acceptodds
Under review as a conference paper at ICLR 2027

CSBOO: A Coherent Schrodinger-Bridge-Inspired Orbit Operator for Reliability-Aware Multi-Period Spatio-Temporal Forecasting

Abstract

Spatio-temporal forecasting often combines recent, daily, and weekly histories, but their reliability changes during incidents and holidays and across forecast horizons; unconstrained mixtures also make orbit weights ambiguous. We present the Coherent Schrödinger-Bridge-Inspired Orbit Operator (CSBOO), which replaces independent reference fusion with sequential inference over the responsibilities of three semantically anchored predictive orbits. A temporally causal backbone constructs horizon states through input-dependent directed spatial propagation, retarded attention, and known calendar phase. The retarded weights also define level, slope, and cross-period mismatch as reliability evidence. A Markov transition propagates responsibility, and a mismatch-informed Feynman–Kac potential corrects it for each sample, node, and horizon. Bounded candidate deformations and target-conditioned calibration give the normalized marginals predictive semantics; their convex barycenter yields the forecast. Across ten tasks, CSBOO achieves the best result in 18 of 20 MAE/RMSE comparisons, including 7.9% lower MAE and 4.0% lower RMSE on PeMS08 than the strongest baseline, with consistent degradation across all reported ablations.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-Temporal Forecasting, Time Series Forecasting, Multi-Period Forecasting, Dynamical Systems, Markov Dynamics, Reliability-Aware Forecasting, Feynman–Kac Inference

Reviews

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

CSBOO forecasts each node and horizon as a convex mix of three anchored candidates: the latest value, the value one day earlier and the value one week earlier, each with a bounded deformation. An HMM-style forward recursion across horizons sets the weights, using a learned transition matrix and mismatch-based evidence terms. A training-time KL loss calibrates the weights. The paper reports the best result in 18 of 20 MAE/RMSE comparisons on 10 tasks.

Critical strengths

Making the periodic references explicit candidates with horizon-coherent weights is a sensible, interpretable bias. On PeMS08, the released record (12.785 / 22.397) matches Table 1, and it beats STAEformer, PDFormer and STD-MAE as reported in their own papers.

Critical weaknesses

The idea is reasonable, but I could not verify the main results from the released material, and the comparison does not isolate the method.

  1. The released records and code do not support the reported results.
  • results/main_table_results.md has records only for PeMS03, PeMS04 and PeMS08. For the other 7 tasks, "matching server JSON records were not found". The PeMS03 record (14.93 vs 14.75) and the PeMS04 record (18.29/30.88 vs 17.98/29.88) disagree with Table 1. With the recorded values, CSBOO is not the best on PeMS03 MAE or on either PeMS04 metric. All records come from one seed, and the paper reports no variance.
  • The experiment-85 README says the checkpoints were trained with MAE only, that is, without the calibration loss given in Table 7. On PeMS08 the calibrated run scores 12.967/22.572, worse than the uncalibrated one. This is the reverse of Table 3.
  • The diagnostic records disagree with Table 4: the daily-slope ρ is +0.053 (Table 4: -0.21), and the mass variation without Markov transport is 0.0119 (Table 4: 0.028). The authors' own report marks C1 and C2 as not supported.
  • In the code (CausalOrbitBackbone.py, CausalSchrodingerBridge.py), the mismatch terms are history means, not horizon-specific (Eq. 6), and the deformations are unbounded linear heads (Eq. 8).
  • Some baseline entries differ from the published values, in both directions. For example, STAEformer on PeMS08 is 13.87/24.01 here vs 13.46/23.25 in its paper.
  1. The softmax-fusion ablation already beats every baseline on PeMS08. So the gain is not shown to come from path inference, contrary to L453-454. CSBOO uses lagged inputs that reach back a week, while STAEformer and PDFormer see 12 steps. The paper compares no long-history methods, and STD-MAE reports better PeMS03/04/07 results than Table 1. Proposition 1 is the Gibbs variational principle with a standard HMM forward recursion, not a Schrödinger bridge. Holidays and incidents are the stated motivation but are never evaluated.
Questions for authors
  1. Please provide records with at least three seeds for all ten tasks, and explain the mismatches with Tables 1, 3 and 4. Did the Table 1 models use the calibration loss? Where does each baseline number come from?
  2. In the evaluated model, are the mismatch terms horizon-dependent, and are the deformations bounded?
  3. Please add a backbone-only variant and a comparison with STD-MAE and STEP.
  4. Please report results on holiday and incident subsets. For example, Presidents' Day 2018 falls in the PeMS04 test split.
Reviewer_6GYROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CSBOO for multi-period spatio-temporal forecasting. Recent, daily, and weekly references define three semantically anchored candidates, whose predictive responsibilities are inferred through Markov transport and Feynman-Kac evidence correction. The formal method includes horizon-dependent level, slope, and cross-period mismatch, scale-bounded candidate deformations, target-conditioned responsibility calibration, and anchor regularization. The paper establishes a connection between a prefix-path free-energy objective and the recursive inference procedure. Experiments cover six highway forecasting tasks and four urban inflow/outflow tasks. Tables 1–2 report the best results in 18 of 20 MAE/RMSE comparisons, with approximately 7.9% and 4.0% improvements on PeMS08.

Critical strengths

S1. This paper tackles a specific and relevant problem. It separates the availability of periodic references from their reliability, and considers level mismatch, slope mismatch, and daily–weekly disagreement. Combined with horizon-dependent historical attention, these signals offer an explicit way to judge when periodic references may mislead and make their use easier to inspect.

S2. The inference structure is compact and mathematically well characterized. The three-state Marko-Feynma-Kac recursion marginalizes state paths exactly without enumerating all sequences. Appendices C–G specify the conditions for normalization, masking limits, conditional monotonicity, and responsibility calibration. The evaluation covers highway flow, highway speed, and urban-grid flow; the 18 best results out of 20 comparisons can be checked against the main tables.

Critical weaknesses

W1. The formal method and supplied implementation differ in several central components, including horizon-dependent mismatch, bounded candidate deformations, and the training objective. It is therefore unclear whether the reported results evaluate the full method described in the paper. These differences need to be reconciled before the gains can be attributed to the proposed mechanisms.

W2. Some supplied experimental records do not match the paper’s tables. In particular, the daily slope mismatch-mass correlation and the full/no-Markov mass variation comparison differ in direction from Table 4. The PeMS03/04 main results also differ from the available run records. Together with the limited seed coverage, these inconsistencies weaken the empirical support for the conclusions.

W3. The paper contains only one figure, and its small labels make the overview difficult to read. A clearer overview and additional visualizations of responsibility trajectories or diagnostic behavior would improve the presentation.

Questions for authors

Q1. Why share horizon-dependent attention weights between state construction and mismatch assessment, and how does this coupling beat horizon-independent mismatch statistics?

Q2. With close semantic anchors or overlapping candidate intervals, how do bounded deformations and responsibility calibration yield meaningful orbit responsibilities, and what evidence separates better interpretation from mere alignment with the calibration target?

Q3. Why do the PeMS03/04 main-table results differ from official_test_seed3031.json, and what are the run records for the final model's main and ablation results?

Q4. Given DAY_STEPS=288 and WEEK_STEPS=2016 in the data-packing helper but 30-/60-minute intervals in CHIBike and T-Drive, how were daily and weekly lags constructed for these urban tasks? Please provide the preprocessing configurations and state whether the baselines used the same splits and periodic information.

Reviewer_qCV9OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CSBOO, a spatio-temporal forecasting model that adaptively combines recent, daily, and weekly historical references. It constructs three anchored candidate predictions and infers their weights sequentially across forecast horizons using Markov transitions and mismatch-informed Feynman–Kac corrections. A responsibility calibration loss encourages the weights to align with candidates that better explain the observed targets. Experiments on traffic flow, speed, and urban inflow/outflow tasks report improvements over existing baselines, supported by component ablations and diagnostics of the learned weights.

Critical strengths

(1) The paper addresses a practical issue in spatio-temporal forecasting. Simply combining recent, daily, and weekly histories may not be enough, since their reliability can change across situations, nodes, and forecast horizons. Taking these differences into account is a reasonable motivation, especially when incidents or holidays make historical patterns less useful.

(2) The model also makes a consistent effort to give the mixture weights meaningful interpretations. It keeps the candidate predictions close to their temporal anchors, calibrates the weights toward candidates that better match the target, and connects the weights across forecast horizons.

(3) The experiments cover different forecasting tasks, including traffic flow, speed, and urban inflow/outflow. The paper reports the best results in 18 of 20 MAE/RMSE comparisons. The ablation results and additional diagnostics also gives useful evidence that the individual components contribute to the results and that the weights behave as intended.

Critical weaknesses

(1) The novelty of the core method compared with existing sequential mixture approaches is not very clear. The method propagates weights from the previous forecast horizon through a Markov transition and then corrects them using current reliability evidence. This looks similar in form to existing sequential filtering methods. More explanation is needed on what is different from these approaches and what additional benefit the Schrödinger-bridge-inspired perspective brings. The mathematical formulation helps explain the computation, but some results on normalization, additive softmax invariance, and KL minimization are relatively basic properties. It is still unclear how these results, together with candidate construction and responsibility calibration, make a distinct methodological contribution.

(2) The experiments do not directly evaluate forecasting performance when periodic references become unreliable, although this is a main motivation of the paper. The paper discusses incidents and holidays as situations where daily and weekly patterns may fail. There is a mismatch ablation and diagnostics showing that larger mismatch is associated with lower weights, but forecasting performance is not separately evaluated during holidays, incidents, or abrupt pattern changes. Good average performance on the full test set is not enough to tell whether the model handles these situations well. A separate evaluation on normal periods and periods with disrupted periodic patterns would make the central claim more convincing.

(3) Statistical evidence for the small performance differences and details of the comparison conditions are limited. For example, MAE improves from 18.06 to 17.98 on PeMS04 and from 17.16 to 17.11 on T-Drive inflow. However, the paper does not report means and standard deviations over repeated runs or significance analyses, so it is difficult to know how stable these improvements are.

(4) The model includes several components, such as directed spatial attention, horizon-specific attention and spatial processing, three candidate predictions, and sequential weight inference, but its computational efficiency is not evaluated. A basic efficiency comparison, for example in inference time or memory usage, would help assess the cost of the accuracy gains, especially since some improvements are small.

Questions for authors

(1) Could you provide a separate comparison of forecasting performance during normal periods and periods when daily or weekly patterns become unreliable, such as holidays, incidents, or abrupt pattern changes? This would help clarify whether the proposed reliability mechanism improves predictions in the situations motivating the paper. (2) Could you report means and standard deviations across repeated runs or a significance analysis for the smaller improvements, such as those on PeMS04 and T-Drive inflow? This would help assess whether these gains are consistent across runs.

(3) Could you provide a basic efficiency comparison with the baselines, for example in inference time or memory usage? Given the model’s multiple components, this would help assess the computational cost of the accuracy gains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.