acceptodds
Withdrawn from ICLR 2027

ExoRAF: Exogenous Residual Retrieval for Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Effectively leveraging exogenous variables remains a longstanding challenge in real-world time series forecasting. The key insight of this paper is that exogenous variables are more effectively utilized by modeling forecasting residuals rather than directly predicting the target series. While an endogenous-only backbone captures the intrinsic temporal dynamics of the target series, exogenous variables mainly explain the scenario-dependent residual variations that cannot be captured by historical observations alone. Moreover, similar exogenous contexts often induce similar residual patterns. Motivated by the insight, we propose ExoRAF, an exogenous residual retrieval method for time series forecasting. ExoRAF builds a residual memory from historical exogenous contexts, retrieves relevant residual information using the current exogenous context, and leverages the retrieved residual information to generate correction terms that refine the backbone forecasts. The resulting decoupling of endogenous temporal modeling and exogenous residual explanation makes ExoRAF a plug-and-play module for existing forecasting models. Extensive experiments on diverse real-world datasets demonstrate that ExoRAF consistently improves strong forecasting backbones across a wide range of application scenarios.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Machine learning, Time series analysis, Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Qkf5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

ExoRAF is a plug-in head for an endogenous-only forecaster. Discrete exogenous variables (holidays, events) and continuous ones (weather) are encoded into two queries, each refined by attention over a learnable key-value memory. The two are fused and used to attend over a second memory of M residual prototypes, and the weighted prototype, scaled by a step-wise and a channel-wise gate, is added to the backbone forecast. The main benchmark is a new Ride-Hailing dataset (200 areas, 30-minute steps, four years) from an anonymous RideBench repository, plus four EventTSF datasets. Adding ExoRAF to endogenous-only backbones lowers overall WMAPE on Ride-Hailing by 4.5% on average (Table 2).

Critical strengths
  1. The goal is practical: add exogenous information to an existing forecaster without changing its architecture.
  2. The module is tested on many backbones (channel-independent, multivariate, exogenous-aware and three foundation models) and in several settings: regular and long-horizon forecasting, unseen areas, cross-backbone transfer, and missing future covariates.
  3. Code is released, and the paper reports where the module helps little, e.g. large-scale events and exogenous-aware backbones (Table 8).
Critical weaknesses
  1. Nothing is retrieved, and the key baseline is missing. Section 3 never describes a step that writes historical residuals into the memory. The query memories are "randomly initialized and learnable", the residual memory holds learned keys and values with between 10 and 50 (Fig. 8a), and all of it is trained end to end with the forecasting loss. This is a softmax mixture over a small set of learned prototypes, i.e. a small parametric output head, which A.2 nevertheless calls "our non-parametric design". The main tables then compare an endogenous-only backbone with the same backbone plus a head that sees every exogenous variable, including time of day, day of week and area id (A.2). Any head with those inputs should help. The paper needs the same backbone given the same exogenous variables in a plain way, as extra input channels or as a linear or MLP correction computed from the exogenous features. The "RRM ⇌ MLP" row in Table 3 keeps both query encoders and both query memories, so it does not answer this.

  2. The paper's own tests show the model barely uses the exogenous signals it is built on. Adding 20% noise to weather changes WMAPE in the weather scenario by at most 0.0001 (Table 4). Without any future covariates, PatchTST still gets most of the holiday gain (0.1507 to 0.1368, versus 0.1346 with them; Table 9), and in long-horizon forecasting it does slightly better without them (0.1282 vs 0.1287; Table 10). This looks like a learned offset by time of day and area, not a model of weather or holidays.

  3. It does not beat existing exogenous models, and it sometimes makes them worse. TiDE beats three of the five ExoRAF models overall and all five on holidays (Table 2). XLinear (Houston) and TimeXer (New York temperature) beat every ExoRAF row (Table 1). Adding ExoRAF to TiDE hurts it (0.1043 to 0.1050 overall, 0.1282 to 0.1330 on holidays; Table 8), and the transferred module hurts four of eight backbones under events (Fig. 5). "Consistently improves" in the abstract is false.

  4. The evaluation cannot be checked or put in context. The main results come from one new dataset in an anonymous repository whose source, city, provider and license are not stated, and the paper does not define how test windows are assigned to the weather, holiday and event scenarios. The EPF electricity-price benchmarks used by TimeXer, the standard testbed for exogenous forecasting, are absent. All backbones run with dropout 0.5 and hidden size 256, with only the learning rate searched (A.2). The tables give no standard deviations (only Fig. 8b shows three seeds), while the overall gains on Ride-Hailing are 0.004 to 0.006 WMAPE. Appendix A.1 is the standard decomposition of squared error around a conditional mean and says nothing specific to this method; it assumes squared loss, while the model is trained with MAE (Fig. 2).

Questions for authors

See weaknesses.

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

This paper proposes a forecasting architecture with an exogenous-conditioned residual correction. Separate encoders process discrete and continuous covariates, and memory-based attention produces residual prototypes that are combined through temporal and channel gates. The corrected forecast adds this output to the backbone prediction. The evaluation includes a large ride-hailing dataset, four smaller datasets, multiple forecasting architectures, and transfer across geographic areas and backbones.

Critical strengths
  1. The cross-backbone transfer experiments test whether a trained correction module can help an independently trained model without further adaptation. This is stronger evidence for the plug-and-play claim than simply retraining the full system with several backbone architectures.

  2. The empirical analysis goes beyond aggregate accuracy. Scenario-specific results, residual-similarity diagnostics, and a parameter-matched MLP replacement help examine why the module works. The larger gains for recurring weather and holiday effects, together with weaker transfer for localized events, are consistent with the proposed reliance on recurring residual structure.

Critical weaknesses
  1. The relationship between the stated residual memory and the implemented memory is insufficiently clear. The paper repeatedly describes storing and retrieving historical forecasting residuals, but Section 3.1 explicitly introduces randomly initialized, learnable key–value banks, and Section 3.2 does not specify how its residual prototypes are constructed from actual historical errors. If these are freely learned parameters, the method is closer to a conditional prototype model than retrieval from an empirical residual database. This distinction matters for both the novelty claim and reproducibility.

  2. The evidence supports a useful correction module more clearly than the broader claim that exogenous information is better modeled through residuals. Appendix A.1 gives a risk decomposition for a fixed backbone, but does not establish superiority over direct exogenous conditioning. The empirical picture is also mixed for stronger exogenous-aware models: in Table 8, adding ExoRAF worsens TiDE’s overall WMAPE from 0.1043 to 0.1050 and its holiday WMAPE from 0.1282 to 0.1330. The MLP ablation is helpful, but only replaces the residual retrieval component; it does not fully resolve this broader comparison.

Questions for authors
  1. Are the residual-memory values measured historical residuals, learned prototypes, or a combination? Please specify their initialization, update rule, and relationship to individual training examples.
  2. In the main experiments, is the backbone pretrained and frozen before fitting ExoRAF, or are both optimized jointly? If training is joint, what preserves the interpretation of the backbone as an endogenous-only predictor?
  3. With the same backbone, covariates, and training budget, does residual correction outperform a direct exogenous-conditioning alternative? Please distinguish this comparison from replacing only the final retrieval module with an MLP.
1 comment

Thanks for your advice.

Reviewer_6grBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ExoRAF, a framework that uses exogenous variables to explain the residual of an endogenous-only backbone model, instead of incorporating them into the backbone architecture. Based on the observation that similar backbone residual patterns recur under similar events (exogenous variables), ExoRAF encodes the exogenous variables as a query and retrieves information from a residual memory.

Critical strengths
  1. The idea is intuitive and makes sense. Exogenous variables are indeed a common cause of backbone failures, which appear as residuals.

  2. Since the backbone is left untouched and only the module is trained, the framework can be used in a plug-and-play manner.

Critical weaknesses
  1. The claimed “new perspective” is not new. Using a separate model to explain the residual of a history-based base model is a classic two-stage structure (e.g., Zhang et al. [1]), and letting exogenous variables explain the residual component in a deep learning model has already been done in NBEATSx [2].

  2. The terms “retrieval” and “memory” are not used in their standard sense. In this paper, the memory consists of randomly initialized, learnable variables rather than the recorded history of the time series. It is therefore closer to a set of trained prototypes than to a direct source of external information.

  3. The experiments need more recent baselines. For example, TimeMixer [3] and TimeMixer++ [4] are missing, even though TimeMixer is cited in the introduction. The paper uses datasets from EventTSF [5], but EventTSF itself is not compared. In addition, the baseline TSFMs are evaluated without exogenous variables.

  4. There is no comparison with existing retrieval-augmented time series forecasting methods (e.g., RAFT [6]). Beyond the comparison, the paper should also examine what happens when ExoRAF is applied on top of these endogenous retrieval-augmented baselines.

[1] Time series forecasting using a hybrid ARIMA and neural network model (Neurocomputing, 2003)

[2] Neural basis expansion analysis with exogenous variables: Forecasting electricity prices with NBEATSx (International Journal of Forecasting, 2023)

[3] TimeMixer: Decomposable multiscale mixing for time series forecasting (ICLR 2024)

[4] TimeMixer++: A general time series pattern machine for universal predictive analysis (ICLR 2025)

[5] EventTSF: Event-aware non-stationary time series forecasting (arXiv, 2025)

[6] Retrieval augmented time series forecasting (ICML 2025)

Questions for authors
  1. Please clarify the novel contribution that distinguishes this work from existing works.

  2. What will happen if we use ExoRAF with existing retrieval-augmented time series forecasting methods like RAFT?

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

Related papers

Discussion (0)

Sign in to comment.