acceptodds
Under review as a conference paper at ICLR 2027

CoFlow: Leaf-Space Factorization for Point and Probabilistic Forecast Reconciliation

Abstract

Forecast reconciliation is usually posed as a point-adjustment problem. When forecasts must also quantify uncertainty, coherence must extend to the joint predictive law, while its center may not optimize the target point-forecast loss. We introduce CoFlow, a post-hoc leaf-space framework that separates a coherent anchor, a joint error law, and a separately optimized point forecast. Mapping both error draws and point corrections through the summing matrix guarantees coherence while allowing predictive shape and point accuracy to be selected independently. This factorization also exposes aggregate dispersion as a central calibration problem. For Gaussian node errors with any predictive-to-true variance ratio, we theoretically derive exact expressions for coverage and expected CRPS and interval-score regret, together with quadratic lower bounds. Under attenuated cross-leaf covariance, these penalties increase with the share of aggregate variance carried by dependence. CoFlow combines Gaussian, copula, count, empirical, and flow-based leaf laws with coherent attention-based and sparse constraint-routing point heads. Across 29 hierarchies, three rolling origins, and five seeds, Tweedie with routing improves relative MAE by 9.8%, nCRPS by 6.2%, and normalized interval score by 16.0% over MinT. An RBIG-Gaussian mixture improves nCRPS on all 29 datasets and both nMIS and coverage gap on 28. Ablations isolate gains from dependence, marginal shape, and point estimation.

Primary area:
learning on time series and dynamical systems
Keywords:
Forecast Reconciliation, Probabilistic Reconciliation, Coherent Subspace, Leaf Space Distribution Modeling, Leaf Space Factorization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_diexOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the problem of reconciling predictions with hierarchical dependencies. For example, total sales in a region must be the sum of the sales for each store in the region, but predicting all of them individually will almost certainly violate that assumption. The paper claims that a baseline solution does not have calibrated prediction intervals due to (co)variance estimation error. The paper develops an alternative method, and contains experiments on many datasets, where the paper's method outperforms baselines on several metrics.

Critical strengths

None

Critical weaknesses

The paper is extremely hard to understand. Many critical concepts are not explained when they are first used, and some are not explained at all. Examples include:

  • Theorem 1: none of the left-hand-side quantities are defined or explained.
  • Corollary 1: rar_a and rqr_q are not defined or explained.
  • The paper claims that separating the "predictive sample" from the "point forecast" is an important part of the method, but does not explain why this is useful.
  • Section 3.5: it is not clear how the point forecaster is constructed. The text presents four options (but Table 1 has six?) at a high level, but the details later in the section are not connected to any of the options.
  • Eq. 4: δθ\delta_\theta is not explained.
  • The terms "anchor" and "centering" are used repeatedly but not explained.
  • Line 44: W is not defined.
  • Line 46: "node wise variance ratio" is not explained.
  • Line 55: H and S not defined.

These make it impossible to learn how the method works and judge what is actually proved in the theorems.

The experiments also have several issues:

  • One of the stated motivations is that MinT prediction intervals do not always have nominal coverage. However, the experiments only report coverage in an ablation (Table 3) but not in the main results of Table 2. The reported coverage is also relative to MinT, while the absolute coverage of MinT is not reported. As a result, the claim that negative relative coverage is closer to nominal could be false.
  • The metrics listed in Section 4 do not fully match those reported in Table 2. The metrics should also be cited.
  • Section 4 claims 72 methods are evaluated, but Table 2 only contains 12. The text claims the method list is in the Appendix A.2, but that does not contain 72 methods.

Minor points:

  • Figure 1 text is too small.
  • Line 207: the citation makes no sense. The text talks about "information patterns", but cites the original transformer paper.
Questions for authors

No questions.

1 comment

Thank you for the review and for the specific list. Most items are symbols used before they are defined, which we will fix. The paper is densely packed and assumes some familiarity with hierarchical forecasting, and we're sorry that made it hard to follow.

Definitions.

  • W, S, H and the nodewise variance ratio (lines 44, 46 and 55): the introduction uses them before Sec. 2 defines them. W is the base-error covariance. S is the summing matrix that builds every series from the leaves, y = Sb. The rows of H encode the aggregation rules (e.g. a region minus the sum of its stores), so Hy = 0 means all rules hold. The ratio is r in Theorem 1, predicted over true variance. We will keep the introduction symbol-free.
  • Theorem 1: fair. "Cov" is the coverage of the central (1−α)(1-\alpha) interval, a poor name next to covariances. ΔCRPS(r)\Delta_{CRPS}(r) and ΔMIS(r)\Delta_{MIS}(r) are expected regrets, the expected CRPS or interval score of N(0,rv)\mathcal{N}(0, rv) minus that of the true N(0,v)\mathcal{N}(0, v). We will define both in the statement. The theorem proves the exact coverage and score cost of a wrong variance, and for r < 1 the cost is at least proportional to (1−r)2(1-\sqrt{r})^2.
  • rar_a and rqr_q: rar_a is Theorem 1's rr at aggregate node aa, and Eq. 12 gives its value from θa\theta_a and κˉa\bar{\kappa}_a, both defined in the statement. rqr_q is the same ratio for a node summing qq leaves with equal variances and correlations. We will name them as variance ratios explicitly.
  • Anchor and centering are defined in Sec. 3.1. The anchor b^\hat{b} is the MinT leaf forecast the samples are built around. Centering subtracts the mean of the K error draws, so the predictive mean is exactly Sb^S\hat{b}. These are also standard terms in hierarchical forecast reconciliation.
  • δθ\delta_\theta (Eq. 4) is the point head, defined at the start of Sec. 3.5, a neural network with parameters θ\theta that maps features f to a leaf correction.

Why separate the samples from the point forecast.

  1. The point metrics are absolute-error losses, minimized by median-type forecasts, while CRPS scores the whole distribution. A single forecast can't be optimal for both.
  2. It makes each choice testable on its own. Across five point heads the predictive scores are identical (nCRPS 0.983, Table 12), and across two error distributions point accuracy is unchanged (relMAE 0.985).
  3. The gains compose. Tweedie plus routing improves all four metrics at once.

In practice (say sales/demand forecasting), planners use a point forecast for replenishment and quantiles for safety stock, and both have to add up.

Sec. 3.5 and Table 1. There are four information patterns: a per-leaf MLP, attention among leaves (LeafAttn), attention among all nodes in three masking variants (FullAttn, TreeAttn, LeafTreeAttn), and hierarchy routing (RouteAttn). That gives the six rows. Every head outputs a leaf correction δ\delta, and the point forecast is S(b^+δ)S(\hat{b}+\delta). The text after Table 1 details RouteAttn, the head used in Table 2. Each aggregate's discrepancy gag_a (its direct forecast minus the anchor sum) is split among its descendant leaves by softmax weights that sum to one (Eqs. 14-15). We will add a one-line summary per head.

Coverage. Fair point. Absolute 90% coverage, averaged over the 29 datasets:

TotalMiddleBottomAll
MinT.36.65.71.70
Bottom-up.44.74.77.76
CoFlow-Gauss.43.73.75.75
CoFlow-Tweedie.49.72.72.72
CoFlow-RBIG+Gauss.45.75.78.77
HINT-BU.62.85.88.88

CovGap in Table 3 is ∣coverage−0.90∣|coverage - 0.90| per dataset and origin, so a negative difference from MinT means closer to 90%, from above or below. The table also shows a limitation plainly. Every residual-based method under-covers, ours included, because errors on the training data understate errors on new data. MinT's Total coverage is below 0.80 on 28 of 29 datasets, and CoFlow narrows the gap without closing it. Coverage alone isn't the target either: HINT covers best but has 1.7 times MinT's CRPS. This table goes into the main text.

Metrics and the 72 methods. Table 2 shows the four metrics. MASE, coverage gap and energy score are in Tables 3, 4 and 15. We will gather all seven in one table with citations (Gneiting & Raftery 2007 for CRPS, interval and energy scores, Hyndman & Koehler 2006 for MASE, Davydenko & Fildes 2013 for relMAE). The 72 methods are 13 baselines (8 standard reconcilers including the unreconciled base, 3 probabilistic MinT variants, 2 HINT configurations) and 59 CoFlow variants, which make up the factor ablations. Table 2 shows ten baselines plus the two final compositions, and the other three are in Table 14. A.2 doesn't list all 72, so we will add the full list.

Minor. We will enlarge the text in Figure 1. The Vaswani et al. citation belongs where attention is introduced, not on "information patterns", and we will move it.

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

The paper claims to formulate forecast reconciliation in leaf space, factorizing the coherent anchor, joint leaf-error law and point forecast. It derives a dispersion-regret theorem and introduced attention-based heads that redistribute aggregate discrepancies. Numerical experiments show the resulting compositions achieve the best point and probabilistic scores.

[If you are lost by my summary above, you are not alone: I don't understand it either.]

Critical strengths

I did not find any positive aspects of this draft.

Critical weaknesses
  • very poorly written, to the extent of impossible to understand.
Questions for authors

The presentation of this work is terrible, and I wonder if the authors actually put in any serious effort in writing it.

The problem under study was never formally defined. The background Section 2 is hardly of any use. For example:

  • line 077, what does "coherent" vector mean?
  • line 080, what does "in-sample residual" mean?
  • line 082, "base-error covariance"?
  • What are the acronyms MAE, EMAPE, MASE, nCRPS, etc. mean? Why do we care about them?

I don't know who can possibly understand the problem setup after reading Section 2. I don't even know what are the inputs or outputs...

More examples:

  • line 101, leaf anchor b^=JPλy^\hat b = JP_\lambda \hat y: what is J here? what does a joint law ν\nu mean?
  • line 106, did you ever define δθ\delta_\theta or ff?
  • line 126, what is a coherent law? what does the notation S_#\nu mean when SS is a matrix as opposed to a transformation?
  • line 129, what do you mean by a Gaussian "base law"?
  • line 138, what do you mean by the claim "this reduction is calibrated"? How is calibration defined in the first place?
  • Theorem 1: what does Cov(r, \alpha) mean when \alpha appears to be just a nominal number, and rr is the ratio of two variances? How are ΔCRPS\Delta_{CRPS} and ΔMIS\Delta_{MIS} defined? What are the consequences of Theorem 1, for the problem you are studying?
  • Corollary 1, Eq (12), what is r_a? In line 183, how can you defer "exact definitions" to Appendix A.1.3?
  • Section 3.4 makes little sense to me. It seems like the authors just put a lot of fancy terms there, without a clear goal. Again, full definitions are deferred to Appendix A.1.4?
  • Eq (17) what does wmed mean?

I can keep listing the presentation issues on and on, but I don't think I need to. The authors should seriously revise their draft before next submission (if at all).

1 comment

Thank you for the review. Your list mixes two kinds of items, and it helps to separate them: standard terms of the field, and notation that is ours. The second group has real gaps, and we will fix every one.

Standard terms. The paper is written for the forecast reconciliation audience. Coherence, base forecasts, in-sample residuals, the base-error covariance and MinT are that literature's basic vocabulary. They are introduced in the first pages of every reconciliation paper we cite and in the standard textbook (Hyndman & Athanasopoulos, Forecasting: Principles and Practice). MAE, MASE and CRPS are textbook forecast metrics. We don't think re-deriving them is a good use of nine pages, but briefly:

  • Coherent (line 077): the forecasts add up, so each aggregate equals the sum of its children. Sec. 2 gives the algebraic form, y = Sb.
  • In-sample residuals (line 080): the base model's fitted errors on the training window. Base model can be any forecaster, a neural network, gradient boosting methods, statistical forecasting, etc.
  • Base-error covariance (line 082): W, the covariance of those errors, estimated by Eq. 3.
  • Metrics: MAE is mean absolute error (just like MSE which is mean squared error!), WMAPE weighted mean absolute percentage error, MASE mean absolute scaled error, CRPS the continuous ranked probability score and MIS the mean interval score. The prefix n means normalized by total volume, and relMAE is MAE relative to the base forecast. We report point metrics because the method outputs a point forecast, and proper scores because it also outputs a distribution.

Inputs and outputs: Sec. 2 states the inputs (S, base forecasts for all n series from any model, and their in-sample residuals). Eq. 4 states the two outputs (one coherent point forecast and K coherent sample paths). This is the standard post-hoc probabilistic reconciliation setup (Panagiotelis et al., 2023). We will still pull it into one explicit problem statement at the top of Sec. 2.


Our notation. Here your points are mostly fair:

  • J (line 101) is used in Sec. 3.1 but defined one subsection later. It selects the leaf rows, J=[0Im]J = [0 \quad I_m]. We will move it up.
  • δθ\delta_\theta and f (line 106) are the point head, a network with parameters θ\theta that maps features f to a leaf correction. Both are defined in Sec. 3.5, and Eq. 4 needs a forward pointer.
  • Law is standard probability usage for distribution. ν\nu is the joint distribution of the m leaf errors, and a coherent law is one whose draws are all coherent.
  • The pushforward (line 126) is the distribution of Sb when b∼νb \sim \nu. A matrix is a linear map, so the notation applies as is.
  • Gaussian base law (line 129): base errors distributed as N(0,W)\mathcal{N}(0, W).
  • Calibrated (line 138) is meant in the usual sense of Gneiting, Balabdaoui & Raftery (2007): outcomes are consistent with the predictive distribution, so 90% intervals cover 90% of the time. The sentence says that with Gaussian errors and the true W, MinT's covariance is exactly right. Our argument is about the estimated W.
  • Theorem 1: this one is on us. "Cov" is coverage, not covariance, which is a poor name in a paper full of covariances, and we will rename it. α\alpha is the miscoverage level (0.1 for 90% intervals) and r is predictive over true variance. ΔCRPS(r)\Delta_{CRPS}(r) and ΔMIS(r)\Delta_{MIS}(r) are expected regrets, the expected score of N(0,rv)\mathcal{N}(0, rv) minus that of the true N(0,v)\mathcal{N}(0, v). The consequence: any variance error costs coverage and score, at least quadratically in 1−r1-\sqrt{r}. Corollary 1 then shows the cost grows at aggregate nodes when cross-leaf covariance is attenuated, which is why CoFlow fits the leaf covariance directly.
  • rar_a (Eq. 12) is Theorem 1's r at aggregate node a. Its value follows from θa\theta_a (the share of the node's variance that comes from cross-leaf covariance) and κˉa\bar{\kappa}_a, both defined in the corollary statement.
  • Sec. 3.4 has one goal: list the candidate distributions for ν\nu, grouped by the property each one changes (dependence, marginal shape, zeros and skew, flexible joint shape). That lets the ablations in Sec. 4.2 attribute each gain to one property.
  • wmed (Eq. 17) is the weighted median. We will define it inline.

This was a packed submission and it was hard to fit everything in 9 pages. We also assumed certain level of familiarity with the topic, i.e. hierarchical forecast reconciliation and forecasting in general. But we understand that it might be hard to follow for someone not familiar with the topic. The revision will add an explicit problem statement and a notation table, spells out acronyms at first use, and removes every main-text dependence on appendix definitions.

Ethics flag: this one genuinely surprised us. Unclear writing is one thing (and a fair reason for low score), but ethics concern is a different thing. We kindly ask that the flag be withdrawn.

Reviewer_7bxLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CoFlow proposes to do post-hoc reconciliation in leaf space. The predictive samples include a MinT-shrink leaf forecast and a leaf-error model. A separate point head are coherent by construction. The paper derives the coverage and score cost of a wrong Gaussian variance, and also relates it to the share of aggregate variance coming from cross-leaf covariance. In experiments, the paper shows that Tweedie with routing head beats MinT-shrink on relMAE, nCRPS, and NMIS, on 29 hierarchies.

Critical strengths
  1. The proposed method can work with any base forecaster, and guarantees coherence for every sample and point forecasts.
  2. RouteAttn is a nice design, where each aggregate discrepancy is split among its descendant leaves with corresponding weights summing up to 1, and the cost scales. This can help WMAPE where dense attention cannot.
  3. The statistics are done carefully. Experiments are per-dataset, and the per-dataset spread and level-wise results are reported. The ablation studies are controlled.
Critical weaknesses
  1. Model selection and size of the gains. It seems like each dataset only has three test origins, and I cannot find separate origins used for model selection. Choices such as Tweedie over ZINB in Tab. 6 are justified with ablation results, but it's not clear to me whether these ablations use the same origins as Tab. 2. In App A.4.6, it says the final model 'is evaluated once on untouched origins', which is why it seems a bit confusing to me. Besides, the MinT-shrink looks worse than Base on relMAE in Tab 2, so the headline numbers might have overstated the gain on point accuracy? Looking at result tables in tab. 15-20, when comparing aginst WLS-var, the relMAE gain shrinks. And the results on six public datasets are also mix, showing a public-only version of Table 2 might help with the claim. In summary, to me, the evaluation may overestimate the improvement because the model-selection protocol is unclear, and the headline gains are measured against a baseline that is not always the strongest.

  2. Are the probabilistic gains from dependence modeling or just wider intervals? The theory itself shows what a wrong variance can cost, but it does not show MinT's predictive is actually too narrow on those data, and the rar_a and θa\theta_a are also not reported. In Eq. 5, the covariance reduction DWD_W is correct only when WW is correct. So, the question becomes whether it is too large in practice. CoFlow-Gauss uses the raw covariance WbbW_{bb} instead of CWC_W, so no node variance can be smaller than MinT. Since EE is made of in-sample residuals, this can underestimate out-of-sample errors (this is also mentioned in the appendix 4.7). Besides, I found two results suggesting the gain might partly come from intervals: a. In Table 7, fitting on MinT-projected residuals loses most of the gains; b. In table 15, plain bottom-up already gets close to RBIG+Gauss. It seems the current experiments does not separate three possible sources of gain: overall variance scale, covariance structure, and marginal shape. I would suggest a simple check: do MinT with a variance scale that is fitted on a held-out origin (the appendix A.4 also suggests this.) In summary, the paper shows better probabilistic scores, but does not isolate where the improvements come from, i.e., from global variance calibration or richer probabilistic modeling.

  3. Decoupling and novelty. While samples are centered at Sb^S\hat b, the predictive distribution will stay at MinT even when the point head gives a better forecast. It is reasonable to keep a separate point forecast, but the paper does not try centering at Sb^+δθ(f)S\\{\hat b + \delta_\theta(f)\\}. This would show whether the point head gains also help nCRPS and nMIS. Besides, (Panagiotelis et al., 2023) has studied similar ideas about writing coherent distributions as pushforwards S_\\#\nu of bottom-level distributions. To me, the main new contributions appear to be the residual modeling, the RouteAttn and the large modular study, instead of the leaf space representation itself. I would suggest authors to state more clearly about the comparison to prior work.

Questions for authors
  1. In Table 2, about the ablations used to choose Tweedie, RBIG+Gauss and λ\lambda, were they run on origins separate from the test origins? Could you provide an additional version of the table 2 on the six public datasets? Is it possible to consider some other useful baselines, e.g., (Panagiotelis et al., 2023), or (Ben Taieb et al., 2017), etc.
  2. How does the CoFlow-Gauss compare to MinT scaled by a var factor fitted on an held-out origin? Could you consider to report absolute 90% of coverage by level (Total, Middle, Bottom) for MinT, Bottom-up, and the two CoFlow models?
  3. How about centering the samples at Sb^+δθ(f)S\\{\hat b + \delta_\theta(f)\\} instead of Sb^S\hat b? What would happen to nCRPS, nMIS and coverage, etc? Could you also consider reporting the training/inference times of the heads and flows compared to MinT-shrink on M5?
2 comments

Thank you for a careful, technical review.

  1. Model selection and gain size. You're right, and the A.4.6 sentence can be misleading. It describes the protocol we recommend, not the one we ran: the ablations and Table 2 share the same three test origins. Why the effect on our conclusions is small:
  • Our claims about where the gains come from rest on planned comparisons that change one component at a time (e.g. diagonal vs full covariance). No winner is picked by test score, so there is no best-of-many bias.
  • The gains are broad. 17 of the 18 raw-residual leaf laws in Table 3 beat MinT on both nCRPS and nMIS, and every routing λ\lambda in Table 11 beats MinT on relMAE and WMAPE.
  • We picked Tweedie over ZINB despite ZINB's better pooled mean (nMIS 0.745 vs 0.840) because Tweedie was more stable.
  • An unreported nested study on the public sets picks λ\lambda on 1 to 4 inner origins and scores once on the untouched outer origin (3 seeds). It gives relMAE 0.982 and WMAPE 0.981 vs MinT, close to fixed λ=0.5\lambda = 0.5 (0.969 and 0.994).

We will also re-score both frozen compositions on three earlier origins that no decision used.

On baselines, agreed. Against WLS-var, Tweedie-Route gives relMAE 0.964 (p=0.31, not significant), WMAPE 0.948 (p=0.009), nCRPS 0.960 (p=0.064) and nMIS 0.878 (p=0.002). The abstract will quote gains against the strongest baseline for each metric.

Public-only Table 2 (ratio to Base, wins vs MinT out of 6):

relMAEWMAPEnCRPSnMIS
BU.990.858.847.812
WLS-var.984.797.799.809
MinT.993.784.791.825
Tw-Route.L50.952 (4).773 (2).726 (5).669 (5)
RBIG+G-Route.L25.956 (4).739 (2).765 (6).688 (6)

The predictive gains hold on public data. RBIG+G wins 6 of 6 on both scores (p=1/32, the minimum possible with 6 datasets). The point gains come mostly from Wiki2 and M5, so we will soften the point claim.

  1. Width vs. structure. Absolute 90% coverage (Total/Middle/Bottom) on all 29 datasets:
  • MinT: .36/.65/.71
  • BU: .44/.74/.77
  • Gauss: .43/.73/.75
  • Tweedie: .49/.72/.72
  • RBIG+G: .45/.75/.78

The compositions' draws are exactly the Tweedie and RBIG+G draws. MinT's Total coverage is below 0.80 on 28 of 29 datasets. Inverting Eq. 7 gives a median r of about .08, .33 and .51 (lower bounds, since bias also lowers coverage). Every residual-based predictive under-covers, Base included (.76), as your in-sample-residual point predicts. So yes, part of the gain is dispersion, and CoFlow narrows the gap without overshooting.

The ablations separate the three sources at a fixed center:

  • DWD_W: Gauss and Gauss.reconres differ only in whether DWD_W is already in the residuals. With it, Total coverage falls from .43 to .36 (= MinT) and nMIS rises from .925 to .975. So Table 7 measures Eq. 5 directly.
  • Covariance: Diag and Gauss share their leaf marginals (Bottom coverage .752 for both). Their Total nMIS relative to MinT is 1.12 vs .90.
  • Shape: Tweedie has narrower leaf intervals than Gauss (Bottom coverage .715 vs .752), yet its Bottom nMIS is .86 vs .95 and nCRPS .94 vs .99 (ratios to MinT). That gain is not width.

Bottom-up in Table 15 tells the same story. Its Gaussian predictive covariance is SWbbS⊤S W_{bb} S^\top, the unconditioned leaf covariance that Gauss also uses, up to the shrinkage intensity (nMIS vs MinT: CF-Gauss-BU .924, BU .922, Tables 3 and 14). CoFlow keeps that dispersion but has a better point (WMAPE .88 vs .96 of Base), and Tweedie reaches nMIS .840.

Agreed that scaled MinT is the clean control. We will add it with a global and a per-level scale, each fitted on the preceding origin, and report θa\theta_a by level.

  1. Decoupling and novelty. Centering at S(b^+δ)S(\hat{b}+\delta) is free to test (the draws shifted by SδS\delta stay coherent), and we will report it. We anchored at MinT to fix location across ablations. Also, the heads minimize absolute-error losses, whose minimizers are median-type, so moving a skewed law's mean onto them is not obviously right. We will test both mean and median alignment.

On Panagiotelis et al. (2023), agreed: the pushforward is theirs, and we will say so at Lemma 1. Our difference is fitting ν\nu directly on leaf errors rather than pushing an all-node law through a projection (which imports DWD_W), plus Theorem 1 and RouteAttn. Normality-MinT and Bootstrap-MinT in Table 2 are their Gaussian and bootstrap methods.

PERMBU (Ben Taieb et al., 2017) ran on every dataset except the grouped Tourism Large. On those 28, Tw-Route beats it on nCRPS .925 (24/28), nMIS .836 (24/28), energy .907 (22/28) and WMAPE .925 (18/28). We will add PERMBU and score-optimised reconciliation.

M5 runtime (median per origin, including about 30 s each method spends drawing and scoring 200 samples): MinT 34 s, Gauss 44 s, Tweedie 37 s, RBIG+G 70 s, Tw-Route 36 s, HINT-BU 55 s. Routing adds a one-off 4 base-model refits for its training origins.

I really thanks the authors for their detailed response and new results. For W1, the admission about the origins, the table and the coverage numbers are helpful. For W3, the additional results on PERMBU also seems good. Authors promised to add a few experiments, including the scaled-MinT control experiment, and the centering at corrected point, etc. I may incline to raise my score if CoFlow wins (in terms of nCRPS and nMIS) on that scaled-MinT control experiment. Though, given that the rebuttal period is reaching its end, I may not see those results.

I also take the time to read reviews from other reviews. I believe every research has its own significance. A senior researcher once reminded me that every strong manuscript eventually finds its home, and random factors often play a role along the way. I hope authors are not discouraged by the scores of this submission. Finger crossed!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.