- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Time Fluxer (TiF), an input-dependent FFT global convolution for high-dimensional forecasting. It uses causal zero-padding along time and circular convolution along variables. Four variants are evaluated on seven datasets and four horizons.
Critical strengths
S1: The method is well motivated through a clear distinction between causal temporal closure and circular variable closure, supported by both theoretical derivation and computational considerations.
S2: The evaluation is broad, and substitutions into PatchTST and iTransformer help isolate the proposed operator.
Critical weaknesses
W1: Each temporal fluxer layer applies one input-dependent lag kernel across the entire window. The method however is not tested under regime changes or other settings where temporal relationships vary within a window.
W2: Causal padding does not clearly improve performance under the standard forecasting protocol, as the observed differences are no larger than the normal variation between repeated training runs. Its benefit is demonstrated mainly through the additional arbitrary-history experiment.
W3: TiF-V assumes a fixed variable set because its circular transform and static kernel depend on . The experiments do not evaluate how its performance is affected when variables are added, removed, or reordered after training.
Questions for authors
Q1: Do the 26/28 and 21/28 win counts remain stable over multiple seeds?
Q2: How robust is TiF-V to variable reordering, insertion, or removal at test time?
2 comments
Dear Reviewer Mtne,
Thank you again for your constructive review.
We have added three-seed results (42/43/44) across all seven datasets at L=H=96, with means and standard deviations: TiF-V achieves lower mean MSE on 7/7 datasets, and TiF-T on 6/7. We also provide frozen-model experiments covering variable reordering, 25% removal, and insertion, documenting the observed behavior under each change. Could you please confirm whether this additional evidence resolves your questions about seed stability and sensitivity to variable-set changes?
We would greatly appreciate your feedback and updated assessment.
Best regards,
The Authors
Response to Reviewer Mtne
Thank you for your thoughtful review and recognition of the axis-specific design and mixer substitutions. We appreciate your constructive suggestions.
W1. Temporal relationships that change within a window
Each temporal layer generates an input-dependent lag kernel shared across positions. Queries and output gates vary by position, giving position-dependent responses.
We evaluate causal-fluxer and causal-attention decoders on Atec/Solar (seed 42, FP32). The diagnostic measures changes in lag-1 input correlation between window halves and defines high-change windows using the 90th percentile fitted to training inputs (post hoc analysis). MSE is:
Under natural within-window variation, fluxer MSE is comparable to causal attention on Atec and lower on Solar. Solar has 8528 correlation-eligible windows; full-test MSE covers all 20,929.
W2. What causal padding contributes
Causal closure enables every temporal position to predict its future without reading wrapped-around targets. Under the same position-wise objective, causal mixing beats circular mixing in 35/35 shorter-history settings, with median circular/causal MSE ratio 1.600. One causal model also beats history-specific models in 26/35 settings (seed 42; nonoverlapping patches, RevIN off). Causal attention supports the same history capability.
The full-window head reads the entire history, and both closures give comparable accuracy there. The position-wise objective uses causal closure to train every position and serve multiple history lengths.
W3. Dependence on the variable set
TiF-V models a consistent variable schema with a C-dependent static kernel and order-dependent circular mixing. Reduced variable sets use a prefix of that kernel. Q2 reports frozen-weight tests of reordering, removal, and insertion.
Q1. Stability across training seeds
We report FP32 MSE mean±sample SD at L=H=96 over seeds 42/43/44. Each arm retains its seed-42 LR, batch, and training recipe across seeds.
The no-static variant uses the same iTransformer references:
TiF-V has lower mean MSE on 7/7 datasets and wins all three paired seeds on 5/7; TiF-T and the no-static variant each have lower means on 6/7. Variable-model global batches are AQ 32/32, Atec 16/16, Meter/Temp/Wind 32/32, Solar 24/12, Traffic 8/8; no-static Atec/Traffic use 32/16. LR is 1e-3. These repeats cover H=96. The original 26/28 and both 21/28 counts describe the seed-42 four-horizon grids.
Q2. Reordering, insertion, and removal at test time
Frozen tests (seed 42, BF16) use five random permutations, shift, and reversal, with inverse-permuted outputs and fixed calendar positions. Deletion changes C by removing 25% of variables. Parentheses show MSE increases over clean inputs on the same retained targets.
TiF-V maintains lower MSE on Solar/Traffic under all seven reorderings and at 25% removal. Worst TiF-V increases are 6.690%/47.783%/4.412% for AQ/Solar/Traffic; iTransformer changes by <0.001%. On AQ, reordered TiF-V has higher MSE than iTransformer.
Deletion uses the existing static-kernel prefix with frozen weights. Appending one zero variable yields a kernel-size mismatch for TiF-V on all three datasets; iTransformer executes the expanded shape. This is an input-shape probe.