- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
TASCT enhances forecasting backbones with history input adaptation and forecast output calibration. The combination of multiscale temporal corrections and corrections from frequency-grouped input views, preserving historical mean in TSA is motivated and relevant. This and the backbone are jointly trained, followed by static and dynamic gate fitting . LPC then learns contextual residual corrections and corrections using neighboring predicted leads and previous-origin forecasts, while preserving each forecast's temporal mean. Across five backbones, seven datasets, four horizons, and three seeds, TASCT improves MSE in 130/140 settings, with a 4.17% average reduction. Ablations and composition/comparison with LIFT and CCM are also evaluated.
Critical strengths
-
The paper provides reasonable evidence of improvement with matched Native/TSA configurations to support the main empirical claim. We see gains across several architecture families with 130/140 MSE wins and 123/140 MAE wins, and average reductions of 4.17% and 2.34%. It is interesting to note that TASCT can add value to LIFT and CCM enhanced forecasters as well. These experiments establish a useful practical contribution.
-
The ablations reported in the paper help address alternative explanations, like the comparison of TSA with a parameter-matched residual MLP, individual branches, and the fixed 0.5/0.5 combination before gate fitting. Appendix F evaluates the contextual and consistency components, removal of centering, and simple temporal-averaging controls which are meaningful controls. TASCT's main contribution is the specific mean-preserving combination (of input/output adapters, spectral adaptation, and contextual revision) and the evidence supporting it.
Critical weaknesses
-
Some implementation details in the paper are missing. I understand the TSA correction-energy penalty and the separate gate penalty. However, the it is not clear how the final LPC correction is fitted. How many pooled history/forecast features are used, whether the regression is shared across channels, and exactly how the consistency coefficients and final mixing weights are estimated. The chronological calibration blocks also need explanation of how overlapping forecast targets are handled.
-
The benefit of the extra fitting stages is less clear than the benefit of TASCT overall. TSA already achieves a 3.49% average MSE reduction; adding LPC raises the total to 4.17%. This is useful, but Table 9 shows that the extra consistency component contributes little in several settings. The paper reports LPC's 0.38 ms average batch latency and 3.2–25.0K deployment coefficients, but it would also be beneficial to add the fitting time and a breakdown showing what the consistency component adds in accuracy and cost. Similarly, Table 8 compares fixed equal weights with fully fitted gates, but does not separate the benefit of learned static weights from the subsequent input-dependent gate. Adding these may improve the practical case for using every stage.
Questions for authors
-
Could you provide one section or algorithm or table listing each stage's objective, fitted parameters, and regularization? For LPC, please specify the pooled features, ridge-penalty selection, coefficient-fitting procedure, including how the overlapping target intervals are handled.
-
Since Table 8 already evaluates fixed equal weights, how much of the remaining improvement comes from learned static channel-wise weights, and how much comes from the dynamic gate?.
-
Can you summarize the incremental improvement and fitting/inference cost of full LPC over contextual correction alone? To justify its added fitting and rolling-inference steps.
-
In Table 12 δ-Adapter worsens all eight ETT backbone–dataset averages. Please explain this behavior, as there is clearly a difference in how the frozen-backbone is handled in δ-Adapter (frozen) v/s TASCT (joint training).
(nice to have) An inference-cost-matched weighted ensemble using an independently trained forecaster would provide another practical reference point, as LPC is a residual correction model and can be viewed as additive model combination.
I am grateful to the authors for clarifying the notation and definitions. However, there are still some problems concerned with me: