-tuning of the model or its prediction head. SOLID (Chen et al., 2024) follows a related strategy by selecting contextually similar historical samples and adapting a model's prediction layer, but it is not compared. Such an experiment would help determine whether the proposed multi-stage correction pipeline offers a meaningful advantage over more direct adaptation.

Questions for authors

Clarification requests

  1. How does ResRAF compare with a simple ResMem-style nearest-neighbor residual corrector, alternative retrieval keys (e.g., input-window or backbone-representation similarity), and direct decoding of the predicted residual representation without retrieval? I am particularly interested in whether the additional residual-similarity retrieval stage improves over a direct learned correction.

  2. Have the authors considered selective fine-tuning on retrieved historical samples, as in SOLID, as a simpler alternative? For the Chronos experiments, how would ResRAF compare against lightweight target-domain adaptation or parameter-efficient fine-tuning?

  3. How were UEC-STD and PIR configured and tuned? Can the authors explain their relatively large degradations on ETTh1 and Exchange, and whether comparable tuning budgets were used?

  4. How are historical residuals in the retrieval memory generated? In particular, are they computed from in-sample predictions or out-of-sample/rolling-origin forecasts? How is it ensured that both the retrieved samples and their corresponding future labels are available at each evaluation forecast origin?

  5. How do fitting time, memory requirements, and inference latency scale with the size of the retrieval memory and the number of channels? Results on higher-dimensional datasets such as Electricity or Traffic would be valuable.

Suggestions for presentation

  • With Tw=106T_w = 10^6, the retrieval weights in Equation 9 appear to be approximately uniform, assuming the similarity scores have moderate magnitude. Could the authors clarify whether this weighting has a meaningful numerical effect? If not, a plain average could simplify the method description.

References

  • Allen, Ferro & Kwasniok (2019). Regime-dependent statistical post-processing of ensemble forecasts. Quarterly Journal of the Royal Meteorological Society.
  • Brath, Montanari & Toth (2002). Neural networks and non-parametric methods for improving real-time flood forecasting through conceptual hydrological models. Hydrology and Earth System Sciences.
  • Chen et al. (2024). Calibration of Time-Series Forecasting: Detecting and Adapting Context-Driven Distribution Shift (SOLID). KDD.
  • Delle Monache et al. (2011). Kalman Filter and Analog Schemes to Postprocess Numerical Weather Predictions. Monthly Weather Review.
  • Delle Monache et al. (2013). Probabilistic Weather Prediction with an Analog Ensemble. Monthly Weather Review.
  • Glahn & Lowry (1972). The Use of Model Output Statistics (MOS) in Objective Weather Forecasting. Journal of Applied Meteorology.
  • Han et al. (2025). Retrieval Augmented Time Series Forecasting (RAFT). ICML.
  • Yang et al. (2023). ResMem: Learn what you can and memorize the rest. NeurIPS.
  • Zhou et al. (2026). Stationarity-Aware Retrieval-Augmented Time Series Forecasting (SARAF).
9 comments

Q1: How does ResRAF compare with a simple ResMem-style nearest-neighbor residual corrector, alternative retrieval keys such as input-window or backbone-representation similarity, and direct decoding of the predicted residual representation without retrieval? In particular, does residual-aware retrieval improve on a direct learned correction?

A1: We thank the reviewer for suggesting these comparisons. We evaluate ResMem, three alternative retrieval strategies, and direct decoding with and without SARC on iTransformer across six datasets and four prediction horizons, as summarized in Table A.

Our ResRAF combines residual-aware retrieval (RAR) with state-adaptive residual calibration (SARC). To assess RAR, we replace its retrieval strategy with backbone-representation similarity (ResMem-style + SARC), input-window similarity (RAFT-style + SARC), or SARAF-inspired stationarity and temporal-structure criteria (SARAF-style + SARC), while keeping the aggregation and SARC unchanged. We also compare with the original ResMem, which retrieves historical residuals using backbone representations and applies distance-weighted correction. To assess the value of retrieval itself, Direct uses the same trained residual-representation predictor as ResRAF but decodes its output into a correction without retrieval. Direct + SARC further calibrates this correction.

Table A. Residual correction and retrieval alternatives on iTransformer: relative MSE change.

DatasetResRAF(Ours)ResMemResMem-style + SARCRAFT-style + SARCSARAF-style + SARCDirectDirect + SARC
ETTh1−1.09%−0.53%−0.20%−2.38%−1.40%+6.63%+4.11%
ETTh2−2.07%−0.95%−1.10%−0.99%−2.80%−1.81%−1.68%
ETTm1−2.07%−1.88%−1.47%−2.03%−2.29%−1.11%−1.83%
ETTm2−2.48%−1.87%−2.01%+3.48%−1.27%−1.81%−1.91%
Exchange−8.22%+13.42%+1.58%+1.63%+1.81%−7.33%−6.25%
Weather−6.32%−2.42%−4.05%−4.70%−3.65%−6.91%−8.32%
Average−3.71%+0.96%−1.21%−0.83%−1.60%−2.06%−2.65%

Note: Each dataset result averages relative MSE changes across four prediction horizons. Negative values indicate improvement over the same frozen iTransformer forecast. Bold indicates the best result in each row.

Our ResRAF achieves the best overall performance among the alternative retrieval strategies, with an average ΔMSE of −3.71%, compared with −1.60% for the strongest alternative (SARAF-style + SARC). This highlights the importance of retrieval criteria and supports the effectiveness of our residual-aware retrieval (RAR). Furthermore, comparing ResMem with ResMem-style + SARC shows that aggregation and calibration help mitigate negative transfer. The average ΔMSE improves from +0.96% to −1.21%, while the Exchange degradation drops from +13.42% to +1.58%. However, ResRAF still outperforms ResMem-style + SARC, indicating that improved aggregation and calibration alone cannot fully compensate for less effective retrieval. Finally, compared with Direct and Direct + SARC, ResRAF achieves better average results on five of the six datasets. This supports the benefit of retrieving historical residuals over directly decoding predicted residual representations.

Implementation details. All methods share the same frozen iTransformer, cached base forecasts, and training and validation splits. For backbone-representation retrieval, we extract the first CC variable tokens after the encoder's final LayerNorm and before projection. The original ResMem selects its retrieval size and distance-weighting temperature based on validation MSE. For direct correction, we freeze ResRAF's trained residual-representation predictor and train an MLP decoder to predict the full HH-step residual, selecting its checkpoint using validation MSE. Direct + SARC additionally fits calibration coefficients using training data.

We will incorporate these additional comparisons and the corresponding analysis into the revised manuscript.

Q2: Have the authors considered selective fine-tuning on retrieved historical samples, as in SOLID, as a simpler alternative? For the Chronos experiments, how would ResRAF compare against lightweight target-domain adaptation or parameter-efficient fine-tuning?

A2: We appreciate the reviewer's suggestion to compare ResRAF with selective fine-tuning and lightweight target-domain adaptation. Inspired by your feedback, we conducted additional experiments with SOLID and Head-only FT on Chronos-Bolt-Small across 6 datasets at a prediction length of 96. For SOLID, we follow its similarity-based retrieval and sample-wise gradient updates, using a causal online pool where retrieved targets are fully observed before each forecast origin. For Head-only FT, we fine-tune the output module (5.29% of Chronos parameters) on the target-domain training set. All methods use the same pretrained checkpoint, data splits, context length (512), and validation-based hyperparameter selection. The results are summarized in Table B.

Table B: Relative MSE/MAE Changes on Chronos-Bolt-Small at a Prediction Length of 96. Negative values indicate improvements over the original Chronos predictions.

DatasetMetricResRAF(Ours)SOLIDHead-only FT
ETTh1MSE-1.02%-3.31%-3.82%
MAE+0.79%-1.05%-0.52%
ETTh2MSE-1.06%0.00%-5.32%
MAE+0.62%0.00%+0.31%
ETTm1MSE-7.62%-0.61%-3.05%
MAE-1.81%+0.30%-0.90%
ETTm2MSE-1.72%0.00%-1.15%
MAE0.00%0.00%0.00%
ExchangeMSE-9.20%+1.15%+2.30%
MAE-4.35%0.00%+1.45%
WeatherMSE-11.05%-1.16%-6.98%
MAE-8.25%-0.49%-0.97%
AverageMSE-5.28%-0.66%-3.00%
MAE-2.17%-0.21%-0.11%

As shown in Table B, our ResRAF improves MSE on all 6 datasets and achieves the largest average reductions in both MSE (5.28%) and MAE (2.17%), demonstrating its consistent effectiveness across different forecasting scenarios.

Two observations further support the advantage of our approach. (1) SOLID can leverage newly observed samples and perform query-specific gradient updates during testing, whereas our ResRAF relies exclusively on a fixed training-set residual memory. Nevertheless, our ResRAF achieves substantially larger average improvements in both MSE (5.28% vs. 0.66%) and MAE (2.17% vs. 0.21%). This shows that our ResRAF achieves stronger improvements without relying on online adaptation. (2) Although Head-only FT outperforms our ResRAF in terms of MSE on ETTh1 and ETTh2, its effectiveness does not consistently extend across datasets. On Exchange, both Head-only FT and SOLID exhibit negative transfer, increasing MSE by 2.30% and 1.15%, respectively, whereas our ResRAF reduces MSE by 9.20%. This demonstrates the robustness of our ResRAF under distribution shifts, where parameter-based adaptation may suffer from negative transfer.

We will incorporate these additional comparisons and discussions into the revised manuscript.

Q3: How were UEC-STD and PIR configured and tuned? Can the authors explain their relatively large degradations on ETTh1 and Exchange, and whether comparable tuning budgets were used?

A3 (part 1): We sincerely thank the reviewer for raising these important concerns about the fairness of our baseline comparisons. We carefully configure and tune UEC-STD and PIR with reference to their original implementations. Below, we clarify the tuning procedures and provide additional evidence to explain the observed degradations.

Configurations and tuning. For UEC-STD , to ensure a fair comparison, we adapt it from autoregressive to direct multi-step forecasting. Its original autoregressive rollout would produce different base forecasts and introduce additional error accumulation, making it difficult to isolate the effects of the correction methods. We therefore apply UEC-STD to the same frozen forecasts as ResRAF, while preserving its original trend–seasonal decomposition, correction architecture, training objective, and hyperparameter settings. In particular, its correction strength β\beta is selected independently for each task from the original grid 0.01,0.03,0.05,0.07,0.1,0.3,0.5,0.7,1.0\\{0.01,0.03,0.05,0.07,0.1,0.3,0.5,0.7,1.0\\} using the prescribed balanced validation pool. For PIR , we use K=50K=50 from the original paper's search space 10,20,50\\{10,20,50\\}. Additionally, all three retrieval sizes are evaluated, with KK independently selected based on validation loss for each of the 120 tasks (5 backbones, 6 datasets, and 4 horizons). As shown in Table C, K=50K=50 performs best among the three fixed settings on 75/120 tasks and achieves the lowest average MSE degradation. However, validation-based tuning still fails to prevent negative transfer on Exchange. For our proposed ResRAF , we similarly tune the retrieval size KK and regularization coefficient λ\lambda on validation data. Although the search spaces differ across methods, all hyperparameters are selected using validation data, and no test labels are used for tuning.

Table C: PIR Retrieval-Size Comparison (Relative MSE Change).

DatasetK=10K=20K=50Tuned KResRAF(Ours)
ETTh1+0.87%-0.06%-0.16%+0.36%-1.47%
ETTh2+2.81%+1.29%-0.13%-0.24%-1.71%
ETTm1-0.13%-0.50%-0.56%-0.75%-2.04%
ETTm2+4.88%+3.20%+3.58%+2.25%-2.46%
Exchange+70.01%+23.75%+9.36%+25.29%-10.17%
Weather+2.71%+0.19%-0.76%-0.69%-5.83%
Average+13.52%+4.65%+1.89%+4.37%-3.95%

Note: Results report average relative MSE changes across 5 backbones and 4 horizons. Negative values indicate improvement. Tuned K is selected per backbone–dataset–horizon combination using validation loss. Bold indicates the best result.

Reviewer_cA8POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ResRAF, a post-hoc correction framework for frozen time-series forecasters. It learns to predict low-dimensional residual representations from historical inputs and base forecasts, retrieves historical residual sequences using residual and forecast similarity, and calibrates their contributions according to current and historical prediction states. The resulting correction is added to the base forecast. Experiments cover six datasets, four forecast horizons, five trained backbones, and Chronos-Bolt-Small. The reported results show broad accuracy improvements, while ablations suggest that residual-aware retrieval drives average gains and calibration reduces negative transfer.

Critical strengths
  • The paired evaluation provides credible evidence of practical usefulness. Each corrected forecast is compared with the exact frozen checkpoint that generated the base prediction and historical residuals. Improvements across several architectures, datasets, and horizons support the usefulness of model-specific historical errors for forecast correction.

  • The evaluation examines negative transfer alongside average accuracy. Table 4 reports that calibration reduces degraded tasks from 29/120 to 8/120 and lowers the worst observed degradation from 12.2% to 2.0%. This is practically consequential because correction modules can otherwise harm forecasts even when their average performance improves.

Critical weaknesses
  • Limited methodological novelty and insufficient evidence of improvement over closely related approaches. The central formulation—augmenting a base predictor by retrieving historical residuals and adding a residual correction—is closely related to ResMem, as acknowledged in the paper. The main distinctions appear to lie in predicting SVD-based residual representations for retrieval and applying state-dependent calibration. These are plausible extensions, but the paper does not convincingly establish the extent of the methodological advance. In particular, it lacks a direct comparison with a ResMem-style residual nearest-neighbor baseline and with direct decoding of the predicted residual representation into a correction. The current ablations establish the usefulness of components within ResRAF, but do not adequately demonstrate an advance over simpler residual-learning and residual-memory approaches. Consequently, both the novelty and the incremental value of the proposed framework remain insufficiently supported.

  • The distinctive contribution of state-dependent calibration is not isolated. SARC combines clustering of observable features with regularized, bounded linear calibration coefficients. Its methodological contribution therefore depends substantially on demonstrating that conditioning on current and historical states provides benefits beyond ordinary correction-strength control. However, the ablation replaces SARC with an unscaled residual average, simultaneously removing state conditioning, shrinkage, regularization, and coefficient bounds. The reduction in negative transfer could therefore arise from more conservative correction strengths. Given that Table 4 reports approximately the same mean MSE improvement with and without SARC, comparison against a fitted channel-wise scalar correction is particularly important for establishing the value of the proposed state-dependent design.

Questions for authors
  • How does ResRAF compare with directly decoding the predicted residual representation into a correction, using the same trained predictor? How does it compare with a ResMem-style residual nearest-neighbor baseline under comparable training and validation-selection protocols?

  • Does SARC retain its advantage in degraded-task frequency and worst-case degradation over a fitted channel-wise scalar correction? Current-state-only and historical-state-only variants would further clarify the contribution of each conditioning term.

  • How was UEC-STD adapted regarding direct versus autoregressive forecasting, correction-training data, and correction-strength selection? These details could materially affect the interpretation of the baseline comparison.

6 comments

Q1: How does ResRAF compare with directly decoding the predicted residual representation into a correction, using the same trained predictor? How does it compare with a ResMem-style residual nearest-neighbor baseline under comparable training and validation-selection protocols?

A1: We appreciate the reviewer's constructive suggestions and introduce two additional baselines, Direct decoding and ResMem, to assess the benefits of our proposed framework.

For Direct decoding, we use the same trained residual-representation predictor as ResRAF but directly decode its output into a residual correction without retrieval. For ResMem, we retrieve historical residuals using frozen backbone representations and aggregate them with distance-based weights. In contrast, our ResRAF predicts low-dimensional SVD residual representations for retrieval and calibrates the retrieved residuals according to current and historical prediction states.

We evaluate these methods across 5 base forecasters, 6 datasets, and 4 prediction horizons, as summarized in Table A.

Table A: ResRAF vs. Direct Decoding and ResMem (Relative MSE Change, %).

BackboneMethodETTh1ETTh2ETTm1ETTm2ExchangeWeatherAverage
DLinearResRAF (Ours)−3.83−2.26−5.85−6.73−7.02−11.68−6.23
Direct+1.73−13.05−6.28+0.63−0.10−11.19−4.71
ResMem+0.30+38.32−1.35+88.66+3.02−2.84+21.02
iTransformerResRAF (Ours)−1.09−2.07−2.07−2.48−8.22−6.32−3.71
Direct+6.63−1.81−1.11−1.81−7.33−6.91−2.06
ResMem−0.53−0.95−1.88−1.87+13.42−2.42+0.96
OLinearResRAF (Ours)−0.79+0.23−1.35−0.90−9.10−2.81−2.45
Direct+5.83+1.68−1.78−0.86−10.06−2.64−1.30
ResMem−0.68+1.23+0.01−0.10+7.46+0.23+1.36
TimeMixerResRAF (Ours)−1.52−1.47+0.21−1.70−15.70−2.09−3.71
Direct+6.59−1.33+0.04−1.09−13.00−1.15−1.66
ResMem+0.05−1.35+0.27+1.60−2.34+1.47−0.05
RAFTResRAF (Ours)−0.10−3.01−1.12−0.50−10.80−6.25−3.63
Direct+5.68−2.13−0.16−0.67−14.35−4.38−2.67
ResMem−0.97+3.02−2.64−1.81+5.91−6.61−0.52

Note: Results are averaged across prediction horizons. Values are relative MSE changes (%) from the corresponding frozen forecasts; negative values indicate improvement. Bold indicates the best result for each backbone–dataset pair and backbone average.

Our ResRAF achieves the best average MSE performance across all 5 backbones and exhibits fewer cases of negative transfer. Compared with Direct decoding, ResRAF performs better in 22/30 dataset–backbone combinations, supporting the benefit of retrieving and calibrating historical residuals rather than directly decoding predicted residual representations. Compared with ResMem, ResRAF performs better in 26/30 combinations. Notably, ResMem exhibits substantial degradations on DLinear–ETTm2 (+88.66%) and DLinear–ETTh2 (+38.32%). To investigate this, we apply SARC to ResMem's retrieved residuals while keeping its retrieval mechanism unchanged. This reduces the degradations to +41.37% and +27.21%, respectively. These results show that adaptive correction strength can substantially reduce negative transfer, but cannot fully compensate for limitations in retrieved residuals. This highlights the complementary roles of residual-aware retrieval and state-adaptive calibration. Overall, average MSE degrades in only 2/30 combinations for ResRAF, compared with 8/30 for Direct and 15/30 for ResMem. These findings support the value of our framework in improving forecasting accuracy while mitigating negative transfer.

Implementation and evaluation details. All methods use identical frozen base forecasters and data splits, with training conducted on training data and model selection based on validation performance. For the ResMem comparison specifically, we follow the original paper's validation-based hyperparameter selection protocol. We jointly tune K∈1,4,16,53,64,96,160,320,512,1024K\in\\{1,4,16,53,64,96,160,320,512,1024\\} and σ∈0.1,0.3,0.7,1,3,10,106\sigma\in\\{0.1,0.3,0.7,1,3,10,10^6\\} using validation forecast MSE. Both ResRAF and ResMem use training-only residual memories, and ResRAF retains its original validation-based hyperparameter selection.

Following the reviewer's suggestion, we also incorporate these additional comparisons into Table 3 of the revised manuscript.

Q2: Does SARC retain its advantage in degraded-task frequency and worst-case degradation over a fitted channel-wise scalar correction? Current-state-only and historical-state-only variants would further clarify the contribution of each conditioning term.

A2: Thank the reviewer for suggesting these controls. We include a fitted channel-wise scalar baseline and two single-state variants to assess the benefits of state-dependent calibration beyond a fitted correction strength.

In full SARC, the transfer coefficient is γi,jc=αgi,cc+δgjc\gamma^c_{i,j}=\alpha^c_{g_{i,c}}+\delta^c_{g_j}, where α\alpha controls the overall correction strength based on the current prediction state, and δ\delta adjusts each retrieved residual's contribution based on its historical state. The w/o State-Adaptive Transfer variant sets all transfer coefficients to one, directly averaging the retrieved residuals. Channel-wise Scalar replaces both terms with one state-independent coefficient per channel. Current-state Only SARC retains α\alpha with δ=0\delta=0, while Historical-state Only SARC retains δ\delta with α=0\alpha=0.

Table B: Ablation Results Across 120 Tasks.

VariantMean ΔMSEMedian ΔMSEMean ΔMAEMedian ΔMAEDegraded TasksWorst ΔMSEResRAF Wins
ResRAF (Ours)−3.95%−2.39%−2.43%−1.31%8/120+2.00%—
w/o State-Adaptive Transfer−3.89%−2.04%−2.36%−0.79%29/120+12.17%89/120
Channel-wise Scalar−3.84%−2.32%−2.43%−1.18%10/120+2.67%87/120
Current-state Only SARC−3.88%−2.35%−2.42%−1.40%7/120+3.06%75/120
Historical-state Only SARC−0.73%−0.55%−0.50%−0.34%4/120+0.21%109/120

Note: ΔMSE and ΔMAE are relative to the corresponding frozen base forecasts; negative values indicate improvement. Degraded Tasks counts cases with increased MSE, Worst ΔMSE reports the largest MSE increase, and “ResRAF Wins” counts tasks where ResRAF achieves lower MSE than the corresponding variant. Bold indicates the best value in each column, including ties.

As shown in Table B, Full SARC retains an advantage over the fitted channel-wise scalar in both degraded-task frequency and worst-case degradation. Compared with Channel-wise Scalar, it reduces degraded tasks from 10/120 to 8/120 and worst-case MSE degradation from 2.67% to 2.00%. It also achieves a greater mean MSE reduction (−3.95% vs. −3.84%) and lower MSE on 87/120 tasks. The fitted scalar itself substantially reduces negative transfer compared with uncalibrated residual averaging, showing the importance of learning the correction strength.

The single-state variants reveal the complementary roles of the two conditioning terms. Current-state Only achieves similar average accuracy and has slightly fewer degraded tasks (7 vs. 8), but its worst-case degradation is larger (+3.06% vs. +2.00%). Historical-state Only achieves the lowest negative-transfer risk, with only 4 degraded tasks and a worst-case increase of 0.21%, but its average MSE improvement is much smaller (−0.73% vs. −3.95%). These results show that the current-state term controls the overall correction strength, while the historical-state term adjusts the contributions of individual retrieved residuals. Combining both allows SARC to preserve substantial forecasting gains while better controlling negative transfer.

We include these ablation results in Table 4 of the revised manuscript.

Q3: How was UEC-STD adapted regarding direct versus autoregressive forecasting, correction-training data, and correction-strength selection? These details could materially affect the interpretation of the baseline comparison.

A3: We appreciate the reviewer's concern about the adaptation of UEC-STD and its implications for the baseline comparison.

Regarding direct versus autoregressive forecasting, the original UEC-STD uses autoregressive forecasting, repeatedly applying a 96-step forecaster to generate longer predictions. In contrast, ResRAF corrects direct H-step forecasts from the corresponding frozen backbone. Comparing these two settings directly would introduce differences in the base forecasts, making it difficult to isolate the effect of the correction methods. We therefore applied UEC-STD to the exact same frozen H-step base forecasts used by ResRAF, without modifying or retraining the backbone. We preserved its trend–seasonal decomposition, two-stage MLP architecture, and training objective, adapting the corrector's temporal interface from 96 to H steps. This allows Table 3 to compare the correction mechanisms on identical base forecasts rather than different forecasting strategies.

Regarding correction training, the UEC-STD corrector is trained on the validation split, using the frozen base forecasts and their corresponding ground-truth observations. For each H-step forecasting instance, the corrector takes the context (aligned to H steps) and the trend and seasonal components of the base forecast as inputs, and learns to predict the corresponding trend and seasonal errors. We deterministically split this validation correction dataset into 70% for training the corrector and 30% for early stopping.

Regarding correction-strength selection, we construct a separate selection pool using two sources: the 30% validation holdout described above and a random sample from the original training split. The training sample contains 70% as many instances as the full validation dataset. Importantly, the sampled training instances are used only for selecting the correction strength, not for training the corrector.

The coefficient β\beta controls how much of the predicted correction is added to the base forecast, i.e., Corrected=base+β⋅perr\text{Corrected}=\text{base}+\beta\cdot\text{perr}, where perr\text{perr} is the sum of the predicted trend and seasonal errors. We search the released 9-value grid β∈0.01,0.03,0.05,0.07,0.1,0.3,0.5,0.7,1.0\beta \in \\{0.01, 0.03, 0.05, 0.07, 0.1, 0.3, 0.5, 0.7, 1.0\\} and select β\beta independently for each backbone–dataset–horizon task by minimizing the validation loss on the balanced selection pool. Thus, β\beta is task-specific rather than a fixed global constant. The test split is accessed only after correction training and β\beta selection are completed; no test labels are used in either process.

We clarify in the revised manuscript that Table 3 evaluates UEC-STD as a direct-forecast adaptation under matched frozen base forecasts, rather than reproducing its original autoregressive evaluation protocol, and explicitly detail its correction-training data and validation-based selection procedure.

Reviewer_hhzsOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors proposed a retrieval approach based on the historical error to correct the time-series forecaster (ResRAF). The approach constructed a residual database which includes the residual pattern representation, then retrieveing with the support of this database. Furthermore, it introduced state-adaptive retrieval calibration to adjust. The experiments are conducted on established benchmarks and showed the high performance of the approach.

Critical strengths
  • The paper is well-presented, the framework is described clearly.
  • The approach (ResRAF) presented in this paper is sound which utilises the error of the frozen forecaster to build the knowledge base and retrieve to predict.
  • Experiments are conducted on various of established benchmarks in time-series forecasting (), compares the performance with recent baselines.
  • The paper also provides ablation study of the proposed approach. The components of ResRAF were analysed in Table 4.
Critical weaknesses
  • In the third contribution, the ResRAF is referred to address the “uncertainty in residual transfer strength”, the paper could be improved if how the ResRAF can have this effect is discussed.
Questions for authors

Could the author discuss about the mechanism of the approach that can deal with uncertainty?

2 comments

Q1: In the third contribution, the ResRAF is referred to address the “uncertainty in residual transfer strength”, the paper could be improved if how the ResRAF can have this effect is discussed. Could the author discuss about the mechanism of the approach that can deal with uncertainty?

A1: We sincerely thank the reviewer for this insightful suggestion. By “uncertainty in residual transfer strength,” we refer to the difficulty of determining how strongly each retrieved historical residual should contribute to correcting the current forecast. Even similar historical instances may differ in residual transferability under different prediction conditions.

Prior work explores uncertainty-aware adjustment of information contributions. SCN learns importance weights to regulate the contributions of samples with varying quality during training [1], while SRASR adjusts the fusion strength of retrieved information based on predictive uncertainty [2]. In time-series forecasting, prediction biases can also vary across temporal contexts [3]. These studies motivate adapting information contributions according to the conditions under which they are used. In our ResRAF framework, RAR identifies relevant historical instances through forecast-shape similarity and proximity in residual representation space. However, similarity alone does not determine the appropriate residual transfer strength. We therefore design SARC to address this uncertainty by calibrating residual contributions according to both current and historical prediction states.

Specifically, our SARC learns state-dependent transfer coefficients γi,jc=αgi,cc+δgjc\gamma^c_{i,j}=\alpha^c_{g_{i,c}}+\delta^c_{g_j}. Here, αgi,cc\alpha^c_{g_{i,c}} controls the overall correction strength based on the current prediction state, while δgjc\delta^c_{g_j} adjusts each retrieved residual's contribution according to its historical state. We learn these coefficients by minimizing corrected forecasting error with ridge regularization (Eq. 13), allowing residual transfer strength to adapt to different prediction conditions.

Our ablation results in Table 4 support this design. Compared with directly averaging retrieved residuals, SARC reduces the number of degraded tasks from 29/120 to 8/120 and the worst-case MSE degradation from 12.2% to 2.0%, while maintaining similar average improvement. These findings demonstrate the practical value of state-dependent calibration in mitigating negative transfer associated with uncertain residual transfer strength.

Following the reviewer's suggestion, we will revise Section 3.3 and Section 4.4 to clarify how SARC addresses uncertainty in residual transfer strength through state-dependent calibration, supported by the ablation findings.

References

[1] Wang, K., Peng, X., Yang, J., Lu, S., & Qiao, Y. (2020). Suppressing uncertainties for large-scale facial expression recognition. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 6896–6905.

[2] Zhou, C., & Huang, H. (2025). Syntax-Aware Retrieval Augmentation for Neural Symbolic Regression. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 13137–13147.

[3] Chen, M., Shen, L., Fu, H., Li, Z., Sun, J., & Liu, C. (2024). Calibration of Time-Series Forecasting: Detecting and Adapting Context-Driven Distribution Shift. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 341–352.

Thank the authors for your discussion during the response phase. The response addressed my comment, and I maintain the current evaluation.

Comments on the paper

1 comment

We sincerely thank the reviewers and the Area Chair for their careful evaluations and constructive feedback. The discussion has helped us better articulate the methodological contribution of ResRAF, strengthen the empirical evidence, and clarify its practical benefits and computational trade-offs. We summarize the main clarifications and additional results below.

1. Clarification of the core methodological contribution (Reviewers cA8P and xHwH).

We have further clarified the specific contribution of ResRAF in relation to existing residual-correction and task-guided retrieval methods. Rather than relearning temporal patterns already captured by a trained forecaster, ResRAF focuses on its remaining model-specific prediction errors and uses historical residual evidence to refine its forecasts.

The central design is to learn low-dimensional representations of historical residual patterns and predict these representations from observable inputs and base forecasts, enabling residual-aware retrieval even when the current prediction error is unavailable. Importantly, the predicted representations guide the selection of historical instances, while their complete observed residual trajectories provide the correction signals. This separates identifying relevant historical errors from directly reconstructing the full correction trajectory. State-Adaptive Retrieval Calibration (SARC) further adjusts the contributions of retrieved residuals according to current and historical prediction states.

2. Stronger experimental evidence and mechanism analysis (Reviewers cA8P and xHwH).

We have substantially expanded the comparisons to include direct residual decoding, ResMem, alternative retrieval strategies, SOLID, and head-only fine-tuning of Chronos-Bolt-Small. Additional ablations examine the effects of residual-aware retrieval, channel-wise scaling, and current-state and historical-state calibration.

Importantly, the additional experiments provide evidence beyond overall accuracy improvements. Removing residual-pattern similarity reduces the mean MSE improvement from 3.95% to 0.50% across 120 tasks. The retrieval-alignment analysis further shows that residual-aware retrieval identifies historical errors more closely aligned with actual query residuals. These findings support the importance of learning representations informed by historical prediction errors.

We have also extended the evaluation to Electricity and Traffic, where ResRAF improves both MSE and MAE across all five evaluated frozen backbones, providing further evidence of its applicability to higher-dimensional forecasting tasks.

3. Correction reliability and computational trade-offs (Reviewers hhzs, cA8P, and xHwH).

We have expanded the computational analysis to include fitting time, peak memory usage, and inference latency, with a direct comparison against residual decoding under matched settings. We acknowledge that ResRAF incurs additional memory and inference costs, particularly at longer prediction horizons.

Nevertheless, the broader evaluation demonstrates that its benefits extend beyond average forecasting accuracy. Across 120 tasks, ResRAF achieves a larger mean MSE reduction than direct decoding (3.95% vs. 2.48%), reduces the number of degraded tasks from 26 to 8, and lowers the worst-case MSE degradation from 22.25% to 2.00%.

The ablation results further indicate that residual-aware retrieval is the main contributor to accuracy improvements, while calibrated correction strengths help reduce negative transfer, with state-dependent calibration providing additional benefits. Together, these findings highlight the value of ResRAF in improving forecasting accuracy while reducing the frequency and severity of harmful corrections across diverse forecasting tasks.

4. Implementation details and reproducibility (Reviewers cA8P and xHwH).

We have clarified the implementation and validation-based hyperparameter selection of the compared methods, the construction of the residual knowledge base exclusively from training data, and the temporal availability of historical residuals. We have also refined the explanation of SARC and simplified the residual aggregation procedure. The base forecaster remains frozen, and all additional fitted components are fixed during evaluation.

We sincerely appreciate the reviewers' thoughtful feedback. The corresponding clarifications and additional results have been incorporated into the revised manuscript and supplementary material, with the latest direct computational comparison provided in our follow-up response. We hope these results offer a clearer assessment of ResRAF's methodological contribution, empirical effectiveness, and practical limitations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.