acceptodds
Under review as a conference paper at ICLR 2027

STGraphFM: Prior-Anchored Flow Matching for Joint Spatiotemporal Forecasting

Abstract

Spatiotemporal forecasting requires accurate central predictions, informative uncertainty estimates, and coherent future scenarios across interacting locations. These goals call for a model that captures predictable structure while organizing variation across nodes and forecast leads. We propose STGraphFM, a prior-anchored flow-matching framework that unifies deterministic and probabilistic forecasting. A node-wise temporal prior provides a future-shaped predictive anchor, graph-correlated perturbations initialize structured alternatives around it, and a sample-dependent graph-time velocity field transports each alternative across nodes and forecast leads. Bounded velocity regression trains this field along sampled source–target paths without unrolling generation. The same learned transport maps the anchor to a central forecast and perturbed sources to complete future trajectories. Across traffic, air-quality, and gridded-temperature benchmarks, STGraphFM achieves strong point accuracy and improves probabilistic forecasts, including 18.4–45.8% lower continuous ranked probability score (CRPS) than the strongest evaluated probabilistic baselines on traffic. Matched air-quality comparisons retain joint-score gains at fixed coordinate-wise empirical marginals, providing evidence beyond marginal accuracy.

Primary area:
learning on time series and dynamical systems
Keywords:
spatiotemporal graphs, flow matching, probabilistic forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CAj9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the probabilistic forecasting modeling problem. The main methodology includes three parts. First, using history sequence to generate a deterministic predictive anhor. Then, add noise to the anchor. Finally, by graph-time velocity field and flow matching, map these initial states into future prediction. The experiments are comprehensive and results are significant.

Critical strengths
  1. The experiments are really comprehensive. This is worthy encouraging.
  2. This paper is well-orgainized and presented. It contains 40 pages and 32 tables.
Critical weaknesses

The experiments are really comprehensive, the limitations come from novelty and methodological contribution that are required by ICLR.

  1. All components are existing tools. Predictive prior, correlated source, GAT and flow matching are already utilized in many related fields. The loss includes three parts: prior, nll and CFM surrogate. It doesn't introduce any generative modeling theory. For example, authors can consider intorducing dynamics mechanisms to regularize the loss.
  2. According to the core of flow matching that transports noise to conditional distribution, this paper seems only use one-step residual, but not the ODE transport. The problem comes from the std from prior nll. Therefore, CRPS of 1/2/3 are 12.6744/12.6745/12.6747, indicating the learned speed is constant. Authors need to further check about this problem.
  3. Based on W2, the noise therefore is directly propagated.
Questions for authors

See Weaknesses.

6 comments
AuthorsOpenReview ↗

Part 1 of the Response (1/3)

W1. What is the contribution beyond existing components?

Thank you for recognizing the breadth of the experiments and for asking us to clarify the methodological contribution. We have revised the Introduction and method presentation to connect the task-specific design with the controlled evidence supporting it.

The method separates three forecasting responsibilities. STGraphFM first estimates the predictable future background, then introduces graph-structured uncertainty around it, and finally learns how each perturbed future should evolve through a sample-state-conditioned graph–time field. The contribution is this coordinated adaptation for probabilistic spatiotemporal graph forecasting, rather than the invention of each underlying operator.

The objectives support these roles. SmoothL1 supervises the predictive center, Gaussian likelihood estimates its node–horizon uncertainty scale for source construction, and bounded velocity regression trains the sample-specific transport. The flow branch uses detached prior outputs, preserving the division between learning the background and updating future scenarios (Section 4.4; Appendix A.1).

Controlled studies support the design. Predictive centering improves mean CRPS by 12.90% with graph noise and 17.92% with IID noise. At identical predictive-center marginals, graph initialization improves GraphVS by 10.41% and regional interval score by 11.28%. Shared retraining and the fixed-weight intervention further support the predictive value of sample-state-dependent updates (Sections 4.1–4.3 and 5.3; Appendix C.1 and C.3).

Additional forecasting comparisons. We adapted GiFlow and FlowCast to the corresponding STGraphFM forecasting configurations. On KnowAirV2, STGraphFM achieves MAE/RMSE of 13.324/22.602, compared with 24.238/40.156 for our GiFlow-F adaptation. On WeatherBench T2M, STGraphFM achieves 1.272/2.164, compared with 1.389/2.331 for our FlowCast-WB adaptation, corresponding to 8.37% lower MAE and 7.19% lower RMSE (Appendices F.1 and F.3).

Physical penalties are domain-specific. We appreciate the suggestion. The benchmarks span traffic, air quality, and temperature, for which a common physical penalty would require additional domain-specific assumptions. We keep the current contribution focused on graph–time statistical generation and identify richer physical context as a natural extension.

AuthorsOpenReview ↗

Part 2 of the Response (2/3)

W2. Is one-step decoding only residual prediction rather than Flow Matching?

Thank you for questioning whether the default decoder uses the learned flow field meaningfully. We clarify the distinction between path-state-conditioned training, the one-evaluation inference rule, and the prior’s role in setting the initial perturbation radius.

The training field is path-state-conditioned. For the linear conditional path,

xs=(1−s)x0+sx1,u∗=x1−x0,x_s=(1-s)x_0+sx_1,\qquad u^{\ast}=x_1-x_0,

the ideal unweighted and unclipped population field is

v∗(x,s,C)=E[X1−X0∣Xs=x,s,C].v^{\ast}(x,s,C)=\mathbb{E}[X_1-X_0\mid X_s=x,s,C].

It depends on the current interpolated state, flow time, and context; the target displacement is therefore not a fixed history-only residual. Our implementation uses bounded endpoints and velocity targets, so we describe it as a bounded CFM objective rather than claiming the exact theory of the unclipped formulation (Appendix A.2).

The one-evaluation decoder is explicit. At inference, the default map is

x+=c(x0+vθ(x0,0.5∣C)),x^+=c\left(x_0+v_\theta(x_0,0.5\mid C)\right),

where cc clips the updated state. We agree that this is algebraically a residual update. Its distinction lies in path-state-conditioned training and in applying the learned update to each source member. We describe it as a one-evaluation decoder, not an exact ODE solution or a second-order midpoint integrator (Section 4.4; Appendix A.4).

Prior scale sets the initial radius, not the final distribution. The prior-derived σs\sigma_s determines the amplitude of the initial graph-correlated perturbation. Each source member is then processed by the sample-state-conditioned field, so the final ensemble depends on both initialization and the learned member-specific update (Eqs. 3 and 8).

Similar CRPS does not imply a constant field. CRPS evaluates coordinate-wise marginal distributions and is relatively insensitive to how nodes and forecast steps are coupled. In our fixed-checkpoint diagnostic, four-evaluation RK4 improves GraphVS by 15.70%–20.75%, temporal variogram score by 54.86%–66.72%, and regional interval score by 5.29%–6.29% over one evaluation. This demonstrates that additional evaluations change measurable aspects of joint scenario organization, at approximately 3.94× the latency; the one-evaluation decoder remains the default (Appendix E.3).

The complete training scheme outperforms Score16. After twenty epochs, STGraphFM improves full-test CRPS by 2.50%, Energy Score by 2.53%, and GraphVS by 20.99% (Section 5.3; Table 3; Appendix C.2).

AuthorsOpenReview ↗

Part 3 of the Response (3/3)

W3. Does source noise simply pass through unchanged?

Thank you for identifying the possibility that the identity path could dominate the generated samples. We address this by distinguishing variation inherited from the source from additional corrections learned by the velocity field, using complementary retraining and fixed-weight interventions.

The identity path is not the learned update. The decoder has the form

x1=c(x0+vθ(x0,0.5∣C)),x_1=c\left(x_0+v_\theta(x_0,0.5\mid C)\right),

so the source contributes an identity component. The key question is whether the velocity field learns meaningful, member-specific corrections beyond this inherited variation.

Shared retraining verifies sample-state dependence. The Shared control retains the source identity path, model capacity, graph, conditioning, and training objective, but replaces the current random-state input with the predictive background. This prevents the field from adapting its correction to individual source members. Compared with this control, STGraphFM improves calibrated CRPS by 3.35% and GraphVS by 21.48%; all six predefined trajectory-functional CRPS scores also improve (Appendix C.3; Table 12).

A fixed-weight intervention isolates learned corrections. To determine whether these gains come from the learned update rather than simply the source’s initial spread, we further compare native member-specific updates with shared-state and common-mean-velocity controls, using identical source draws and conditions. After removing the source identity contribution, native updates retain an across-member RMS dispersion of 0.526 ± 0.039, whereas both controls exhibit only floating-point residual dispersion (Appendix C.3; Figures 4–5).

These corrections improve joint forecasts. At identical native empirical marginals, native updates improve GraphVS by 8.14%, temporal variogram score by 13.09%, and regional interval score by 8.07%. Thus, the learned field contributes member-specific changes with measurable predictive value beyond the variation already present in the source (Figure 4; Appendix C.3).

The result concerns complete updates. These controls evaluate the full sample-state-conditioned mapping. Appendix E.1 separately examines graph-attention routing, so the demonstrated benefit is not attributed to dynamic routing probabilities alone.

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

This paper considers the spatiotemporal forecasting problem and proposes a flow-matching approach named STGraphFM that combines a predictive anchor, structured alternatives, and sample-dependent transport. Experiments show competitive results compard with other methods.

Critical strengths

The paper identifies the problem that identical node-wise event probabilities can imply different probabilities of simultaneous exceedance.

Critical weaknesses
  1. The positioning of the work over existing literature is a bit unclear. The related work section is not very well structured, as it is just a list of existing papers without much discussion. Also, some parts are hard to understand: 'adaptive connectivity support point forecasting'; 'STID and GraphSparseNet address computation'; 'estimate distributions or quantiles'; 'generate graph futures'.
  2. The motivation of the proposed approach is not well justified in the introduction. It can be better explained.
  3. The use of the clip operation is not justified.
  4. The use of renorm is not justified.
Questions for authors
  1. The paper only compares with TSFlow, which is designed for univariate time series forecasting. Are there any existing flow matching approaches for spatiotemporal forecasting?
  2. The training forces the perturbed source samples to transport to the same target; will this cause the potential problem of loss of variability in the generated samples?
6 comments
AuthorsOpenReview ↗

Part 1 of the Response (1/3)

W1 / W2 / Q1. Related work, motivation, and flow-based forecasting

Thank you for pointing out that the original Related Work and Introduction did not make the task distinction sufficiently clear. We have reorganized the discussion around the prediction task and generated output, and added forecasting adaptations of GiFlow and FlowCast under the corresponding STGraphFM task configurations.

The task is joint graph forecasting. Section 2 now distinguishes time-series forecasting, spatiotemporal imputation, grid-based nowcasting, and probabilistic forecasting of graph-indexed future fields.

Related methods address different settings. TSFlow studies probabilistic time-series forecasting; FlowCast applies conditional flow matching to radar precipitation nowcasting; and GiFlow studies graph-informed spatiotemporal imputation. These works establish flow matching in related temporal and spatiotemporal domains, but do not formulate the same task as STGraphFM: generating the complete, unobserved future node–horizon field on a graph, conditioned on historical observations and known future context. Our contribution is this graph-forecasting formulation and its coordinated predictive source and graph–time generation design, not flow matching itself.

Additional forecasting comparisons. We adapted the released GiFlow implementation and FlowCast to spatiotemporal forecasting under the corresponding STGraphFM experimental configurations.

DatasetMethodMAE ↓RMSE ↓
KnowAirV2STGraphFM ensemble mean13.32422.602
KnowAirV2GiFlow-F (our adaptation)24.23840.156
WeatherBench T2MSTGraphFM central forecast1.2722.164
WeatherBench T2MFlowCast-WB (our adaptation)1.3892.331

For GiFlow-F, we adapted the released GiFlow implementation to future forecasting under the same experimental configuration as STGraphFM (Appendix F.1). For FlowCast-WB, we retained its released autoencoder and conditional-flow architecture and completed the AE60/CFM16 adaptation budget. It uses the same WeatherBench fields, year splits, eight-step history, forty-step horizon, and latitude-weighted evaluation as STGraphFM. On all 721 test windows, STGraphFM achieves 8.37% lower MAE and 7.19% lower RMSE (Appendix F.3).

Marginal forecasts do not determine joint events. Accurate node-wise predictions do not specify whether multiple locations exceed a threshold together, how regional peaks evolve, or how long an event persists. STGraphFM addresses this by combining a predictive background, graph-correlated initial alternatives, and updates conditioned on each complete sample state. The source factorial and Shared/fixed-weight controls connect these roles to forecast placement and joint predictive utility (Sections 1, 4, and 5.3).

The revision clarifies the task distinction. The revised Related Work compares methods by their prediction targets and generated outputs, rather than presenting them as a list of architectures.

AuthorsOpenReview ↗

Part 2 of the Response (2/3)

W3 / W4. Why are clipping and Renorm needed?

Thank you for asking for the rationale behind these implementation choices. We have clarified the numerical role of clipping, the statistical role of Renorm, and how each relates to the source–target generation process.

Clipping protects the regression from rare extremes. Even in normalized coordinates, a small number of extreme source or target values can produce disproportionately large squared regression errors. We therefore bound source and target states, as well as predicted and supervised velocities, to [−8,8][-8,8]. At inference, the raw velocity is applied before the updated state is projected to the same interval. Clipping is a safeguard for extreme states and updates, rather than a mechanism intended to alter ordinary predictions.

The implemented objective is a bounded CFM surrogate. For a linear path, the exact path derivative is x1−x0x_1-x_0. Clipping the supervised velocity bounds this target for extreme source–target differences, so it no longer equals the exact path derivative when the bound is active. We therefore describe and evaluate the bounded regression actually used, rather than assuming the exact transport properties of unclipped CFM (Appendix A.2). In sampled training probes, velocity clipping was rarely activated, consistent with its role as a guardrail rather than the main source of model performance (Appendix C.4).

The ablation supports this safeguard. Removing clipping increases mean CRPS by 0.43% and regional interval score by 1.85%. Removing both clipping and Renorm increases these scores by 2.18% and 18.88%. The modest marginal effect is consistent with clipping mainly protecting against rare extremes (Appendix C.4).

Renorm separates graph correlation from source amplitude. Graph smoothing mixes neighboring noise values. This introduces spatial dependence, but also changes the overall spread of the perturbation. Without scale restoration, changing the graph-mixing ratio would unintentionally change both correlation and source radius. For a linear graph operator SS and IID noise with identity covariance,

Cov⁡(Sϵ)=SS⊤.\operatorname{Cov}(S\epsilon)=SS^\top.

Renorm rescales each example’s smoothed or mixed perturbation to the original noise’s empirical standard deviation over all node–horizon–channel entries, without subtracting its mean. It is a global scale restoration, not coordinate-wise whitening (Appendix A.2).

Preserving source scale benefits regional uncertainty. Removing Renorm increases mean CRPS by 1.07% and regional interval score by 15.13%; regional coverage also falls from 58.57% to 41.97%. Some variogram scores improve when Renorm is removed, so we describe it as a source-scale control that benefits marginal and regional uncertainty, not as a uniformly beneficial operation (Appendix C.4).

AuthorsOpenReview ↗

Part 3 of the Response (3/3)

Q2. Does one observed future per history cause sample collapse?

Thank you for raising the central difficulty of probabilistic forecasting with only one observed future per historical context. We clarify why the CFM field is not simply a history-only residual and connect this mechanism to controlled tests of member-specific updates.

Single futures limit direct identification. A forecasting dataset typically provides one realized future for each historical window, rather than repeated futures under exactly the same condition. This limits direct identification of the full conditional multimodal distribution.

The CFM field reads each path state. For ideal, unweighted conditional flow matching with a linear path,

Xs=(1−s)X0+sY,v∗(x,s,C)=E[Y−X0∣Xs=x,s,C].X_s=(1-s)X_0+sY,\qquad v^{\ast}(x,s,C)=\mathbb{E}[Y-X_0\mid X_s=x,s,C].

Although a training history has one observed future, different source draws occupy different interpolated states. The learned field therefore predicts updates conditioned on each current state, flow time, and context, rather than assigning one fixed residual to every member with the same history. Our bounded and weighted implementation modifies this idealized setting, so the formula explains the standard CFM mechanism rather than guaranteeing exact recovery of the conditional distribution.

Shared retraining verifies the value of state input. Replacing each member’s current-state input with the predictive background removes the field’s ability to adapt its correction to individual source members. Relative to Shared, STGraphFM reduces calibrated CRPS by 3.35% and GraphVS by 21.48%; all six predefined trajectory-functional CRPS scores also favor STGraphFM (Appendix C.3).

A fixed-weight intervention isolates learned corrections. The retraining comparison establishes the predictive value of sample-state conditioning. To distinguish learned corrections from variation already present in the source, we further compare native updates with shared-state and common-mean-velocity controls using identical source draws and conditions. After subtracting the source identity path, native updates retain an RMS dispersion of 0.526 ± 0.039, whereas both controls exhibit only numerical residual dispersion. At identical empirical marginals, native updates improve GraphVS by 8.14%, temporal variogram score by 13.09%, and regional interval score by 8.07% (Appendix C.3; Figure 4).

Together, these results show that the model does not merely preserve initial source variation: its learned, sample-dependent corrections contribute predictive structure to complete future scenarios.

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

This paper studies probabilistic forecasting over spatially connected time series. It proposes STGraphFM, which combines a node-wise MLP predictor, graph-correlated random perturbations, and a graph–time attention network trained through flow matching. The MLP provides an initial forecast, while the attention network transforms perturbed forecasts into complete samples covering all nodes and forecast horizons. The same network also produces a deterministic forecast from the unperturbed initial prediction.

The evaluation covers traffic, air quality, and gridded temperature forecasting. In addition to forecasting accuracy and sampling cost, the paper evaluates regional events and compares sample dependence while keeping the empirical marginal predictions fixed. Same-backbone controls examine the contributions of the training objective and sample-dependent updates.

Critical strengths
  1. The common-marginal comparisons in Appendix E hold every coordinate’s empirical predictive values fixed and change only their pairing across samples. STGraphFM retains advantages in joint scores and regional event prediction under this comparison. The transformed-trajectory evaluations also examine changes, peaks, and exceedance duration. These experiments directly address the paper’s central motivation: useful joint forecasts require more than accurate predictions at individual locations.

  2. Results span several domains, and sampling is fast: generating 16 samples takes approximately 195 ms versus 1,483 ms for CSDI on the same device (Table 4).

Critical weaknesses
  1. Regional uncertainty remains unreliable despite improvements over controls. Nominal 80% regional intervals achieve coverage as low as 53.6% (Table 24), and calibrated PM2.5 samples contain 3.17% and 6.53% negative values (Table 23). These acknowledged limitations weaken the evidence for regional risk applications.

  2. The additional benefit of the proposed source design over simpler alternatives remains unclear. Although component ablations support using graph information, output scaling outperforms source-temperature adjustment at matched validation coverage (Table 25). Historical-rank pairing also achieves better graph-variogram and regional-event scores in Table 26, although STGraphFM achieves a better Energy Score. These mixed results leave the specific advantage of the source design insufficiently established.

Questions for authors
  1. Do validation curves indicate convergence in the Score16 comparison? Does the advantage persist with longer training and consistent training/test output transformations?

  2. What explains the regional undercoverage? Would nonnegative output processing followed by validation-only recalibration preserve the joint-score and event-prediction gains?

  3. Given Tables 25 and 26, what specific benefit does the proposed source design provide beyond simple output scaling or historical-rank pairing? Please clarify which results establish this benefit and under what conditions.

5 comments
AuthorsOpenReview ↗

Part 1 of the Response (1/3)

W1 / Q2. Negative PM2.5 values and regional uncertainty

Thank you for highlighting physical validity and regional reliability. We revised the output-processing description and added a regional-query calibration analysis. We address the physical support of PM2.5 samples, whether joint-prediction gains survive correction, and why node- and region-level coverage require separate treatment.

PM2.5 samples satisfy nonnegative support. We apply nonnegative correction in physical space, followed by node-scale recalibration on the calibration split.

Joint gains persist after marginal correction. Under exact common marginals, STGraphFM improves over Score16 by 5.88%–11.56% in GraphVS, 5.75%–7.68% in regional interval score, and 1.83%–2.49% in regional-event Brier. Against Shared, the corresponding gains are 9.11%–14.92%, 7.01%–8.67%, and 2.85%–3.15%. Since coordinate-wise empirical distributions are fixed, these comparisons isolate how values are assembled into complete futures (Table 4; Appendix D.1).

Regional coverage depends on joint uncertainty. Even when each node has a calibrated interval, aggregation can produce a different uncertainty because cross-node correlations affect the spread of regional outcomes:

Var⁡(∑iwiYi)=∑iwi2Var⁡(Yi)+2∑i<jwiwjCov⁡(Yi,Yj).\operatorname{Var}\left(\sum_i w_iY_i\right) = \sum_i w_i^2\operatorname{Var}(Y_i) + 2\sum_{i < j}w_iw_j\operatorname{Cov}(Y_i,Y_j).

After nonnegative processing and node recalibration, mean regional coverage is 65.06%. Fitting region–lead spread factors on validation queries raises it to 80.64% and reduces regional interval score from 59.492 to 54.919. This adjusts regional query intervals while preserving the node-level ensemble; it is a query-specific calibration rather than simultaneous calibration of every possible joint query (Appendix D.3).

The appendix is more focused. We consolidated repeated protocols and overlapping tables, retaining the evidence needed to assess the method while making the supplementary presentation more concise.

Q2. Do nonnegative processing and recalibration preserve joint gains?

Yes. The common-marginal comparisons show that joint-score gains persist after nonnegative processing and node recalibration. Coordinate-wise empirical distributions remain identical; only cross-coordinate member pairing changes. Regional query calibration is reported separately because it adjusts regional intervals, not the node-level sample ensemble.

AuthorsOpenReview ↗

Part 2 of the Response (2/3)

W2 / Q3. What do the source components contribute beyond scaling and historical ranks?

Thank you for asking us to clarify the specific role of the source design. We distinguish three operations that act at different stages: placing the source around a useful forecast, organizing its initial spatial variation, and controlling scenario strength before transport.

Predictive centering reduces the generation burden. Historical observations already reveal node-specific levels, recent-state inertia, and temporal rhythms. Rather than asking the generator to recover this predictable background from unstructured noise, we center the source on a learned future reference, allowing transport to focus on the remaining deviations. Predictive centering reduces mean CRPS by 12.90% with graph noise and 17.92% with IID noise.

Graph correlation organizes initial uncertainty. Independent noise perturbs neighboring nodes separately, whereas graph smoothing introduces locally coherent variations while retaining an independent component. Because smoothing also changes overall amplitude, Renorm restores the source scale. At identical predictive-center marginals, graph initialization improves GraphVS by 10.41% and regional interval score by 11.28% over IID initialization (Table 2; Appendix C.1).

Historical-rank isolates dependence reconstruction. Historical-rank transfers coordinate-wise rank patterns from complete training futures to STGraphFM’s generated values. It preserves every coordinate’s empirical marginal, so it is a dependence control rather than an independent forecaster; CRPS and point forecasts are unchanged. In the full-test comparison, historical pairing is competitive on GraphVS and regional Brier, while STGraphFM has lower Energy Score (Appendix E.2).

Conditional templates test the value of learned pairing. On the same frozen STGraphFM marginals, native pairing improves over IID pairing by 6.21%–9.28% in regional-mean peak CRPS, 5.64%–5.88% in joint-exceedance Brier, 9.86%–10.49% in longest-duration CRPS, and 7.46%–8.14% in exceedance-fraction CRPS. Conditional historical templates remain competitive, including on several outcomes in high-volatility and spatially heterogeneous groups. We therefore establish the value of structured pairing over independence, rather than claiming universal superiority over empirical historical dependence (Appendix E.2).

Temperature controls pre-transport scenarios. Source temperature scales the structured perturbation before transport, changing the state entering the nonlinear velocity field. Output scaling acts after decoding and preserves within-coordinate member ranks. Figure 3 illustrates the functional effect: with the central forecast fixed, changing temperature changes the event-risk map; increasing τ\tau from 0.3 to 1.0 yields six additional true positives with no additional false positives in the illustrated query. A separate fixed-model comparison confirms that temperature changes 60.65%/66.38% of member-rank entries, unlike positive affine output scaling (Figure 3; Appendix E.4).

Q3. How is the source more than output scaling or historical pairing?

The components act at different stages. Predictive centering places alternatives around a useful future reference; graph-correlated perturbations organize their initial dependence; temperature changes the pre-transport state and can alter generated member assignments. The source factorial isolates the first two roles. Historical-rank remains a fixed-marginal dependence control, not a competing full forecasting model.

AuthorsOpenReview ↗

Part 3 of the Response (3/3)

Q1. Does the Score16 gap persist with longer training?

Thank you for asking whether the original comparison was affected by a short training schedule or inconsistent output transformations. We extended the comparison and report its validation behavior together with raw and calibrated test results.

The advantage persists after extended training. Both methods are trained for 20 epochs and evaluated on all 14,003 KnowAirV2 test windows, using the same backbone, initialization rules, data order, and optimizer settings. Score16 uses the complete 16-member endpoint-scoring map, including the transported central forecast and its gradients.

Validation performance reaches a plateau. The selected checkpoints occur around epochs 11–12, and neither method improves its selected validation score during the final five epochs.

Metric ↓STGraphFMScore16Improvement
Raw CRPS10.15510.4162.50%
Raw Energy Score13.80314.1612.53%
Raw GraphVS2.8893.65620.99%
Calibrated CRPS10.20210.4702.56%
Calibrated GraphVS3.6404.32315.81%

The output comparison is aligned. Both methods use the complete output map in training and evaluation. The advantage therefore persists under extended training and consistent output processing (Section 5.3; Appendix C.2).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.