acceptodds
Under review as a conference paper at ICLR 2027

DAShrink: Disagreement-Adaptive Shrinkage for Multivariate Forecast Combination

Abstract

Combining multivariate forecasters requires deciding how strongly to depart from equal weights. A learned router can exploit differences between experts, but a large weight change can also produce a large forecast change when their predictions disagree. We propose disagreement-adaptive shrinkage (DAShrink), which pulls contextual combination weights toward the uniform average. The shrinkage strength depends on the spread of history-normalized expert predictions. A scalar chosen on validation windows controls the overall departure from equal weighting and can recover the uniform average exactly. The resulting weights have a positive lower bound, and the forecast deviation from the uniform average is bounded by the range of expert predictions. We evaluate a model with three heterogeneous experts on ten multivariate benchmarks with forty dataset-horizon settings. Against eleven baselines, DAShrink achieves an overall average rank of 1.262 and ranks first in 34 settings for MSE and 34 for MAE. With trained parameters fixed, removing either combination control increases both errors in all eight ETT ablation settings.

Primary area:
learning on time series and dynamical systems
Keywords:
multivariate time series forecasting, forecast combination, ensemble learning, adaptive shrinkage, expert routing

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_9g83OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DAShrink, a method for combining horizon forecasts from multiple time-series forecasting models by balancing a uniform combination with a learned router. Given k pre-trained frozen models, DAShrink first combines their predictions using two Mixture-of-Experts (MoE) routers: (1) a uniform router, assigning 1/k to each forecaster prediction (the mean operation), and (2) a learned MoE router, which is trained to predict assignment weights from normalized encodings from each expert and optimized with an MSE loss and a regularization term. The weights obtained by these two routers are balanced via linear interpolation based on two parameters: γ\gamma, which is obtained by fitting on a validation set, and β\beta, which is set per time point and per variable based on the disagreement between the two routers. The method is implemented using three experts: DLinear, TimeFilter, and an expert combining decomposition, lag-based extrapolation, and variable-token attention (denoted as a ‘contextual expert’). It is evaluated on the ETT datasets (excluding ETTh1), PEMS datasets, Electricity, Traffic, and Solar-Energy, and compared to time-series baselines as well as to MoE-like methods (e.g., MoLE-DLinear and TimeFuse).

Critical strengths

The method outperforms several time-series baselines and MoE-like methods on standard time-series benchmarks by notable margins in the majority of cases.

Critical weaknesses
  1. The method is framed as general, yet it is evaluated using only three specific experts, without showing how the identity of the specific experts affects performance gains or whether the proposed balancing works across various configurations. Crucially, it is not clear whether the performance gain comes from the combination method or from the specific experts chosen. In addition, the paper is missing an apples-to-apples comparison with MoE-like baselines and the baseline of the contextual expert and the uniform router.

  2. The conceptual motivation and empirical justification for key design choices is missing or only partially presented. In addition, for some parts, the paper appears rushed (missing notation definitions, missing references).

Questions for authors

Questions

  1. Have the authors evaluated their method using different configurations (varying the number of experts and expert identities)?
  2. Could the authors please provide an apples-to-apples comparison using the same set of experts while varying only the combination method? For example, running TimeFuse with the same expert set used by DAShrink, or running DAShrink under the same configuration as MoLE-DLinear (i.e., using only DLinear experts)?
  3. Could the authors please include the results of the uniform router and of the contextual expert for the main experiments in the paper (Tables 1-2)?
  4. Could the authors please provide a clear motivation for their main design choices and empirically motivate them?
3 comments

1. Expert configurations and attribution. We independently fit every trainable combiner on eight banks of 2-5 experts on ETTh2/ETTm1 H96 with three seeds each. The banks comprise the contextual/DLinear/TimeFilter (N/D/T) bank, all three two-expert subsets, three independent DLinear models, two DLinear models plus TimeFilter, five experts, and an exact duplicate. Within each task, seed and bank, every rule receives identical frozen forecasts. This fixes expert quality and expert cost across combination rules (Tables 7-10, D.1-D.2).

DAShrink's mean paired MSE reductions from each bank's uniform average are 0.444% for three DLinear models and 2.348% for five experts (six task-seed units each; Table 8). These configurations establish that the rule operates beyond the particular contextual expert. The contextual expert contributes differently by task: adding it to DLinear/TimeFilter changes uniform MSE from .9416 to .9711 on ETTh2 and .9687 to .9285 on ETTm1.

2. Matched TimeFuse and MoLE. The matched TimeFuse uses its upstream linear-softmax fusion head and 22 history features on exactly the shared N/D/T reports. On ETTh2/ETTm1, DAShrink MSE is .9703/.8946 and matched TimeFuse is .9787/.8942. On three shared, frozen DLinear forecasts, DAShrink/MoLE-gate MSE is 1.1036/1.1092 on ETTh2 and .8925/.8927 on ETTm1 (Tables 7, 10). MoLE's adapter uses its four calendar features and zero head dropout; DAShrink uses report/history features. This tests combination heads with expert identity fixed. Original MoLE trains heads/gate jointly; D.2 specifies the frozen-bank adapter.

Each neural combiner is fitted under its own objective and a shared 30-epoch budget with router-stop checkpoint selection. Plain routing minimizes MSE; the two shrinkage heads minimize MSE plus .05 KL. Static stacking solves a simplex-constrained MSE quadratic program. Expert fit, expert stop, router fit, router stop, scale calibration and testing use six disjoint chronological blocks, with standardization estimated from expert-fit data (Tables 6-7, D.1). These separately fitted alternatives complement the original fixed-parameter deletions.

3. Standalone experts and uniform averaging. Table 7 adds the requested controls on ETTh2/ETTm1 H96, with three-seed MSE/MAE means and sample SD. Mean MSE/MAE:

RuleETTh2ETTm1
Contextual expert1.1450/.70931.0158/.5682
DLinear1.1201/.7357.8942/.5292
TimeFilter.9338/.62781.2655/.6298
Uniform.9711/.6520.9285/.5335
DAShrink.9703/.6519.8946/.5229

The rows distinguish standalone expert quality, equal pooling and the incremental combination rule on identical test windows per seed. All use D.1's disjoint splits and fit-only target scaling. Tables 1-2 retain the original forty-task system benchmark; the added control coverage is these two H96 tasks. Table 7 supplies same-bank attribution; Section 3 and Tables 15-19 establish the spread-dependent movement constraint and its decision consequence.

4. Design motivation, evidence and notation. Uniform weights are fixed at 1/K1/K. Spread is the sample SD of clipped, history-normalized expert reports at one lead/channel. Original training alternates expert/router updates; the router uses detached reports. Experts stay fixed after checkpoint selection (Section 3, Figure 1).

The post-softmax control enforces a budget for every proposal. Local coefficients shape the correction; the task scale sets its amplitude, with zero recovering exact averaging. Equation 6 derives the quadratic validation-loss path (C.4). The .25 cap gives wbhc(k)w_{bhc}^{(k)} in [1/4,1/2][1/4,1/2] for three experts. With report and spread clips inactive, proposal sensitivity is bounded proportionally to se−ss e^{-s}, decreasing above unit spread.

Fixed-parameter removals increase both errors in all eight original ETT settings. Independent refits evaluate accuracy and departure jointly: DAShrink improves on uniform in four core runs and ties in two; plain routing has lower MSE. Seed-mean p99 absolute departures from uniform on unclipped coordinates are .017/.102 history SD for DAShrink, .081/.538 for constant shrinkage and .306/1.397 for plain routing on ETTh2/ETTm1 (Section 4.3). Tables 15-18 compare inverse spread, direct clipping, parameter settings, no-KL and reversed response.

Proposition 2 identifies the value of observing spread: at matched sensitivity, zero-preserving scalar rules given only aggregate correction and fixed history/mean context retain at most one eighth of DAShrink's unit-spread correction, under its parameter, unclipped-report and margin conditions. All 44 references were verified against primary sources. We invite reassessment of novelty and soundness using the enforced movement budget, its information analysis and the matched tests.

I thank the authors for their time and response, which addressed some of my concerns. However, critical issues remain: the main evaluations are limited to two datasets (Etth2/Ettm1), with ETTh1 still omitted without clarification. Reporting on all four ETT datasets (alongside additional benchmarks) is the minimum standard in recent literature. Furthermore, the matched comparisons with TimeFuse and MoLE show only marginal gains. Taken together, the limited benchmark coverage and modest improvements under matched protocols indicate that more work is needed, and I thus maintain my score.

  1. Benchmark coverage and ETTh1. The main evaluation tests the complete DAShrink system on 10 datasets and 40 dataset–horizon settings, with four horizons per dataset (Tables 1–2). The suite includes ETTh2, ETTm1, ETTm2, Electricity, Traffic, Solar-Energy, PEMS03, PEMS04, PEMS07, and PEMS08. ETTh2 and ETTm1 are the two datasets used in the additional matched comparison at horizon 96, where the combination heads are independently fitted on identical frozen expert forecasts with three seeds per task (Tables 7 and 10).

ETTh1 was outside the original benchmark grid. In the revision, we retained that grid and used the new experiments to address the attribution question on the two series from the original component study. Holding the expert forecasts fixed lets us measure how changing the combination rule affects the issued forecast on those series. That scope accounts for why ETTh1 was not added in the revision. We agree that extending the matched protocol to all four ETT datasets would strengthen the empirical evidence for the combination rule.

  1. Matched results and the contribution of DAShrink. The mean MSE differences against matched TimeFuse and MoLE are small. On ETTh2 and ETTm1, DAShrink and matched TimeFuse achieve 0.9703 versus 0.9787 and 0.8946 versus 0.8942, respectively. On the shared DLinear bank, DAShrink and the matched MoLE gate achieve 1.1036 versus 1.1092 and 0.8925 versus 0.8927.

DAShrink uses expert disagreement at each forecast step and channel to set an explicit bound on how far the learned combination can move from the uniform forecast. With the experts held fixed, the 99th-percentile absolute departures on ETTh2 and ETTm1 are 0.017 and 0.102 history standard deviations for DAShrink, 0.081 and 0.538 for constant shrinkage, and 0.306 and 1.397 for plain routing. These values are averaged across seeds on unclipped coordinates (Section 4.3). Proposition 2 explains why observing disagreement matters. Under its conditions, DAShrink retains a larger correction at unit spread than a scalar rule given the aggregate correction and fixed context at the same worst-case sensitivity. In practice, the sensitivity bound lets a forecaster determine at issue time whether a specified perturbation to the proposed weights could flip an alarm decision (Figure 8).

Reviewer_c6dqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies multivariate forecast combination: how far should learned and input-dependent weights depart from equal averaging. It proposes DAShrink, which shrinks a router's weight proposal toward the uniform average at each forecast coordinate. The shrinkage grows stronger as the history-normalized expert predictions disagree more. A task-level scale chosen on paired validation losses with a mean–dispersion eligibility criterion sets how much the combination adapts overall. The construction also gives two deterministic guarantees: every expert keeps a positive minimum weight, and the forecast's deviation from the uniform average is bounded by the range of the expert predictions. The system combines DLinear, TimeFilter and a new contextual expert. The method outperformed other baselines across diverse benchmarks on averaged ranks.

Critical strengths
  1. The paper is studying a practically important problem, no single deep forecaster wins everywhere, so combining them is a practical hedge. Picking one model per dataset is fragile, it is more flexible to study the forecasting combination problem.

  2. The method's shrinkage strength can vary with the input, horizon step and channel, which earlier equal-weight shrinkage methods do not do.

  3. The experiment comparisons are broad and has compared with competing combination methods; the paper also has a clean ablation design, which are consistent and roughly additive.

Critical weaknesses
  1. The paper never tests what happens when one expert is weak, broken or redundant. The rule keeps every expert at 25–50% weight.

  2. All results use the same three experts. There are no experiments varying the number of experts, swapping in different experts. It is therefore unclear whether the rule generalizes beyond this particular bank.

  3. We never see when the safety fallback actually triggers. So cannot tell how often the method reverts to equal weights, or whether the safe default matters in practice.

  4. It seems that online expert aggregation already offers adaptive weights with stronger guarantees, what is the unique contribution compared with the following method? https://github.com/dralliag/opera

  5. The baselines appear substantially under-trained (many baselines are far worse than their published results), so the state-of-the-art claim is suspicious.

Questions for authors

Figure 1 contains too much content and is hard to interpret. I suggest plot with larger font, remove the redundant parts, and only convey essential information.

1 comment

1. Weak, broken or redundant experts. Tables 20-21 test bias, noise, persistence, zero output, stale reports and NaNs with trained parameters/scales fixed. At +10 history-SD bias, DAShrink/plain-router MSE is 8.164/9.276 on ETTh2 and 18.217/8.212 on ETTm1. With reports (0,0,M)(0,0,M) and target zero, the weight floor implies loss at least M2/16M^{2}/16. The routing bound fixes the reports; fault tests measure changes to those reports. NaNs at every tenth origin give 90% valid-origin coverage, counting failed origins in its denominator; full-test MSE is undefined. Table 8 includes exact duplication, which changes the uniform reference (D.6).

2. Expert number and identity. Tables 8-10 independently refit combiners on eight banks of 2-5 experts, with three seeds per ETTh2/ETTm1 H96 task. Banks include all three two-expert subsets, three independent DLinear models, two DLinear models plus TimeFilter, five experts and exact duplication. Each bank/task/seed shares identical frozen forecasts across rules. Mean paired MSE reductions over each bank's uniform average are 0.444% for homogeneous DLinear and 2.348% for five experts. The homogeneous bank tests the rule without the contextual expert.

3. Fallback frequency and effect. In the 48 disjoint-calibration bank-task-seed units, 9 select zero, all on ETTh2. The core-bank rates are 2/3 on ETTh2 and 0/3 on ETTm1. Holding the ETTh2 routers fixed, selected scales give .9703 mean test MSE versus .9721 with scale forced to one. The fallback selects the equal-weight endpoint of the correction path. Equation 6 and Appendix C.4 identify how residual alignment and correction magnitude determine that path; Tables 11-13 report the selections.

4. Contribution relative to OPERA. DAShrink constrains proposal-induced forecast movement at issuance. Expert spread sets a budget at every lead/channel, enforced after routing; validation chooses its task amplitude, with zero giving exact averaging. For fixed reports, Equation 12 maps a proposal-TV radius to a forecast interval. A threshold outside the interval has an invariant decision before the outcome arrives. OPERA supplies feedback-based cumulative-loss guarantees over time; DAShrink adds this issue-time control.

Proposition 2 identifies the information supplied by spread. Close reports and cancellation among distant reports can share the same mean/correction. For zero-preserving scalar rules observing only that correction and fixed history/mean context, matching worst-case sensitivity over unclipped reports limits the unit-spread correction to one eighth of DAShrink's, under the stated parameter and clipping-margin conditions. Table 7 separately compares OPERA and neural gates empirically. On 237,600 unclipped ETTm1 step-96 coordinates, at calibration 90th-percentile thresholds and TV radius .10, DAShrink's pooled worst-case error count is 17,258, below this class's outcome-informed unperturbed lower bound of 17,439; the gap holds in each seed (Figure 8; Table 19). Frozen/delayed BOA have lower full-test mean MSE on both tasks. The decision comparison connects report geometry to the adaptation budget.

5. Baseline training and comparison basis. On Traffic H96, iTransformer MSE is .664 in Table 1 and .395 in its published Table 10. Tables 1-2 report the displayed system-level comparison. The rule-level evidence fixes the expert bank: Table 7 independently fits each trainable combiner on identical forecasts under six disjoint blocks and expert-fit-only scaling. Plain routing and the two shrinkage heads share the router architecture. Neural controls use AdamW at 10−310^{-3}, weight decay 10−410^{-4}, gradient limit one, a 30-epoch cap, patience seven and router-stop MSE selection. Plain routing fits MSE; the shrinkage heads fit MSE plus .05 KL. Static stacking solves the simplex MSE QP to a KKT residual below 10−510^{-5} (D.1). Table 7 reports three-seed MSE/MAE mean/SD, including TimeFuse and OPERA. Expert quality and expert cost are shared across these rules. This comparison evaluates the combination mechanism; the coordinate-wise movement bound and matched-sensitivity result establish its distinct capability.

6. Figure 1. We retain the original expert motifs, curves and simplex, enlarge the admissible-weight region and labels, align the proposal/control branches, and distinguish original-unit fusion from normalized routing inputs. Figures 3 and 8-9 explain scale selection and the sensitivity mechanism. We ask you to reassess novelty and soundness using the explicit coordinate-wise constraint, its matched-sensitivity analysis and these controlled comparisons.

Comments on the paper

1 comment

The revision separates combination effects, expert identity and calibration through shared forecasts and independent fits. Sections 3-4 and Appendix D connect these controls to DAShrink's budget on forecast movement before the outcome is known.

1. Matched combiners and computational budget. Rules share frozen forecasts per task/seed/bank. Uniform uses 1/K1/K; static stacking solves a simplex MSE QP. Plain routing independently minimizes MSE; constant shrinkage and DAShrink independently minimize MSE plus .05 KL, sharing the router architecture, 30-epoch cap and router-stop checkpoint selection (D.1). Table 7 gives three-seed MSE/MAE means and sample SD for ETTh2/ETTm1 H96. MSE mean (SD):

RuleETTh2ETTm1
Uniform.9711 (.0146).9285 (.0277)
Static stack.9780 (.0362).8796 (.0099)
Plain router.9548 (.0249).7943 (.0137)
Constant shrinkage.9665 (.0175).8768 (.0223)
DAShrink.9703 (.0157).8946 (.0265)

DAShrink improves on uniform in four of six runs and ties in two (1.87% mean paired MSE reduction). The plain router's lower MSE accompanies larger departures: seed-mean p99 absolute departures from uniform on unclipped coordinates are .017/.081/.306 history SD on ETTh2 and .102/.538/1.397 on ETTm1 for DAShrink/constant/plain. These readouts quantify accuracy and departure jointly; Equation 12 separately bounds sensitivity to proposal perturbations (Section 4.3). Same-bank comparisons share expert cost. Table 22 reports DAShrink full-bank median latency of 65.03/78.39 ms at batch one with resident inputs on the RTX 3070 Ti Laptop GPU; the cached head takes about .35 ms.

2. Contribution of the contextual expert. Table 7 includes every standalone expert, both metrics and seed SD. Their MSE mean (SD) is:

ExpertETTh2ETTm1
Contextual1.1450 (.0751)1.0158 (.0328)
DLinear1.1201 (.0343).8942 (.0058)
TimeFilter.9338 (.0157)1.2655 (.1085)

Adding the contextual expert to the same DLinear/TimeFilter pair changes uniform MSE from .9416 to .9711 on ETTh2 and .9687 to .9285 on ETTm1. Its contribution therefore varies by task. Eight refitted banks cover two to five experts, including independent DLinear models and exact duplication (Tables 8-10). The homogeneous DLinear bank evaluates the combination rule without the contextual expert. These controls separate expert identity from the rule applied to its forecasts.

3. Independent validation and temporal dependence. The follow-up uses six chronological, disjoint target/history blocks: expert fit, expert stop, router fit, router stop, scale calibration and test. Dataset standardization uses expert-fit data only. With checkpoints fixed, 999 resamples use aligned 96-origin blocks across all five candidate scales. The penalty uses the per-resample maximum centered mean across nonzero candidates, accounting approximately for dependence and joint screening (D.3).

Holding the rebuilt router/scaler fixed, reused-stop 1.645, disjoint 1.645 and disjoint block scores select ETTh2 scales (1,0,0)(1,0,0), (1,1,1)(1,1,1), (0,0,1)(0,0,1), with test MSE .9715/.9721/.9703. ETTm1 selects (1,1,1)(1,1,1) throughout, giving .8946 (Table 13). Thus the follow-up directly separates checkpoint selection from scale selection and tests the effect of their temporal reuse.

4. Variability and the Table 1/Table 3 difference. Table 7 reports training variability across seeds 2027-2029 for the two H96 tasks; Figure 7 fixes errors and resamples dataset composition. ETTh2 H96/H192 MSE is .217/.266 in Table 1 and .228/.296 in Table 3's Full row. A cross-table difference alone does not identify a run, checkpoint or protocol effect. Component effects use Table 3's own Full row: at H96, joint removal raises MSE from .228 to .242 (6.14% using printed values), with experts and router fixed. Table 7 then independently refits the rules under declared splits, scaling and seed protocols (D.7). This distinguishes fixed-parameter mechanism tests from retrained-combiner comparisons.

The methodological value is an enforced, spread-dependent correction budget. Proposition 2 compares its defined scalar rules at matched sensitivity: spread distinguishes close reports from cancellation among distant reports with the same aggregate correction and history/mean context. For unclipped ETTm1 H96 final-step coordinates, at calibration 90th-percentile thresholds and TV radius .10, DAShrink's worst-case error count is below that class's outcome-informed, unperturbed lower bound in each seed (Figure 8, Table 19). This links the constraint to a decision consequence. We invite reassessment of the contribution using the matched attribution, independent calibration and explicit issue-time guarantees.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.