acceptodds
Withdrawn from ICLR 2027

FCDC: Forecast-Conditioned Delay Correction for Traffic Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Traffic changes can be propagated between connected locations with a delay, so observations available at the forecasting origin may remain useful during the prediction interval. When these observations are used for correction, the already-predicted trajectory must be taken into account. In Forecast-Conditioned Delay Correction (FCDC), the complete multistep forecast is supplied directly to a compact signed readout, alongside local history and delayed evidence. The base forecast is formed by adding a global backbone prediction and a shared linear own-history prediction. Historical values are routed into an extended arrival window through learned edge-lag distributions; available lags are combined by a history-dependent gate, and messages are normalized over valid sources. All branches are trained jointly, without a cache of realized forecast errors. With local evidence fixed, correction signs can be reversed by changing the forecast input; such reversals are excluded by the corresponding unconditioned readout. FCDC and seven controlled variants are evaluated with five seeds on four PEMS datasets. FCDC's mean MAE is 0.62–2.29% lower than the best recorded external baseline on each dataset. Its average relative MAE reductions against variants without forecast input and without delay correction are 0.63% and 1.38%, respectively, with lower overall MAE in every matched-seed comparison.

Primary area:
learning on time series and dynamical systems
Keywords:
Forecast-conditioned correction, Delay-aware modeling, Spatiotemporal learning, Traffic forecasting

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_1QdrOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FCDC, a traffic forecasting model equipped with a delay-correction mechanism based on lags between nodes. FCDC determines the magnitude and the sign of a correction to the base forecast by taking into account the local history, the delayed evidence, and the base forecast itself. FCDC outperforms baseline methods on four real-world datasets.

Critical strengths
  • S1: The idea of using lags to correct a forecast is interesting.
  • S2: The effectiveness is confirmed on real-world datasets.
  • S3: The authors release code for reproducing the experiments.
Critical weaknesses
  • W1: The design rationale of the method is not sufficiently explained in the manuscript. In particular, I have the following questions.
    • W1-1:Why is it necessary to separate the global prediction and the self (own-history) prediction to obtain the base forecast?
    • W1-2: Why does the global prediction not use standard spatial attention and temporal attention?
    • W1-3: Why do lags need to be used to correct the base forecast? It is also possible to build a single forecasting model that takes all lag information as input. The motivation of this work would be clearer if the authors explained why they consider such an approach insufficient.
  • W2: The performance gains over the baselines and the ablation variants are not necessarily large. In particular, even the "Parameter-matched MLP", in which the proposed delay branch is entirely replaced with an MLP, is only about 1% worse than the full model (Table 2), although its mean performance is lower on all datasets. Moreover, the ablation variants would in most cases rank second best in Table 1. These results suggest that most of the advantage of the proposed method over the baselines in Table 1 is attributable to the base forecast, and that the contribution of the delay branch, which is the core of the proposal, may be small.
Questions for authors
  • Q1: Please explain the experiment in Figure 2. What does the horizontal axis "condition-only shift" represent?
  • Q2: What is the purpose of gating δ\delta with η\eta? It seems that only δ\delta could lean the scale of the correction. Please explain the intention behind introducing the gate.
  • Q3: The magnitude of the lag may vary with the time of day and with traffic conditions. Currently, a single π\pi is shared across all times. Consequently, the aggregation weights in Eqs. (5) and (6) are time-invariant. If the lag varies over time, ν\nu and ω\omega may be computed from incorrect lags. Is accounting for the time dependence of lags outside the scope of this work?
Reviewer_cSjdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors propose FCDC to correct traffic flow predictions with time-delayed neighbor evidence from historical observations. The base prediction is obtained by adding the predictions of the bidirectional cross-attention backbone and the shared own-history linear map. Neighbors' historical observations are routed through learnable edge-level lag distributions and history-dependent dynamic lag gating to arrival times that extend into the prediction interval, and the resulting messages are jointly normalized over the valid sources. The signed correction readout simultaneously receives the local history, delayed messages, validity masks, and the complete base prediction. Its correction is added back to the base prediction after sigmoid gating, and all modules are trained jointly. The authors compare FCDC over five seeds against eleven external baselines and seven controlled variants on four PEMS datasets.

Critical strengths

The seven controlled variants share initialization and match the number of parameters, including a local MLP with an equal number of parameters and two lag controls fixed to uniform weights. Running five seeds per configuration, FCDC outperformed each variant in all 20 matched dataset and seed combinations, and in 1,190 out of 1,200 per-horizon comparisons. The credibility of the internal comparison is clearly higher than that of typical traffic prediction work.

Critical weaknesses

W1. The No delay correction variant, which only keeps the new backbone network and the own-history path, has MAE of 18.14, 18.96, and 13.32 on PEMS04, PEMS07, and PEMS08, which is already lower than all external baselines in Table 1. However, the correction branch only brings a 1.38% reduction on average compared to this variant. The backbone network is neither introduced as a contribution nor ablated separately, so the external comparison cannot prove that the delay correction conditioned on the forecast is effective by itself.

W2. The MAE of the internal base prediction of the jointly trained model in Table 15 is 20.21, 29.02, 25.80, and 31.72. This is much worse than that of the separately trained no-correction variant, indicating that y^base\hat{y}^{base} is an intermediate quantity co-adapted with the correction branch, rather than a prediction being corrected. Permuting delayed messages across sensors in Table 17 only increases MAE by 0.02 to 0.09. The edge-level priors in Table 19 are close to uniform, and 71% to 94% of the dynamic gate mass falls on lag 1. Thus, neither the narrative of correcting an existing prediction nor the time-delay propagation explanation is supported, and the gain more likely comes simply from an additional readout stage.

Questions for authors

Q1. After joint training, the MAE of the internal base on PEMS08 is 31.72, while that of the separately trained no-correction variant is only 13.32. What is the reason? How does the result of FCDC change if y^base\hat{y}^{base} is detached before being fed into the readout, or if an auxiliary prediction loss is added to it?

Q2. Permuting delayed messages hardly affects the error, and the priors are also close to uniform. So which part of the delay branch does the gain relative to the Parameter-matched MLP come from? What happens if you randomly shuffle the graph during training?

Sug1. Implement a ResCAL-style correction on the same backbone, and transplant the forecast-conditioned readout to existing backbones such as STAEformer, to separate the contributions of the backbone and the correction.

Sug2. Supplement the results on speed datasets such as METR-LA and PEMS-BAY, as well as under larger values of K, to examine the effect of delay coverage.

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

FCDC combines a base traffic forecast with a learned correction conditioned on the complete forecast, local history, and delayed neighbor observations. Historical observations are routed through learned edge-lag distributions and a history-dependent gate. All components are trained jointly. Experiments on four PEMS datasets report consistent MAE improvements across five seeds, supported by seven controlled variants and input-sensitivity analyses.

Critical strengths

1.The evaluation includes useful controls: a parameter-matched MLP, hidden-state conditioning, and uniform lag weights. Five-seed results provide evidence that the MAE gains over these variants are consistent.

2.The paper clearly specifies the information boundary and uses only observations available at the forecasting origin. The diagnostics also report limitations and unfavorable results rather than presenting uniformly positive evidence.

Critical weaknesses

1.The technical contribution is limited. The core design feeds the base prediction into a residual correction network. This is reasonable, but the paper offers limited insight into why it is preferable to conditioning on the underlying representation. Proposition 4 establishes a sign restriction for a particular architecture where hidden states influence correction only through a positive gate. It does not establish a general limitation of hidden-state-conditioned correction. The empirical comparison favors forecast input, but the theoretical argument does not explain that advantage.

2.The contribution of delay modeling remains unclear. In Table 17, permuting delayed messages across sensors changes MAE only slightly. Table 19 further shows that most edge-lag priors are close to uniform, while the dynamic gates strongly favor the shortest lag. These findings leave open whether the gains require adaptive delay routing or mainly come from simple aggregation of recent neighbor observations. Uniform-weight controls are useful, but do not rule out fixed short-lag or globally shared lag weighting.

3.The distinction between forecast correction and ordinary residual prediction is unclear. The base predictor and correction branch are trained jointly, with supervision applied only to their summed output. Nothing explicitly requires the base output to be an accurate standalone forecast or the correction branch to learn its remaining errors. The two branches can instead learn complementary parts of the prediction. The paper therefore needs to clarify what makes this a distinct error-correction mechanism beyond adding a forecast-conditioned residual branch.

Questions for authors

1.What explains the advantage of decoded forecast input over hidden-state input beyond the restricted architecture considered in Proposition 4?

2.How does FCDC compare with fixed one-step lag aggregation or globally shared learned lag weights? How should the small effect of message permutation be interpreted?

3.Does the correction remain effective with a separately trained, frozen backbone or across different forecasting architectures? Can the authors identify cases where gains specifically arise from delayed propagation?

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

Related papers

Discussion (0)

Sign in to comment.