acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Retrieval-Augmented Error Correction for Time-Series Forecasting

Abstract

Retrieval-augmented time-series forecasting has emerged as a promising way to complement learned forecasters with historical analogues. However, directly transferring a retrieved future can overwrite level, phase, and dynamics that the backbone forecasting model already predicts correctly. This tension raises a natural question: can retrieval preserve what the forecaster already gets right while correcting what it tends to miss? To investigate this question, we introduce TRACE (Time-series Retrieval-Augmented Correction from Errors), a post-hoc framework that retrieves model-specific historical forecast errors. Rather than treating retrieved futures as replacement predictions, TRACE anchors its correction on the base forecast and retrieves errors from states characterized by both the observed context and the forecaster's prediction. It further shrinks unreliable corrections and adaptively controls how much additional trajectory information to borrow from historical analogues, without modifying the forecasting backbone. Across 585 controlled settings spanning 16 datasets and 13 forecasting backbones, TRACE strictly improves all reported metrics in 92.5% of cases, demonstrating consistent gains across diverse forecasting conditions. Complementary end-to-end evaluation further confirms that error-based correction generalizes across independently trained forecasters. More importantly, our analysis reveals that error retrieval becomes increasingly favorable as the underlying forecaster grows stronger, while the benefit of borrowing retrieved trajectories diminishes with forecast horizon. These results suggest that as forecasting models improve, retrieval is increasingly valuable for identifying how the model is likely to be wrong, rather than for supplying another complete future.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, retrieval-augmented generation

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_QrArOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TRACE corrects a frozen forecaster after the fact by retrieving its own past errors. It keeps a memory of validation residuals and finds neighbors with a descriptor built from the context and, optionally, the base forecast. It shrinks the averaged residual where neighbors disagree. A dial (beta) can also mix in part of the neighbors’ trajectories, interpolating between error retrieval and analog (future) retrieval. A validation gate then picks, per setting, among this retrieval correction and simpler ones: a seasonal blend, an overlap blend, a static or rolling bias, or no change. Across backbones and datasets, it improves on uncorrected forecasts and on RAFT, SARAF and analog retrieval applied to the same fixed forecasts.

Update: I see no rebuttal from the authors, as the deadline is less than 12 hours. Therefore, I vote reject for this paper.

Critical strengths
  • The question is well framed. The paper separates case relevance from what the forecast already captures, and nests analog and error retrieval in one dial

  • Evaluations seem comprehensive to me

Critical weaknesses
  • The main weakness is that it is hard to tell where the gain comes from. TRACE picks, per setting, among identity, a seasonal blend, an overlap blend of earlier forecasts, static and rolling bias, and retrieval. In the main comparison, non-retrieval candidates are selected in a lot of long-term cells (Section 4.2). The rolling bias also uses test labels once their targets have passed, which the static-memory baselines never see. The paper itself says Table 8 is descriptive, so the key ablation is missing: the same selector without the retrieval candidates. Without it, the results mostly show that validation-selected post-processing helps, not that retrieving errors does

  • I’m not convinced the paper’s own experiments support its main message. When everything else is held fixed, ANALOG-KNN actually has lower average error than RESIDUAL-KNN on six of the seven long-term datasets and on all four PEMS datasets. And when TRACE does choose retrieval, it nearly always keeps some of the analog drift: beta > 0 in 435 settings, out of 449 where retrieval is selected. If I read this correctly, pure error retrieval is chosen in at most 14 settings. The 15.6% headline also leans heavily on weak, lightly trained backbones. For the stronger backbones (TimeXer, TimeMixer, PAttn and PatchTST), the long-term MSE gain is much less, and TimeMixer actually gets worse on PEMS

Questions for authors
  • What do the results look like with the same selector but no retrieval candidates?

  • How often is the causal rolling bias selected in the main comparison? What changes if candidates that use test-period labels are removed?

  • Section 4.2 suggests beta > 0 in 435 of 449 retrieval selections. Where is pure error retrieval (beta = 0) actually chosen, and which configuration produced Table 1?

1 comment

Dear Reviewer QrAr,

W1 & Q1: Ablation Study

What do the results look like with the same selector but no retrieval candidates?

Removing retrieval and rerunning the same selector raised mean test MAE by 2.00% for long-term forecasting, 23.01% for traffic, and 3.43% for electricity prices on the settings with paired test results.

We removed both retrieval candidates (historical errors alone and historical errors with trajectory drift) and repeated validation selection across all 585 settings. We kept the base forecasts, remaining correction options, two chronological validation folds, scoring rule, and tie-breaking rule fixed.

TaskNumber of Settings with paired resultsFull TRACE MAEMAE without retrievalChange after removal
Long-term forecasting260/3640.387930.39569+2.00%
PEMS traffic53/15618.7949623.11969+23.01%
Electricity prices42/650.295100.30521+3.43%

Each row compares mean MAE on the same settings. For the long-term MSE metric reported in Table 1, removal raised mean error from 0.40585 to 0.42229 (+4.05%). In 134 of the 260 paired long-term settings, both MSE and MAE were unchanged.

Removing retrieval left the validation winner unchanged in 136 settings and forced a switch in 449. Test results cover all 136 unchanged settings and 219 of the 449 replacements, giving 355 paired comparisons. The largest observed increase in MAE was on traffic. The effect across the full benchmark remains open because test errors are unavailable for the other 230 replacements.

Q2: Rolling bias

How often is the causal rolling bias selected in the main comparison? What changes if candidates that use test-period labels are removed?

TL;DR: Rolling bias was selected in 9 of the 585 settings (1.54%), all long-term forecasting settings. Removing rolling bias leaves most predictions unchanged and in the remaining nine settings the selector chooses the next highest ranked option on validation.

Under the policy reported in the main comparison, rolling bias was selected in 9 of the 364 long-term settings, 0 of the 156 short-term traffic settings, and 0 of the 65 electricity-price settings. The nine selections occurred on ETTh2 (3), ETTm2 (3), Traffic (2), and Weather (1). Seven were at the longest horizon, 720 steps. The 56 rolling-bias selections in Appendix Table 8 come from a separate experiment with a different candidate pool.

With the same validation scores and tie-breaking rule, removing rolling bias can change the selected forecast only in those nine settings.

W2 & Q3: Error Retrieval, Trajectory Drift, and Configuration

When everything else is held fixed, ANALOG-KNN actually has lower average error than RESIDUAL-KNN on six of the seven long-term datasets and on all four PEMS datasets.

The two methods share the same frozen base forecasts, although their neighbor descriptors and tuning grids differ. We tested the transfer rules using the same eight neighbors and weights on German DLinear/24. With correction strength tuned on the same validation grid, test MSE was 0.4978 for historical errors, 0.4920 for aligned trajectories, and 0.4829 for errors plus half the trajectory drift. Aligned trajectories had lower MSE than historical errors (0.4920 versus 0.4978), but higher MAE (0.4461 versus 0.4445). The result supports the framework’s ability to combine the two: the intermediate rule had the lowest error. Across the main benchmark, validation selected an intermediate drift weight in 299 of 449 retrieval selections.

Where is pure error retrieval (beta = 0) actually chosen, and which configuration produced Table 1?

TL;DR: TRACE selected pure error retrieval in seven long-term forecasting settings. Table 1 reports full TRACE. In settings where validation selected retrieval, it also determined β.

The table shows the breakdown of the 14 settings by dataset, backbone, and forecast horizon, where pure error retrieval is chosen.

DatasetBackboneForecast horizonCount
TrafficPatchTST961
TrafficDLinear1921
WeatherDLinear, FEDformer, FreTS, LightTS, TimesNet7205
PEMS04, PEMS07TimeMixer6, 12, 246
German electricity pricesDLinear241

Table 1 reports p_v2_folded, which allows all four drift weights at every horizon. The same selector chooses the correction and its hyperparameters in each setting. It evaluates each candidate on two consecutive validation periods, computes its error relative to the uncorrected forecast for each reported metric, and selects the candidate with the lowest worst-case ratio across metrics and periods. The uncorrected forecast wins ties. The selected configuration is then fixed for test evaluation. For example, on the same 24-hour German electricity-price task, this rule selected β = 0 for DLinear, 0.5 for PatchTST, and 1 for TimeXer.

Reviewer_s5GaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Historical cases can improve time-series forecasts, but transferring their entire future trajectories may overwrite patterns that the forecasting model already captures correctly. It may bring an old level, phase, or rhythm into today’s forecast. If those differ from today’s actual future, accuracy decreases. The paper asks: what information from a retrieved historical case complements an existing forecast?

TRACE addresses this question through a post-hoc correction framework that retrieves the model’s errors in similar historical situations. It attenuates residual corrections when retrieved errors disagree and uses validation data to select how much historical trajectory information to retain, whether to apply a simpler correction, or whether to preserve the original forecast.

The evaluation covers 16 datasets and 13 forecasting backbones. TRACE reports reductions in both long-term MSE and short-term traffic MAE relative to uncorrected forecasts.

Critical strengths

The paper makes a clear and practically relevant contribution by distinguishing between retrieving a similar historical case and identifying information that complements the current forecast. TRACE is a flexible correction framework that improves existing forecasters without retraining them. This is especially valuable as forecasting models, including foundation models, become stronger and already capture useful structure in historical data.

The evaluation covers 16 datasets and 13 forecasting backbones. Improvements across diverse backbones, together with additional ensemble and foundation-model experiments, provide strong evidence of practical usefulness.

Critical weaknesses
  1. The authors show that TRACE improves forecasting performance. However, the research question concerns which information from historical cases complements an existing forecast, and the proposed explanation is that retrieving errors can correct missing information while preserving patterns the model already predicts accurately. Experiments that directly test this explanation would strengthen the paper. For example, the authors could use the same historical neighbors and compare corrections based on errors, trajectories, and their combination. They could then examine which corrections help or harm parts of the forecast that were initially accurate or inaccurate. This grouping should be used only for analysis, not to select corrections.

  2. The contribution of each correction option is unclear. The authors could remove each option in turn, including retrieval, seasonal blending, overlapping forecasts, and bias corrections, repeat validation selection among the remaining options, and compare test performance with full TRACE. It is particularly important for retrieval, as it relates to the paper’s main contribution, to show how much it adds beyond simple corrections.

Together, these experiments would better connect the reported gains to the proposed mechanism and method components.

Questions for authors

Please refer to weaknesses.

1 comment

Dear Reviewer s5Ga,

We sincerely thank you for your valuable feedback and for recognizing the strengths of our work! We address the concerns and questions raised in your review below:

W1: Errors and Trajectories

the authors could use the same historical neighbors and compare corrections based on errors, trajectories, and their combination ... then examine which corrections help or harm parts of the forecast that were initially accurate or inaccurate.

TL;DR: With the same historical neighbors, a combination of past errors and trajectory drift performed best in our German DLinear study. Error-only transfer disturbed fewer initially accurate predictions than trajectory transfer, but both frequently made those predictions worse.

We held the base forecast, eight retrieved windows, and neighbor weights fixed at each of 10,460 test origins. At common unit correction strength, test MSE was 0.5232 for historical errors, 0.6369 for aligned trajectories, and 0.5183 for errors plus half drift. All three exceeded uncorrected DLinear’s MSE of 0.5081. When each transfer rule could select its strength on the same validation grid, all three improved on the base forecast, with MSE of 0.4978, 0.4920, and 0.4829, respectively. Thus, both the information transferred and the strength of the correction mattered.

We set the low- and high-error thresholds from validation errors. After fixing all test predictions, we grouped individual forecast values by DLinear’s realized squared error and evaluated the same groups for every transfer rule.

Transfer rule at unit strengthLow-error values harmed (n = 37,241)High-error values helped (n = 122,121)
Historical errors80.02%60.50%
Aligned trajectories86.00%63.59%
Errors + half drift81.86%67.44%

Error-only transfer caused the least harm to initially accurate values. The hybrid helped the largest share of initially inaccurate values and was the only rule to reduce mean squared error within that group. All three worsened most low-error values at unit strength.

W2: Contribution of each correction option

The authors could remove each option in turn, including retrieval, seasonal blending, overlapping forecasts, and bias corrections, repeat validation selection among the remaining options, and compare test performance with full TRACE.

TL;DR: We removed each option and repeated validation selection across all 585 settings. Removing retrieval increased mean test error in all three task families with paired results.

We kept the original validation folds, scoring rule, and tie-breaking fixed. The table distinguishes how often removal changed the selected configuration from how often both configurations have test results.

Option removedoption changesPaired test resultsObserved effect relative to full TRACE
Both retrieval options449355/585Mean error rose 4.05% for long-term MSE, 23.01% for traffic MAE, and 2.39% for price MSE.
Seasonal blend75512/585Mean long-term MSE rose 0.000096 over 291 pairs.
Overlap blend48541/585Mean long-term MSE rose 0.000188 over 322 pairs.
Static bias1585/585584 settings tied. In the remaining setting, Traffic/FreTS/720, switching to overlap blending lowered MSE from 0.6422 to 0.6410.
Rolling bias9576/585All 576 unchanged winners tied.
Identity (keep the base forecast)3582/585All 582 unchanged winners tied.

The distinction matters particularly for seasonal and overlap blending: their small average changes are measured mostly in settings where the original winner remained available. We therefore cannot conclude that these options contribute little in the settings where TRACE selected them.

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

This paper proposes TRACE, a framework that retrieves the forecaster’s past prediction residuals instead of the actual past patches.

Critical strengths
  1. The experiments are extensive and well controlled.

  2. Figure 3 is interesting. The finding that a stronger backbone benefits more from residuals than from analogs gives a useful insight for future work on retrieval-augmented time series forecasting.

Critical weaknesses
  1. The conceptual novelty is limited. Residual correction is a classic method, Eq. (1) is a simple subtraction, and the gate is a selection on the validation set. The main contribution of the paper is closer to a combination of existing methods.

  2. The paper overclaims with numbers hidden behind the statistics. The average performance gain of 15.6% is driven by weaker backbones. For example, Informer is improved by roughly 50%, while stronger backbones such as TimeXer are improved by only about 1%.

  3. The ablation study on the components of the framework is missing.

Questions for authors
  1. What is the methodological novelty that distinguishes this paper from existing components?

  2. What would be the median (instead of average) improvement? What would be the actually expected improvement that a user employ TRACE on their system with recent, strong models?

1 comment

Dear Reviewer qhKp,

We sincerely thank you for your thoughtful feedback and suggestions for improving our work! We address the concerns and questions raised in your review below:

W1 & Q1: Contributions and Methodological Novelty

What is the methodological novelty that distinguishes this paper from existing components?

TL;DR: TRACE is a forecast-aware retrieval framework that separates what to borrow from a historical episode from how strongly to correct the current forecast.

Moving the current forecast toward an aligned historical outcome combines two quantities: the error the forecaster made on that episode and the difference between its aligned historical forecast and its current forecast. An ordinary blend scales both together. TRACE stores the forecaster’s own historical errors and gives the second quantity, trajectory drift, a separate weight β. The framework can therefore transfer a recurring error while borrowing only part of the historical trajectory. Validation selects β alongside the overall correction strength α.

W2 & Q2: Performance Gain and Backbone Strength

TL;DR: The median reduction in test MSE was 4.17% across all 364 long-term settings and 1.62% across the 112 settings using the four strong backbones.

What would be the median (instead of average) improvement?

Across the 364 settings in Table 1, the median percentage reduction in test MSE is 4.17%.

What would be the actually expected improvement that a user employ TRACE on their system with recent, strong models?

Across seven datasets and four horizons per backbone, the median MSE reductions are 1.29% for TimeXer, 1.75% for TimeMixer, 2.33% for PAttn, and 1.62% for PatchTST. Across these 112 settings, the median MSE reduction was 1.62% and the mean was 1.98%. TRACE improved both MSE and MAE in 85 of 112 settings.

W3: Ablation Study

The ablation study on the components of the framework is missing.

TL;DR: Removing retrieval increased average test error in each of the three task families with paired results. Additional ablations on frozen DLinear forecasts showed the separate effects of retrieval features, trajectory drift, and disagreement shrinkage.

We first removed both retrieval options and reran the same validation selector over the remaining corrections. Under the main evaluation protocol, removal raised mean long-term MSE by 4.05% (260 out of 364 settings), traffic MAE by 23.01% (53 out of 156), and electricity-price MSE by 2.39% (42 out of 65). Each percentage compares the same settings with and without retrieval. Test coverage depends on which replacement forecast was available, so these figures describe the observed pairs.

We then examined how the retrieval components work on 24-hour German electricity-price forecasts from DLinear. Parentheses show the change in test error relative to the reference in each group. Positive values mean higher error.

ComparisonConfigurationMSEMAE
Component removal with validation reselectionFull matched grid0.4955 (ref.)0.4450 (ref.)
Remove calendar features0.5071 (+2.35%)0.4496 (+1.04%)
Remove forecast features0.4938 (−0.33%)0.4449 (−0.02%)
Remove first-difference features0.4949 (−0.12%)0.4448 (−0.06%)
Same eight neighbors and unit correction strengthError only, no shrinkage0.5232 (ref.)0.4511 (ref.)
Error only, shrinkage λ = 40.5070 (−3.09%)0.4447 (−1.43%)
Errors + half drift, β = 0.50.5183 (−0.94%)0.4392 (−2.65%)
Full drift, β = 10.6369 (+21.73%)0.4903 (+8.68%)

Calendar features helped in this setting, while removing forecast or first-difference features slightly lowered test error. At fixed correction strength, shrinkage reduced error and half drift performed better than either endpoint. The full matched search itself selected zero shrinkage and zero drift, so removing either option individually preserved its predictions. Channel pooling had no effect in this single-channel setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.