acceptodds
Under review as a conference paper at ICLR 2027

XMatch: Enhancing Covariate-Aware Time Series Forecasting through Tree-Structured Exogenous Matching

Abstract

Future exogenous variables provide valuable information for forecasting endogenous time series. Existing covariate-aware methods primarily learn the direct influence of exogenous variables on endogenous variables. However, these effects can be complex and change with the pattern of the exogenous variables, making them difficult to capture. Beyond this perspective, we observe that a given exogenous pattern often co-occurs with only a small set of endogenous response patterns. These associations motivate a strategy that matches future and historical exogenous patterns and uses the corresponding endogenous patterns to enhance forecasting. However, in real-world forecasting scenarios with multiple exogenous variables, each exogenous variable provides a distinct dimension for matching, creating a dilemma for this strategy between precise matching and sufficient historical support. To bridge this gap, we propose **XMatch** (**EX**ogenous **MATCH**ing), a covariate-aware forecasting model that realizes the aforementioned strategy through a tree-structured matching process that adaptively adjusts the number of exogenous variables used as matching conditions. Specifically, we first introduce the *ProtoTree Creator*, which organizes historical correspondences between exogenous and endogenous patterns into a *ProtoTree*, whose deeper levels incorporate additional exogenous variables for matching. For forecasting, we then design the *ProtoTree Matcher*, which uses future exogenous variables to query the *ProtoTree* and adaptively determines how many exogenous variables to use for matching based on exogenous pattern similarity and historical support. Finally, the matched endogenous patterns are used as explicit historical evidence to enhance forecasting. Extensive experiments on 12 real-world datasets demonstrate that XMatch outperforms state-of-the-art baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Covariate-Aware Forecasting, Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_tLqeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

XMatch augments covariate-aware forecasting by retrieving historical endogenous response patterns using future exogenous inputs. Soft-DTW clustering and information-gain ordering organize training-set associations into a ProtoTree. A matcher distributes weights across branches and depths according to similarity and historical support, and fuses response prototypes with a Transformer representation. Experiments cover 12 datasets and 10 baselines; averaged across two forecasting settings, XMatch reports the best MSE on nine datasets and MAE on all twelve.

Critical strengths
  1. The paper observes that exogenous patterns often co-occur with a small set of endogenous response patterns in real data, highlighting reusable Exo-Endo correspondences that conventional covariate-aware forecasting leaves underexplored. The problem is clearly defined, and Exo-Endo Association Augmentation follows naturally from this observation.
  2. XMatch uses a ProtoTree Creator to organize historical associations and a ProtoTree Matcher for adaptive matching, yielding an interpretable design. Results on 12 datasets against 10 baselines, together with ablations and parameter analyses, provide a reasonably comprehensive experimental framework.
Critical weaknesses
  1. The method appears to combine existing retrieval, prototype memory, Soft-DTW clustering and attention routing techniques, without sufficiently establishing a substantive advance over prior retrieval/prototype methods. Adaptive matching depth also relies primarily on heuristic thresholds, with limited theoretical or formal analysis.
  2. Experiments concentrate on electricity, energy and wind-power scenarios with few exogenous variables and known future covariates, potentially favoring the proposed strategy. Analyses of tree depth, node support, matching depth, variable-ordering stability, clustering quality and computational overhead remain insufficient to explain the performance gains.
Questions for authors
  1. The method does not appear inherently tied to a particular Transformer. Could it use alternative backbones such as PatchTST, iTransformer, DLinear or Mamba to demonstrate the generality of the ProtoTree Matcher?
  2. Could the current reliance on patch shapes and Soft-DTW clustering be extended with frequency-domain or wavelet features, calendar information, conditional dependencies among variables, or uncertainty in exogenous forecasts? Such extensions may broaden applicability and mitigate the limitations of fixed variable ordering and patch length.
10 comments
AuthorsOpenReview ↗

Response to Reviewer tLqe (Part I)

We are encouraged to hear that Reviewer tLqe finds our observation of recurring exogenous–endogenous associations to be well motivated and clearly defined, and considers the ProtoTree-based design to be interpretable and supported by a reasonably comprehensive experimental framework. We hope the following discussions can address the reviewer’s remaining questions.


W1-a: Novelty beyond existing retrieval/prototype methods

Our contribution starts from the observation that recurring exogenous–endogenous associations can provide explicit historical evidence for forecasting. Exploiting these associations requires addressing two connected challenges: organizing them into a usable memory and balancing matching specificity against sufficient historical support when multiple exogenous variables are involved. We design XMatch as a unified framework which addresses the precision–support dilemma through a tree-structured association memory and adaptive matching conditioned on future exogenous patterns. Clustering, prototype construction, retrieval, and routing serve complementary roles in realizing this framework. Our main contribution therefore lies in how these components are jointly designed to support Exo–Endo Association Augmentation, with further methodological refinements within individual modules, as detailed below:

  1. How the memory is organized. Prior memories are either flat instance banks or coarse-to-fine prototypes of the same signal. In the ProtoTree, each level adds one more conditioning variable. Each node stores an empirical conditional distribution pv(k)p_v(k) together with its support nvn_v. Depth means a larger conditioning set, not a finer resolution.
  2. How many variables each query uses. Existing methods condition on a fixed set of variables. XMatch decides per query and per patch how many exogenous variables to condition on. This directly addresses the precision–support dilemma that arises with multiple exogenous variables.

To further isolate the contribution of the tree structure, we followed the suggestions of Reviewers bXsN and MJsD and added a fairer, controlled flat-retrieval baseline. It uses the same prototypes, encoders, response mapping and backbone. It only replaces hierarchical routing and stopping with top-kk scoring over all non-root nodes, with kk and the temperature tuned on validation data (Table 2, Lines 403–414; Lines 427–431).

DatasetBEDESdwpfh1Sdwpfh2Energy
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
(f) matched flat retrieval0.4370.2800.3760.3750.4150.4480.4230.4780.0930.236
XMatch (Full)0.4200.2730.3490.3690.4060.4420.4130.4510.0890.231
AuthorsOpenReview ↗

Response to Reviewer tLqe (Part II)

W1-b: Justification for adaptive depth

We thank the reviewer for the opportunity to clarify the motivation for adaptive matching depth. As discussed in the Introduction, matching more exogenous variables provides more specific historical references, but leaves fewer samples to support them; matching fewer variables retains more evidence but may overlook relevant conditions. Adaptive depth allows XMatch to adjust this balance for each query and patch, using more specific matches when the evidence permits and retaining broader historical support otherwise.

The two gates implement this design through complementary checks: the similarity gate checks whether a candidate branch matches the query's exogenous pattern, while the support gate checks whether it has sufficient historical samples. Weight proceeds to deeper nodes only when both checks pass; otherwise, it remains at a shallower node. Together, these gates select relevant and sufficiently supported historical evidence to assist forecasting. Our contribution lies in this evidence-selection perspective and its realization through adaptive tree matching; the thresholds are practical controls selected on validation data. The newly added separate gate ablations (c1 and c2; Table 2, Lines 403–414; Lines 417–423) support this design: removing either gate degrades forecasting performance across the evaluated datasets.

DatasetBEDESdwpfh1Sdwpfh2Energy
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
(c) w/o search stopping0.4300.2830.3800.3750.4190.4540.4470.4840.0990.246
(c1) w/o similarity gate0.4210.2760.3770.3740.4220.4540.4370.4830.0970.238
(c2) w/o support gate0.4230.2760.3740.3730.4570.4620.4490.4910.0960.239
XMatch (Full)0.4200.2730.3490.3690.4060.4420.4130.4510.0890.231

W2-a: Few exogenous variables may favor the method

The results do not suggest that XMatch is favored by having few exogenous variables: relative to the strongest baseline, its average MSE/MAE reductions are 10.0%/7.3% on the five datasets with at least five exogenous variables (Energy, Sdwpfh1, Sdwpfh2, Sdwpfm1 and Sdwpfm2), compared with 0.3%/2.7% on the seven datasets with two (NP, PJM, BE, FR, DE, Colbun and Rapel). This is consistent with adaptive matching being more useful when more exogenous combinations make the precision–support trade-off harder. These 12 datasets are also standard benchmarks used by DAG, KITE and GCGNet. Their widespread adoption in covariate-aware forecasting research also provides indirect support for their relevance to real-world forecasting scenarios.

AuthorsOpenReview ↗

Response to Reviewer tLqe (Part III)

We thank the reviewer for highlighting this important practical concern. In real-world settings, future covariates may be unavailable or affected by noise and disturbances, and handling these situations is essential. To address the case of unavailable future covariates, we have added an evaluation of the setting without future exogenous inputs in Appendix C.2 (Table 9, Lines 1247–1261; Table 10, Lines 1264–1294), where future covariates are replaced by mask tokens. XMatch still ranks first on 11/12 datasets in both MSE and MAE. This shows that the gains do not depend only on having future covariates. The table below reports mean MSE over two horizons for XMatch and four baselines; full MSE/MAE results are in Appendix C.2.

Dataset (MSE)XMatchDAGKITEGCGNetTimeXer
NP0.4050.4190.4050.4300.460
PJM0.1230.1260.1270.1310.142
BE0.4520.4620.4530.4590.501
FR0.4280.4350.4290.4350.465
DE0.5970.6030.6010.6140.683
Energy0.1600.1500.1460.1540.174
Sdwpfm10.6860.6890.6980.7110.744
Sdwpfm20.7810.7860.8000.8070.867
Sdwpfh10.7240.7330.7600.7410.784
Sdwpfh20.8600.8700.8800.8860.925
Colbun0.1120.1170.1360.1190.166
Rapel0.2380.2610.3020.2650.380
1st Count110200

Following Reviewer bXsN's comment, we found that the TimeXer results in this setting, which we had taken from DAG [1], contained errors. We have rerun TimeXer with the public DAG code and replaced these entries in Tables 9 and 10 with the corrected results, as shown above. XMatch's ranking and the conclusions are unchanged.

[1] DAG: A Dual Correlation Network for Time Series Forecasting with Exogenous Variables. ICML 2026.

Reviewer_MJsDOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

XMatch forecasts time series by retrieving historical response patterns that occurred with similar external-variable patterns. It builds a tree from training patches, orders external variables by information gain, and lets matching stop at a shallower level when a deeper match is too dissimilar or has too few examples. The retrieved response is added to a Transformer forecast. Results on twelve datasets show strong overall accuracy, and the paper includes retrieval, ordering, stopping, and alternative-retrieval ablations. The paper mainly provides experimental evidence rather than formal theoretical claims.

Critical strengths
  1. The paper gives evidence for the problem described in the Introduction. Figure 1 shows recurring external-variable and target-pattern associations, and Figure 4 directly examines the tradeoff between matching similarity and historical support. The tree and stopping rule address this tradeoff in a clear way. Appendix A.5 also accounts for overlapping parent and child support when calculating effective support.

  2. The experimental results are useful and cover twelve datasets. For example, average MSE on Energy falls from 0.111 for the strongest listed baseline to 0.089 for XMatch. Table 2 shows that removing retrieval, changing the variable order, or removing stopping worsens the tested results.

Critical weaknesses
  1. The reliability of the smaller gains is unclear. The main and full result tables do not report variation across random seeds or state the number of repeated runs. For example, average MSE is 0.349 versus 0.350 on DE and 0.420 versus 0.423 on BE. These small differences need more support before they can be treated as clear improvements. The tradeoff analysis in Figure 4 also includes only successful matches, so reporting how often each method fails to match would help judge the claimed balance between precision and support.

  2. The ablations cover the main branches, but they do not fully separate the proposed mechanisms. The stopping ablation removes the similarity and minimum-support checks together, so it does not show what each check contributes. The direct-retrieval comparison also removes both clustering and the tree. A comparison keeping the same prototypes but removing only the tree would better isolate the tree's contribution. There is also a gap in the method description: Equation 12 specifies a hard inference rule, while Appendix B.1 studies a soft training gate and its temperature without giving that gate's formula. Clarifying this and providing any existing separate gate ablations would strengthen the link between the method and the stated problem.

Questions for authors
  1. How many random-seed runs produced the main results, and are their mean and spread available, especially for BE and DE where the MSE gaps are small?

  2. For Figure 4, what fraction of queries obtains a valid match for each method and each number of variables? Are the methods compared on the same query set?

  3. Could you give the formula for the soft stopping rule used during training and explain how it relates to Equation 12? If separate similarity-only and support-only stopping ablations, or a flat retrieval comparison using the same prototypes, already exist, please point to them.

5 comments
AuthorsOpenReview ↗

Response to Reviewer MJsD (Part I)

We thank Reviewer MJsD for the positive assessment and constructive suggestions, particularly for taking the time to read the appendix, which is typically optional during reviewing. We sincerely appreciate your time and thoughtful feedback. We hope the following discussions can address the reviewer’s remaining questions.


W1 & Q1: Seed count and result variability

The main results were obtained with a single fixed seed, following the common protocol of covariate-forecasting benchmarks such as DAG, KITE and GCGNet, which also report single runs. To assess reliability, we rerun XMatch and the three strongest baselines using three random seeds. We additionally include Sdwpfh2 and Energy, the two datasets with XMatch's largest gains in the main results, and report the mean ± std:

DatasetXMatch MSEXMatch MAEDAG MSEDAG MAEKITE MSEKITE MAEGCGNet MSEGCGNet MAE
BE0.429±0.0050.280±0.0020.429±0.0050.286±0.0060.438±0.0020.291±0.0010.454±0.0100.305±0.005
DE0.372±0.0220.375±0.0090.391±0.0150.378±0.0070.390±0.0060.383±0.0030.414±0.0210.403±0.015
Sdwpfh20.452±0.0260.491±0.0130.579±0.0450.560±0.0250.506±0.0060.531±0.0040.521±0.0600.529±0.032
Energy0.102±0.0180.245±0.0190.157±0.0490.296±0.0400.130±0.0070.276±0.0090.149±0.0100.292±0.017

When evaluated across multiple seeds, all models show some decline in average performance relative to the original single-seed results. XMatch shows no clear improvement on BE, but a more pronounced gain on DE. Nevertheless, XMatch retains an overall advantage, including on Sdwpfh2 and Energy.


W1 & Q2: Figure 4 coverage and query sets

We agree with the reviewer. Following the flat-retrieval setup, Direct-kk now uses the same full training-patch pool as Tree (Appendix A.6, Lines 829–833). We address the matching and query-set questions in three parts.

  1. Coverage (Figure 4). To quantify match failures, we added coverage rates to the revised Figure 4 (Lines 450–453). Direct-kk covers a query when at least one training patch shares its prototype labels on the top-kk exogenous variables. Tree covers a query when its unnormalized non-root weight Mq=∑v≠vrootwq,v≥0.1M_q=\sum_{v\neq v_{\text{root}}}w_{q,v}\ge0.1. The rates over all test queries, rounded to whole percentages, are:
Coverage over all test queriesDirect-1Direct-2Direct-3Direct-4Direct-5Direct-6Tree (Mq≥0.1M_q\ge0.1)
Energy100%58%33%19%8%N/A (5 exog.)62%
Sdwpfh1100%100%99%94%84%72%100%

Tree's exact coverage is 61.6% on Energy and 99.6% on Sdwpfh1; the latter rounds to 100%.

  1. Non-root weight distribution. We also report Tree's unnormalized non-root mass over all test queries:
Non-root mass MqM_qEnergySdwpfh1
[0,0.2)[0,0.2)47.4%4.4%
[0.2,0.4)[0.2,0.4)17.1%7.6%
[0.4,0.6)[0.4,0.6)19.8%12.1%
[0.6,0.8)[0.6,0.8)12.6%21.4%
[0.8,1.0][0.8,1.0]3.1%54.5%
Mean MqM_q (all / covered queries)0.28 / 0.450.73 / 0.73

Energy's low direct-matching success rates coincide with a lower mean non-root mass for Tree, whereas Sdwpfh1 has higher matching success and a higher mean mass. Tree still improves forecasting on both datasets, suggesting that it adapts its use of retrieved evidence to different matching conditions.

  1. Same-query experiment. To compare methods on the same queries and include weak and failed retrieval omitted by Figure 4's conditional boxplots, we added an analysis on the full test query set(Appendix B.4 and Figure 6, Lines 1019–1087). Direct-kk uses uniform weights over matched training patches, or the training marginal when none match; Tree keeps its actual inference weights, including root mass, without renormalizing non-root weights. Failed retrieval contributes zero similarity lift and zero weighted historical support. On both datasets, Tree achieves higher mean similarity lift than every tested Direct-kk, while retaining more support than Direct-22–Direct-55 on Energy and Direct-55–Direct-66 on Sdwpfh1. The comparison, conducted on the same query set, therefore supports the claimed balance after weak and failed matches are included.
AuthorsOpenReview ↗

Response to Reviewer MJsD (Part II)

W2 & Q3 (a): Separate gate ablations and flat retrieval

Separate gates and flat retrieval. We split the stopping ablation into (c1) w/o similarity gate (support check only) and (c2) w/o support gate (similarity check only). Also incorporating Reviewer bXsN's suggestion, we add (f) matched flat retrieval, which keeps the same prototypes, encoders, response mapping and backbone, and only replaces hierarchical routing and stopping with top-kk scoring over all non-root nodes (kk and temperature selected on validation data). These variants have been added to Table 2 (Lines 403–414) and are described in Section 4.2 (Lines 417–423, 427–431).

DatasetBEDESdwpfh1Sdwpfh2Energy
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
(c) w/o search stopping0.4300.2830.3800.3750.4190.4540.4470.4840.0990.246
(c1) w/o similarity gate0.4210.2760.3770.3740.4220.4540.4370.4830.0970.238
(c2) w/o support gate0.4230.2760.3740.3730.4570.4620.4490.4910.0960.239
(e) mix exo retrieval0.4250.2850.3630.3770.4700.4870.4220.4760.0940.237
(f) matched flat retrieval0.4370.2800.3760.3750.4150.4480.4230.4780.0930.236
XMatch (Full)0.4200.2730.3490.3690.4060.4420.4130.4510.0890.231

Removing either gate degrades forecasting, while XMatch outperforms the matched flat-retrieval variant using the same prototypes. These results support the effectiveness of both stopping checks and adaptive tree matching.


W2 & Q3 (b): Soft gating during training and hard gating at inference

We thank the reviewer for pointing out this gap; we have added the formula to Appendix A.5 (Lines 800–816), with a pointer in Section 3.3.2 (Lines 317–319). To train support-qualified tree nodes more fully, we soften only the similarity check: an existing branch with at least nmin⁡n_{\min} historical patches has gate value σ((as,l,j−τsim)/τgate)\sigma((a_{s,l,j}-\tau_{\mathrm{sim}})/\tau_{\mathrm{gate}}); otherwise, its gate is zero.

At inference, the sigmoid is replaced by I[as,l,j≥τsim]\mathbb{I}[a_{s,l,j}\ge\tau_{\mathrm{sim}}], which recovers Equation (12). The soft gate lets gradients reach the similarity scores and routing weights of every support-qualified branch; the weight not passed to a child stays at the parent in both cases. τgate\tau_{\mathrm{gate}} only controls the smoothness of this gate, unlike τroute\tau_{\mathrm{route}}, which controls the softmax over prototypes.


We once again thank the reviewer for the positive assessment and constructive feedback. We hope these additional results help clarify the points raised by the reviewer. We would be glad to discuss any further questions!

AuthorsOpenReview ↗

Kindly Request for Reviewer's Feedback

Dear Reviewer MJsD,

Since the End of the Rebuttal comes soon, we would like to inquire if our response addresses your primary concerns. If they do, we would greatly appreciate it if you could take them into account in your final assessment. We would also be glad to discuss any further questions or suggestions. Thank you again for your positive and constructive review!

Kind Regards, Authors

Reviewer_bXsNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes XMatch for time series forecasting with known future exogenous variables. It clusters aligned exogenous and endogenous patches and organizes their associations into a ProtoTree. Successive tree levels condition on additional exogenous variables, while similarity- and support-based stopping balances matching specificity against historical support. Retrieved endogenous response representations are fused with a Transformer backbone. Experiments on 12 datasets compare XMatch with 10 baselines and include component ablations, hyperparameter sensitivity studies, and analyses of retrieval quality and support.

Critical strengths
  • The paper addresses a meaningful challenge in covariate-based retrieval: conditioning on more exogenous variables produces more specific matches but reduces historical support. The hierarchical memory and adaptive stopping mechanism provide a coherent approach to this tradeoff.

  • The reported forecasting results are promising. XMatch achieves the best horizon-averaged MAE on all 12 datasets and the best MSE on nine, with substantial improvements on Energy and Sdwpfh2. Comparisons with models that natively incorporate future covariates, together with component ablations, provide useful evidence for the effectiveness of the overall approach.

Critical weaknesses
  • The experiments do not sufficiently isolate the contribution of adaptive tree matching. The five-neighbor retrieval baseline in Table 2 differs from XMatch in representation, memory construction, and aggregation. The ordering and stopping ablations provide useful evidence, but do not establish an advantage over comparably learned and tuned flat retrieval. Moreover, these ablations omit Energy and Sdwpfh2, where the largest gains are highlighted. This limits attribution of the reported improvements to the central methodological contribution.

  • The mechanism analysis does not establish the claimed balance across all queries. Appendix A.5 conditions on successful matches and renormalizes non-root weights, whereas prediction uses their unnormalized mass. Consequently, favorable similarity and support metrics can coexist with negligible retrieval contributions. Without coverage and actual non-root mass, Figure 4 provides incomplete evidence for the proposed mechanism.

Questions for authors
  • Does XMatch retain its advantage over flat retrieval using the same learned representations and historical prototypes, with neighborhood size or bandwidth selected on validation data? Can this comparison and the component ablations be provided for Energy and Sdwpfh2?

  • What fraction of queries contributes to Figure 4, and what is the distribution of total non-root retrieval mass? Do the conclusions hold on a common query set when weak or absent retrieval contributions are included?

13 comments
AuthorsOpenReview ↗

Response to Reviewer bXsN (Part I)

We thank Reviewer bXsN for carefully reviewing our work, particularly for taking the time to read the appendix, which is typically optional during reviewing. We sincerely appreciate the time and thoughtful feedback, which have helped us strengthen the paper. We hope the following discussions can address the reviewer’s remaining questions.


W1 & Q1: Matched flat-retrieval baseline and ablations on Energy/Sdwpfh2

We agree that (e) changes representation, memory and aggregation at once. To address the lack of a controlled comparison with flat retrieval, we add (f) matched flat retrieval: it keeps the same prototypes, encoders, response mapping and backbone, but scores all non-root nodes directly (average path-wise query–key cosine similarity) and aggregates the top-kk with softmax weights. k∈{1,3,5,10,20,all}k\in\lbrace 1,3,5,10,20,\text{all}\rbrace and the temperature (bandwidth) are selected on validation data. We also split the stopping ablation into (c1)/(c2), and extend all ablations to Energy and Sdwpfh2. The split follows Reviewer MJsD's suggestion. The full results are also provided in Table 2 and Section 4.2 of the revised manuscript (Lines 403–431).

DatasetBEDESdwpfh1Sdwpfh2Energy
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
(a) w/o tree retrieval0.4400.2980.3710.3730.4420.4770.4790.5050.1090.255
(b) w/o variable ordering0.4380.2890.4100.3960.4570.4810.4730.5010.1050.248
(c) w/o search stopping0.4300.2830.3800.3750.4190.4540.4470.4840.0990.246
(c1) w/o similarity gate0.4210.2760.3770.3740.4220.4540.4370.4830.0970.238
(c2) w/o support gate0.4230.2760.3740.3730.4570.4620.4490.4910.0960.239
(d) endo history retrieval0.4330.2990.3560.3730.4780.4910.4200.4780.1030.244
(e) mix exo retrieval0.4250.2850.3630.3770.4700.4870.4220.4760.0940.237
(f) matched flat retrieval0.4370.2800.3760.3750.4150.4480.4230.4780.0930.236
XMatch (Full)0.4200.2730.3490.3690.4060.4420.4130.4510.0890.231
  • Tree vs. flat retrieval. XMatch outperforms (f) on all five datasets in both metrics (MSE −2.2% to −7.2%), so the advantage comes from hierarchical routing and adaptive stopping rather than from the shared prototypes alone.

  • Energy and Sdwpfh2. The gains on these two datasets come mainly from the core method. Removing retrieval (a) causes the largest degradation in the table. Variable ordering and stopping also play a clear role on these two datasets. Overall, the full model is best in every column, confirming that each component contributes.

AuthorsOpenReview ↗

Response to Reviewer bXsN (Part II)

W2 & Q2: Coverage and non-root retrieval mass in Figure 4

We agree with the reviewer. To address these concerns, we report matching coverage and non-root mass distributions, and add a full-query comparison on Energy and Sdwpfh1. Following the flat-retrieval setup, we also revised Direct-kk to use the same full training-patch pool as Tree for a fairer comparison (Appendix A.6, Lines 829–833).

  1. Coverage (Figure 4). We updated Figure 4 in the revised manuscript (Lines 450–453) to show matching coverage. Direct-kk succeeds when at least one historical patch shares the query's prototype labels on its top-kk exogenous variables. Tree succeeds when its total unnormalized non-root mass Mq=∑v≠vrootwq,vM_q=\sum_{v\neq v_{\text{root}}} w_{q,v} is at least 0.1. Coverage is the percentage of all test queries satisfying the corresponding success criterion. The percentages below Figure 4 report these success rates, rounded to whole percentages:
CoverageDirect-1Direct-2Direct-3Direct-4Direct-5Direct-6Tree (Mq≥0.1M_q\ge0.1)
Energy100%58%33%19%8%N/A (5 exog.)62%
Sdwpfh1100%100%99%94%84%72%100%

Direct matching covers fewer queries as kk increases. On Energy, Direct-5 covers about 8% of queries, whereas Tree meets the Mq≥0.1M_q\ge0.1 criterion for 61.6%. On Sdwpfh1, Direct-6 covers about 72%, compared with 99.6% for Tree.

  1. Distribution of non-root mass. The following distribution is computed over all test queries:
Non-root mass MqM_qEnergySdwpfh1
[0,0.2)[0,0.2)47.4%4.4%
[0.2,0.4)[0.2,0.4)17.1%7.6%
[0.4,0.6)[0.4,0.6)19.8%12.1%
[0.6,0.8)[0.6,0.8)12.6%21.4%
[0.8,1.0][0.8,1.0]3.1%54.5%
Coverage (Mq≥0.1M_q\ge0.1)61.6%99.6%
Mean MqM_q (all / covered queries)0.28 / 0.450.73 / 0.73

Energy’s low direct-matching success rates coincide with a lower mean non-root mass for Tree, whereas Sdwpfh1 has higher matching success and a higher mean mass. Tree still improves forecasting on both datasets, suggesting that it adapts its use of retrieved evidence to different matching conditions.

  1. Same-query analysis with failed matches. To address the concern that Figure 4 includes only successful matches and may compare different query sets, we add an analysis on the identical full test query set for both methods(Appendix B.4 and Figure 6, Lines 1019–1087). Direct-kk uses its actual uniform weights over matched training patches, or the unconditional training distribution when none match; Tree uses its original inference weights, including root mass, without renormalizing non-root weights. We compare similarity lift over the common root baseline and weighted historical support, with failed retrieval contributing zero lift and zero support. On both datasets, Tree achieves higher mean similarity lift than every tested Direct-kk; it also retains more support than Direct-22–Direct-55 on Energy and Direct-55–Direct-66 on Sdwpfh1. Thus, the advantage persists after weak and failed queries are included.

We hope the controlled ablations and full-query analyses address your concerns about the tree's contribution, match coverage, and weak or failed retrieval. We would be glad to discuss any further questions!

AuthorsOpenReview ↗

Kindly Request for Reviewer's Feedback

Dear Reviewer bXsN,

Since the End of the Rebuttal comes soon, we would like to inquire if our response addresses your primary concerns. If it does, we kindly request that you reconsider the score. If you have any further suggestions, we would be happy to discuss them and improve the paper. Thank you again for your time and valuable feedback!

Kind Regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.