- 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.
- 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.
- 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
- 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?
- In the evaluated model, are the mismatch terms horizon-dependent, and are the deformations bounded?
- Please add a backbone-only variant and a comparison with STD-MAE and STEP.
- Please report results on holiday and incident subsets. For example, Presidents' Day 2018 falls in the PeMS04 test split.