acceptodds
Under review as a conference paper at ICLR 2027

TS-CAN: Time Series Forecasting via the Clifford Geometric Product

Abstract

Deep time series forecasters have advanced rapidly by redesigning where representations interact, whether across patches, variables, or scales. The interaction itself, however, remains largely limited to dot-product similarities or learned weighted sums. These classical primitives solely measure how strongly two representations agree. Classical time series analysis instead suggests that relating two signals requires both an in-phase component (co-movement) and a quadrature component (oriented, out-of-phase structure). In this work, we show that the geometric product of Clifford algebra, , yields exactly this dual decomposition within a single operation. Its generalized inner product captures scalar agreement, while its antisymmetric wedge term serves as a discrete, learnable counterpart of the Lévy area from rough path theory, capturing higher-order directional structure. Building on this principled mathematical foundation, we propose TS-CAN, a forecaster whose sole token interaction is the geometric product itself. Applied causally along time and sparsely across the feature dimension, this geometric interaction is representationally dense enough to internalize nonlinear channel mixing, entirely eliminating the need for self-attention and heavy feed-forward networks (FFNs). Operating at a strict linear complexity of , TS-CAN achieves the lowest MSE in 28 of 32 long-term forecasting settings across eight benchmarks, while performing strongly on short-term tasks. These findings establish a new accuracy-efficiency Pareto frontier and suggest that how representations interact is a fundamental, unexplored design dimension for time series modeling.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Clifford Algebra, Geometric Product, Sparse Interaction, Long-Term Forecasting

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_rbt8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TS-CAN, a time series forecasting model built around Clifford geometric products. It uses inner-product and wedge-product terms to model alignment and directional relationships, and combines temporal and channel-wise interactions with multi-scale context. The paper also relates the wedge term to discrete Lévy area and the quadrature spectrum of cross-covariance. TS-CAN reports the best or tied MSE in 28 of 32 long-term forecasting settings and competitive results on short-term tasks.

Critical strengths
  1. Novel idea. I find the use of Clifford geometric products for forecasting interesting. The symmetric and antisymmetric terms offer a clear way to think about directional information, and the links to Lévy area and spectral analysis help motivate the design.

  2. Strong results. TS-CAN performs well on the tested benchmarks, particularly Traffic and Exchange. The attention free design and reported inference efficiency are also appealing.

Critical weaknesses
  1. Ablations are limited to one main settingMost component ablations are run on ETTh1-96. I would like to see whether the temporal and channel-wise wedge terms also help on Traffic and Exchange, where the model shows its largest gains. Table 5 also suggests that the benefit of the wedge term varies across ETT datasets. In addition, the unstructured bilinear baseline in Table 3(a) changes the activation as well as the parameterization, so it does not completely isolate the effect of explicit antisymmetry.

2.Incomplete efficiency comparisonThe reported memory use and inference latency look promising. However, TimeMixer++, the strongest accuracy baseline in the main experiments, is missing from Tables 6 and 9. Including it, preferably using the configurations behind the accuracy results, would make the claimed accuracy-efficiency trade-off easier to assess.

Questions for authors
  1. Could the authors report separate temporal- and channel-wise wedge ablations on Traffic and Exchange, preferably for a longer prediction horizon as well?

  2. In Table 3(a), does the explicit antisymmetric parameterization still outperform the unstructured bilinear baseline when the SiLU activation and other components are kept the same?

  3. Could the authors include TimeMixer++ in the efficiency comparison and report accuracy and latency for the configurations used in the forecasting experiments?

8 comments
AuthorsOpenReview ↗

Response to Reviewer rbt8 [1/4]

Many thanks to the reviewer for the detailed and insightful comments, which helped us improve the paper. In response, we (i) ran separate temporal and channel wedge ablations on all eight datasets and four horizons, including Traffic and Exchange (W1, Q1). We also (ii) added activation-matched, parameter-matched antisymmetry controls on the same 32 settings (W1, Q2), and (iii) added TimeMixer++ to the efficiency comparison under the same protocol as the other models (W2, Q3).

Q1: Ablations on Traffic and Exchange

We thank the reviewer for this suggestion. We ran the separate temporal and channel wedge ablations on all eight datasets at all four horizons, with the same configuration and training protocol as the main experiments. Each variant removes only the wedge term of the specified branch and keeps the in-phase term and all other components unchanged. The table reports MSE, with the relative change with respect to TS-CAN in parentheses.

DatasetFFTS-CANw/o temporal wedgew/o channel wedgew/o both wedges
ETTh1960.3610.382 (+5.8%)0.380 (+5.3%)0.380 (+5.3%)
1920.4090.434 (+6.1%)0.430 (+5.1%)0.415 (+1.5%)
3360.4290.443 (+3.3%)0.443 (+3.3%)0.461 (+7.5%)
7200.4490.448 (−0.2%)0.455 (+1.3%)0.479 (+6.7%)
ETTh2960.2800.297 (+6.1%)0.297 (+6.1%)0.298 (+6.4%)
1920.3450.377 (+9.3%)0.386 (+11.9%)0.365 (+5.8%)
3360.3470.392 (+13.0%)0.394 (+13.5%)0.404 (+16.4%)
7200.3920.430 (+9.7%)0.424 (+8.2%)0.423 (+7.9%)
ETTm1960.2860.305 (+6.6%)0.304 (+6.3%)0.306 (+7.0%)
1920.3320.335 (+0.9%)0.338 (+1.8%)0.338 (+1.8%)
3360.3690.368 (−0.3%)0.372 (+0.8%)0.369 (0.0%)
7200.4220.423 (+0.2%)0.420 (−0.5%)0.428 (+1.4%)
ETTm2960.1680.178 (+6.0%)0.177 (+5.4%)0.169 (+0.6%)
1920.2280.234 (+2.6%)0.234 (+2.6%)0.237 (+3.9%)
3360.2800.287 (+2.5%)0.289 (+3.2%)0.285 (+1.8%)
7200.3570.385 (+7.8%)0.374 (+4.8%)0.375 (+5.0%)
Electricity960.1350.148 (+9.6%)0.148 (+9.6%)0.143 (+5.9%)
1920.1470.161 (+9.5%)0.161 (+9.5%)0.164 (+11.6%)
3360.1640.177 (+7.9%)0.178 (+8.5%)0.178 (+8.5%)
7200.2080.215 (+3.4%)0.216 (+3.8%)0.219 (+5.3%)
Weather960.1510.163 (+7.9%)0.162 (+7.3%)0.162 (+7.3%)
1920.1960.203 (+3.6%)0.205 (+4.6%)0.203 (+3.6%)
3360.2370.259 (+9.3%)0.260 (+9.7%)0.259 (+9.3%)
7200.3110.337 (+8.4%)0.335 (+7.7%)0.336 (+8.0%)
Traffic960.3610.403 (+11.6%)0.402 (+11.4%)0.412 (+14.1%)
1920.3700.423 (+14.3%)0.420 (+13.5%)0.428 (+15.7%)
3360.3880.434 (+11.9%)0.436 (+12.4%)0.441 (+13.7%)
7200.4270.467 (+9.4%)0.470 (+10.1%)0.476 (+11.5%)
Exchange960.0820.095 (+15.9%)0.093 (+13.4%)0.092 (+12.2%)
1920.1720.177 (+2.9%)0.176 (+2.3%)0.189 (+9.9%)
3360.3070.329 (+7.2%)0.336 (+9.4%)0.334 (+8.8%)
7200.7920.907 (+14.5%)0.890 (+12.4%)0.938 (+18.4%)
Average0.3090.332 (+7.2%)0.331 (+7.1%)0.335 (+8.1%)

TS-CAN has lower MSE than the ablated variant in 92 of the 96 comparisons. There is one tie, and the three exceptions are all within 0.5%. Averaged over the 32 settings, removing the temporal wedge, the channel wedge, or both increases MSE by 7.2%, 7.1%, and 8.1%.

AuthorsOpenReview ↗

Response to Reviewer rbt8 [2/4]

Traffic and Exchange. Both wedge terms help at every horizon on both datasets.

  • Traffic. Removing either wedge costs 9.4%–14.3%, and removing both costs 11.5%–15.7%. These are the largest and most stable effects among all datasets. Traffic sensors share daily and weekly cycles with sensor-specific phase offsets, which is the lead–lag structure that the antisymmetric term is designed to capture.
  • Exchange. The effect is largest at the longest horizon. At F=720F=720, removing the temporal wedge, the channel wedge, or both increases MSE by 14.5%, 12.4%, and 18.4%.

Complementarity at long horizons. On ETTh1 at F=720F=720, removing only one wedge costs at most 1.3%, but removing both costs 6.7%. A similar pattern holds on ETTh1 at F=336F=336 and on Exchange and Traffic at F=720F=720. The two branches therefore carry partly overlapping directional information, and losing both is clearly worse than losing either.

Variation across ETT datasets. The wedge terms help on all four ETT datasets, and the size of the benefit depends on both dataset and horizon.

  • ETTh1 and ETTh2. The gain is clear at every horizon, reaching 16.4% on ETTh2 at F=336F=336.
  • ETTm2. The gain is moderate at intermediate horizons and larger at F=720F=720 (5.0%–7.8%).
  • ETTm1. The effect is smallest beyond F=96F=96, where the variants stay within 1.8% of TS-CAN. At F=96F=96 the gain is still 6.3%–7.0%.
  • ETTh2 in the surrogate study (Table 6). The full model did not reach its main-table accuracy (0.305 vs. 0.280), because that study uses a single training configuration (App. B). Under the main configuration, the wedge terms help on ETTh2 at all four horizons.

This pattern is consistent with the wedge acting as an estimator of directional structure (Props. 1–2). It helps where such structure is present and remains predictable over the forecast window, and on ETTh1 and ETTm1 its gain disappears once IAAFT surrogates remove this structure.

AuthorsOpenReview ↗

Response to Reviewer rbt8 [3/4]

Q2: Activation-matched antisymmetry comparison

Yes. To answer this directly, we designed controls that keep the SiLU activation, the context, and all other components of TS-CAN fixed, and ran them on all eight datasets at all four horizons. Let A=S⊙TkC~A=\mathbf S\odot\mathcal T_k\tilde{\mathbf C} and B=C~⊙TkSB=\tilde{\mathbf C}\odot\mathcal T_k\mathbf S. TS-CAN feeds [A−B, SiLU(A)][A-B,\ \mathrm{SiLU}(A)] to the projection P\mathcal P. The controls are:

  • V1, sign flip: [A+B, SiLU(A)][A+B,\ \mathrm{SiLU}(A)];
  • V2, activation-matched unstructured interaction: [SiLU(A), SiLU(B)][\mathrm{SiLU}(A),\ \mathrm{SiLU}(B)];
  • V3, ungated: [A−B, A][A-B,\ A];
  • V4, symmetric in-phase term: [A−B, SiLU(12(A+B))][A-B,\ \mathrm{SiLU}(\tfrac12(A+B))];
  • V5, gated wedge: [SiLU(A−B), SiLU(A)][\mathrm{SiLU}(A-B),\ \mathrm{SiLU}(A)];
  • V6, strict superset: [A, B, SiLU(A)][A,\ B,\ \mathrm{SiLU}(A)].

V1–V5 have exactly the same parameters and FLOPs as TS-CAN. V6 has 26% more parameters and can represent TS-CAN exactly. The table reports the average MSE of TS-CAN on each dataset and the relative change of each control.

DatasetTS-CANV1V2V3V4V5V6
ETTh10.412+5.9%+8.9%+7.5%+8.8%+6.5%+7.1%
ETTh20.341+8.7%+9.6%+8.7%+9.1%+9.5%+8.2%
ETTm10.352+2.2%+2.2%+2.7%+2.6%+1.7%+2.3%
ETTm20.258+3.8%+3.8%+3.4%+3.1%+3.3%+3.7%
Electricity0.164+6.7%+6.9%+6.2%+5.9%+6.4%+5.8%
Weather0.224+7.9%+7.6%+7.2%+7.3%+6.9%+7.8%
Traffic0.387+12.7%+13.0%+12.1%+11.6%+12.5%+11.5%
Exchange0.338+3.5%+10.2%+2.1%+11.6%+10.9%+17.9%
Average MSE0.309+6.5%+8.1%+6.4%+7.8%+7.5%+8.3%
Average MAE0.338+3.2%+3.8%+3.0%+3.6%+3.5%+4.0%

Across all 32 dataset–horizon settings, TS-CAN has lower MSE than the control in 189 of the 192 comparisons, and the three exceptions are within 0.6%.

Sign-flip control (V1). It is the most direct test of explicit antisymmetry. It keeps the activation, parameters, context, and every other component, and changes only whether the explicit linear term is odd or even. It is worse in 31 of 32 settings, by 6.5% MSE on average, and on ETTh1-96 it is worse on each of three seeds (0.3662 ± 0.0015 vs. 0.3614 ± 0.0014).

Activation-matched control (V2). It applies the same SiLU to both raw products, which is the comparison the reviewer asks for. It is worse in all 32 settings, by 8.1% on average. On ETTh1-96, the setting of the original ablation, it is 2.8% worse, close to the 2.4% gap of the original control. The gap therefore remains once the activation is matched.

Not a capacity effect. The strict superset (V6) is worse in all 32 settings as well. Its largest losses are on Exchange, where the additional freedom overfits a small, non-stationary dataset.

Where the effect is largest. On Traffic, every control is 9.0%–15.4% worse at every horizon. The effect remains clear on Exchange, especially for the controls that remove or gate the explicit wedge (10%–18% on average).

How the two components should be treated. Removing the gate from the in-phase term (V3), symmetrizing it (V4), and moving the gate onto the wedge (V5) each cost 6%–8%, similar to flipping the parity. The best configuration keeps the antisymmetric term explicit, odd, and linear, and applies the nonlinearity only to the in-phase term. This agrees with our analysis: the wedge changes sign under time reversal, whereas SiLU would add an even, reversal-invariant component to it.

Reviewer_c9yBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces Clifford algebra's geometric product to capture both in-phase agreement and out-of-phase directional structure in time series. Since this single interaction is representationally rich enough, the proposed model TS-CAN requires no self-attention or heavy feed-forward networks. Experiments show it achieves the lowest MSE in 28 of 32 long-term forecasting settings across eight benchmarks. The idea is interesting, and the evaluation seems to be solid.

Critical strengths

(1) It is novel to introduce the Clifford algebra geometric product into time series modeling to capture both in-phase agreement and out-of-phase directional structure.

(2) The manuscript is well-written, and the evaluation is solid. According to the evaluation, TS-CAN achieves state-of-the-art long- and short-term forecasting accuracy.

Critical weaknesses

(1) The references seem to be outdated, with most dating from before 2024 and only one survey and one other paper from 2026. It is also unclear whether methods from other lines, such as diffusion-based approaches, should be included as comparison baselines. (2) In terms of writing, the abstract lacks a definite evaluation and conclusion. (3) Long-term forecasting is one of the key challenges, yet Table 1 shows that the proposed method appears to have no advantage over TimeMixer, and is notably worse in MAE. In addition, the paper does not clearly discuss its relation to Clifford Transformer-style work, which uses multivectors as the carrier of attention and exploits the equivariance of the geometric product to inject geometric inductive bias. This line of work suggests that the geometric product is a valuable alternative at the level of how representations interact, and the manuscript should clarify how it differs from or improves upon it. (4) Finally, is there any theoretical basis or reasoning for introducing Clifford algebra, and is its scientific soundness sufficiently justified? Implementation details are also missing, making the results difficult to reproduce.

Questions for authors

see in weakness

6 comments
AuthorsOpenReview ↗

Response to Reviewer c9yB [1/4]

We truly appreciate the reviewer's thoughtful and constructive feedback. In response, we (i) added two recent diffusion-based baselines, ARMD and NsDiff, and two geometric-algebra baselines, a Clifford Transformer and CliffordNet-1D (see our response to Reviewer aB3n), all evaluated on the 32 long-term settings (W1, W3). We also (ii) revised the abstract to state concrete results and the conclusion (W2), (iii) clarified the comparison with TimeMixer and the relation to Clifford Transformer-style work (W3), and (iv) added the theoretical justification and a complete forward-pass algorithm (W4).

W1: Outdated references and missing baselines

We thank the reviewer for this suggestion. We agree that diffusion-based forecasting is an important recent line of work. We have added two diffusion-based baselines, ARMD (Gao et al., AAAI 2025), an auto-regressive moving diffusion model, and NsDiff (Ye et al., ICML 2025), a diffusion model for non-stationary series. We have also added the corresponding references and a discussion to the related work. Both models were trained on all eight datasets at all four horizons using the official implementations, with lookback 96 and the same data splits as TS-CAN.

DatasetFFARMD MSE / MAENsDiff MSE / MAETS-CAN MSE / MAE
ETTh1960.456 / 0.4440.485 / 0.4720.361 / 0.393
1920.510 / 0.4870.601 / 0.5420.409 / 0.425
3360.591 / 0.5400.547 / 0.5260.429 / 0.434
7200.741 / 0.6320.638 / 0.5680.449 / 0.467
ETTh2960.451 / 0.4750.379 / 0.4180.280 / 0.334
1920.645 / 0.5710.510 / 0.5060.345 / 0.383
3360.932 / 0.6930.420 / 0.4700.347 / 0.393
7201.535 / 0.9020.477 / 0.5130.392 / 0.426
ETTm1960.603 / 0.5200.513 / 0.4500.286 / 0.345
1920.662 / 0.5540.554 / 0.4770.332 / 0.370
3360.725 / 0.5930.559 / 0.4960.369 / 0.395
7200.830 / 0.6560.679 / 0.5710.422 / 0.426
ETTm2960.332 / 0.4180.193 / 0.3040.168 / 0.258
1920.516 / 0.5210.247 / 0.3390.228 / 0.298
3360.804 / 0.6510.281 / 0.3640.280 / 0.330
7201.446 / 0.8770.523 / 0.4780.357 / 0.388
Exchange960.193 / 0.3540.209 / 0.3230.082 / 0.203
1920.338 / 0.4610.241 / 0.3740.172 / 0.296
3360.557 / 0.5830.538 / 0.5320.307 / 0.403
7200.994 / 0.7841.059 / 0.7000.792 / 0.666
Traffic960.966 / 0.5030.677 / 0.4060.361 / 0.261
1920.843 / 0.4730.693 / 0.4170.370 / 0.271
3360.863 / 0.4860.699 / 0.4190.388 / 0.279
7200.940 / 0.5190.759 / 0.4390.427 / 0.292
Electricity960.272 / 0.3450.204 / 0.3220.135 / 0.231
1920.282 / 0.3680.218 / 0.3360.147 / 0.243
3360.335 / 0.4160.234 / 0.3420.164 / 0.266
7200.526 / 0.5420.265 / 0.3670.208 / 0.296
Weather960.227 / 0.2920.215 / 0.2970.151 / 0.198
1920.257 / 0.3210.293 / 0.3190.196 / 0.241
3360.304 / 0.3600.309 / 0.3510.237 / 0.275
7200.389 / 0.4300.364 / 0.3840.311 / 0.328
Average0.627 / 0.5240.456 / 0.4320.309 / 0.338

TS-CAN outperforms both diffusion models in all 32 settings on both MSE and MAE. Averaged over the 32 settings, it reduces MSE by 30.3% relative to NsDiff and by 46.4% relative to ARMD. NsDiff is the stronger of the two and comes closest on ETTm2, where the two are nearly tied at F=336F=336 (0.281 vs. 0.280). The gaps are largest on Traffic, where TS-CAN reduces MSE by 44%–47% relative to NsDiff at every horizon. Diffusion models are designed to model the predictive distribution, whereas TS-CAN is a deterministic forecaster optimized for point accuracy. Under the point-forecast protocol of this paper, TS-CAN is clearly more accurate, and it also avoids iterative sampling at inference.

W2: Abstract lacks an evaluation and conclusion

We thank the reviewer for this suggestion. We have added two sentences to the abstract. The first states the main results with concrete numbers on accuracy, short-term performance, and efficiency. The second states the conclusion supported by our ablations and surrogate tests: the gain comes from the explicit antisymmetric term and depends on directional structure in the data. The closing statement of the abstract now follows directly from the reported evidence.

AuthorsOpenReview ↗

Response to Reviewer c9yB [2/4]

W3: No advantage over TimeMixer; relation to Clifford Transformer work

We thank the reviewer for these comments. We first clarify Table 1, and then discuss and evaluate Clifford Transformer-style models.

Comparison with TimeMixer. TS-CAN is better than TimeMixer on all eight datasets in both average MSE and average MAE. The average MSE reduction is 9.8%, with the largest gains on Traffic (0.387 vs. 0.484, −20.0%) and Exchange (0.338 vs. 0.391, −13.6%). The average MAE reduction is 4.8%. Per horizon, TS-CAN has lower MSE in all 32 settings and lower MAE in 30, with one tie and one loss.

The strongest baseline in Table 1 is TimeMixer++. TS-CAN has lower average MSE than TimeMixer++ on 7 of the 8 datasets, by 2.9% on average and by 7.0% on Traffic. The advantage is clearest at F=720F=720, the most demanding long-term setting. There TS-CAN has the lowest MSE on 7 of 8 datasets, ties on the eighth, and is best in MAE on 4 of 8. In average MAE the two models are close: TS-CAN is better on ETTh1, ETTm2, and Weather, and the differences on the other five datasets range from 0.3% to 4.5%. TS-CAN is trained with the MSE loss, which weights large errors more heavily, so its gains are concentrated on large deviations. This is the regime that matters most at long horizons, where errors accumulate.

Relation to Clifford Transformer-style work. We agree that this line of research supports the view that the geometric product is a valuable choice for how representations interact. TS-CAN differs in what the product is applied to, which symmetry it encodes, and what it replaces.

  • What it is applied to. GATr and Clifford group equivariant networks use multivectors as the data type to represent points, planes, or physical fields. They exploit the E(n)-equivariance of the geometric product and retain attention and MLP blocks. Latent time series features have no physical frame, so spatial equivariance is not the relevant inductive bias.
  • Which symmetry it encodes. The symmetry that matters in forecasting is time reversal. The wedge term flips sign under reversal, whereas every symmetric second-order interaction is invariant (Prop. 6). TS-CAN therefore uses the geometric product as a causal interaction between real-valued token states, rather than as a carrier inside attention.
  • What it replaces. A full multivector over Rd\mathbb{R}^d has 2d2^d components, and its bivector part alone has d(d−1)/2d(d-1)/2. TS-CAN instead samples ∣K∣|\mathcal K| bivector diagonals at O(∣K∣d)O(|\mathcal K|d) cost per token, and it replaces both attention and the FFN, so its cost is linear in sequence length.
AuthorsOpenReview ↗

Response to Reviewer c9yB [3/4]

Clifford Transformer baseline. To make the comparison concrete, we implemented a Clifford Transformer for forecasting. Patch tokens are projected into a real-valued embedding space and processed by bidirectional Clifford-inspired attention, with scores scoreij=dotij+β (dotij−dotji)\mathrm{score}_{ij} = \mathrm{dot}_{ij} + \beta\,(\mathrm{dot}_{ij} - \mathrm{dot}_{ji}), where dotij=si⊤cj/dh\mathrm{dot}_{ij} = \mathbf{s}_i^\top \mathbf{c}_j / \sqrt{d_h}. Each encoder layer then applies a standard Transformer feed-forward layer. It shares the normalization, patching, and linear head with TS-CAN. The configuration is dmodel=128d_{\text{model}}=128, 16 heads, 3 layers, dff=256d_{\text{ff}}=256, patch length 16, stride 8, and dropout 0.2, trained for 10 epochs with early stopping (patience 3), following the Time-Series-Library protocol. The full results are as follows.

DatasetFFClifford Transformer MSE / MAETS-CAN MSE / MAEΔMSEΔMAE
ETTh1960.390 / 0.4020.361 / 0.393−7.4%−2.2%
1920.434 / 0.4300.409 / 0.425−5.8%−1.2%
3360.467 / 0.4500.429 / 0.434−8.1%−3.6%
7200.463 / 0.4710.449 / 0.467−3.0%−0.8%
ETTh2960.290 / 0.3410.280 / 0.334−3.4%−2.1%
1920.371 / 0.3930.345 / 0.383−7.0%−2.5%
3360.417 / 0.4310.347 / 0.393−16.8%−8.8%
7200.429 / 0.4490.392 / 0.426−8.6%−5.1%
ETTm1960.353 / 0.3850.286 / 0.345−19.0%−10.4%
1920.380 / 0.3970.332 / 0.370−12.6%−6.8%
3360.411 / 0.4140.369 / 0.395−10.2%−4.6%
7200.465 / 0.4440.422 / 0.426−9.2%−4.1%
ETTm2960.177 / 0.2620.168 / 0.258−5.1%−1.5%
1920.243 / 0.3050.228 / 0.298−6.2%−2.3%
3360.304 / 0.3460.280 / 0.330−7.9%−4.6%
7200.402 / 0.4020.357 / 0.388−11.2%−3.5%
Exchange960.087 / 0.2070.082 / 0.203−5.7%−1.9%
1920.180 / 0.3020.172 / 0.296−4.4%−2.0%
3360.337 / 0.4210.307 / 0.403−8.9%−4.3%
7200.877 / 0.7070.792 / 0.666−9.7%−5.8%
Traffic960.511 / 0.3500.361 / 0.261−29.4%−25.4%
1920.514 / 0.3500.370 / 0.271−28.0%−22.6%
3360.532 / 0.3590.388 / 0.279−27.1%−22.3%
7200.568 / 0.3730.427 / 0.292−24.8%−21.7%
Electricity960.204 / 0.3060.135 / 0.231−33.8%−24.5%
1920.211 / 0.3140.147 / 0.243−30.3%−22.6%
3360.227 / 0.3280.164 / 0.266−27.8%−18.9%
7200.266 / 0.3540.208 / 0.296−21.8%−16.4%
Weather960.180 / 0.2230.151 / 0.198−16.1%−11.2%
1920.229 / 0.2620.196 / 0.241−14.4%−8.0%
3360.279 / 0.2990.237 / 0.275−15.1%−8.0%
7200.356 / 0.3490.311 / 0.328−12.6%−6.0%

TS-CAN outperforms the Clifford Transformer in all 32 settings on both MSE and MAE. Averaged over the 32 settings, it reduces MSE by 14.1% and MAE by 8.9%. The largest gaps are on Traffic and Electricity, between 21.8% and 33.8% in MSE at every horizon. These datasets have the most variables and the strongest periodic structure, and Traffic in particular has sensor-specific phase offsets. This is lead–lag structure along time, which a causal temporal wedge captures directly, whereas attention over multivector tokens mixes it into a single similarity score per head. The comparison shows that using the geometric product is not sufficient by itself; what matters is how it is used. In TS-CAN it is a causal, directional interaction that replaces attention, rather than a component placed inside attention.

Reviewer_aB3nOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TS-CAN, which combines patch embeddings, multiscale contexts, and multiplicative interactions containing gated products and antisymmetric differences. These interactions operate across time and latent features, followed by a linear forecasting head. The paper's main claim is that explicitly modeling directional structure improves forecasting while avoiding attention and a standalone FFN. Experiments have validate the performances of the proposed method.

Critical strengths
  1. Strong reported long-term MSE results. Table 15 supports the claimed 28/32 wins when ties are included: 22 strict wins and six ties. The Traffic improvement is particularly substantial.

  2. Useful mathematical interpretation. The discrete-area identity and the relationship between expected wedge features and cross-covariance provide an interpretable account of the antisymmetric operator.

  3. Meaningful attempts to explain the mechanism. Component ablations, parameter-matched alternative mixers, surrogate data, and direct delay-detection tests go beyond reporting benchmark accuracy.

Critical weaknesses
  1. The novelty relative to CliffordNet needs stronger separation. CliffordNet already contains the shifted products, antisymmetric differences, differential context, global mean context, and gated residual update used here. TS-CAN acknowledges this source, but the overlap extends beyond an implementation trick. The main additions appear to be temporal shifts, adaptive multiscale context, and forecasting-specific analysis. A direct one-dimensional CliffordNet baseline would clarify their contribution.

  2. The implemented interaction is not literally the Clifford geometric product. The wedge branch extracts valid selected bivector coefficients. However, the “in-phase” branch applies SiLU to shifted outer-product entries; it does not compute the scalar inner product S^TC. The manuscript itself acknowledges that this branch is not symmetric and loses the exact spectral correspondence. Consequently, describing the interaction as the “geometric product itself” overstates its algebraic properties.

  3. The key ablation does not isolate antisymmetry. The unstructured bilinear control removes both antisymmetric subtraction and SiLU. Its 2.4% degradation could therefore reflect the changed activation. An activation-matched comparison is needed to attribute the improvement specifically to antisymmetry. The inner-only and wedge-only variants also have fewer parameters, and the detailed architectural ablations are concentrated on ETTh1 at horizon 96.

Questions for authors

See weakness. I might adjust my score based on the author response.

6 comments
AuthorsOpenReview ↗

Response to Reviewer aB3n [1/4]

We sincerely thank the reviewer for the careful and constructive review. In response, we (i) ran a one-dimensional CliffordNet baseline on all 32 long-term settings (W1), (ii) clarified which grades of the geometric product our layer computes and where the nonlinearity acts (W2), and (iii) added controls matched in parameters, FLOPs, and activation on eight datasets and four horizons (W3).

W1: Novelty relative to CliffordNet

We thank the reviewer for suggesting the 1-D baseline, which we have now run on all 32 settings. We adopt CliffordNet's shifted realization as an efficient way to compute sampled geometric products, and we credit it as such. Our contribution lies elsewhere: we establish what this product means for temporal data, design a forecaster around that meaning, and test the resulting predictions. The relation is similar to that between WaveNet and image CNNs. Both use convolution, but what makes WaveNet work on sequences is its causal, dilated, multi-scale design.

In vision, spatial orientation carries no intrinsic statistical asymmetry, which is why horizontal flipping is a standard augmentation. In time series, reversing time changes the statistics; the irreversibility index of the ETT datasets ranges from 0.22 to 0.48 (Table 6). The antisymmetric term therefore has a meaning in forecasting that it does not have in CliffordNet's setting. Under causal shifts, the temporal wedge is a discrete Lévy area of the state–context path (Prop. 1). Its expectations are the sine coefficients of the quadrature spectrum (Prop. 2). It also flips sign under time reversal, whereas every symmetric second-order interaction is invariant. These results are new to this work, and they lead to predictions that we verify. The untrained operator recovers the true delay and its sign (Table 10), and the wedge gain vanishes on surrogates that preserve all second-order statistics (Table 6).

The architecture follows from this analysis. CliffordNet has no interaction along time. The temporal Clifford branch, the causal zero-padded shifts that anchor the path at the origin, the multi-lag shift set, and the adaptive dilated multi-scale context are therefore specific to TS-CAN. We also give the differential context a new role: it leaves the wedge unchanged and only removes the self-energy of the in-phase term (Remark 1). The ablations show that the gains come from the forecasting-specific components rather than the shared ones. Removing the temporal branch increases MSE by 2.9%, and restricting the context to a single scale increases it by 4.9%. In contrast, replacing the differential context with an absolute one changes MSE by only 0.2%, and replacing the gated update with direct addition leaves it unchanged.

Following the reviewer's suggestion, we implemented CliffordNet-1D with the same hyperparameters as TS-CAN. The full results are as follows.

DatasetFFCliffordNet-1DMSE/MAETS-CANMSE/MAEΔMSEΔMAE
ETTh1960.428/0.4410.361/0.393−15.7%−10.9%
1920.483/0.4780.409/0.425−15.4%−11.2%
3360.510/0.4880.429/0.434−15.9%−11.1%
7200.574/0.5370.449/0.467−21.7%−13.1%
ETTh2960.341/0.3800.280/0.334−18.0%−12.1%
1920.428/0.4340.345/0.383−19.4%−11.8%
3360.454/0.4540.347/0.393−23.6%−13.5%
7200.473/0.4740.392/0.426−17.0%−10.1%
ETTm1960.356/0.3890.286/0.345−19.6%−11.2%
1920.409/0.4170.332/0.370−18.8%−11.4%
3360.426/0.4300.369/0.395−13.4%−8.2%
7200.522/0.4880.422/0.426−19.2%−12.8%
ETTm2960.200/0.2840.168/0.258−15.9%−9.2%
1920.275/0.3310.228/0.298−17.1%−9.9%
3360.325/0.3580.280/0.330−13.9%−7.8%
7200.470/0.4490.357/0.388−24.0%−13.5%
Exchange960.103/0.2290.082/0.203−20.3%−11.5%
1920.201/0.3250.172/0.296−14.3%−8.8%
3360.334/0.4200.307/0.403−8.0%−3.9%
7201.060/0.7870.792/0.666−25.3%−15.3%
Traffic960.726/0.4190.361/0.261−50.3%−37.7%
1920.750/0.4330.370/0.271−50.7%−37.4%
3360.718/0.4310.388/0.279−46.0%−35.2%
7200.775/0.4360.427/0.292−44.9%−33.0%
Electricity960.201/0.3090.135/0.231−33.0%−25.1%
1920.210/0.3130.147/0.243−30.0%−22.5%
3360.223/0.3250.164/0.266−26.5%−18.2%
7200.251/0.3460.208/0.296−17.3%−14.5%
Weather960.173/0.2270.151/0.198−12.6%−12.6%
1920.240/0.2840.196/0.241−18.3%−15.2%
3360.287/0.3150.237/0.275−17.3%−12.6%
7200.363/0.3640.311/0.328−14.4%−9.9%

TS-CAN outperforms CliffordNet-1D in all 32 settings on both MSE and MAE, reducing MSE by 22.4% and MAE by 15.4% on average. The gap is largest on Traffic (45%–51%), which has lead–lag structure along time that an interaction acting only across features cannot capture. The gain therefore comes not from geometric algebra as such, but from using the product as a causal, directional interaction between temporal states.

AuthorsOpenReview ↗

Response to Reviewer aB3n [2/4]

W2: The interaction is not literally the geometric product

We thank the reviewer for this careful reading and welcome the chance to state the algebraic properties precisely. Before the activation, TS-CAN computes the components of the geometric product exactly, in channel-resolved form. A nonlinearity is then applied to the scalar grade only, as is standard in geometric-algebra networks. In this sense, the interaction is a realization of the geometric product rather than an approximation of it.

In the temporal branch, the shifted entries Sn,jCn−k,jS_{n,j}C_{n-k,j} form the diagonal of the outer product SnCn−k⊤\mathbf S_n\mathbf C_{n-k}^\top. Their sum over channels is Sn⋅Cn−k\mathbf S_n\cdot\mathbf C_{n-k}, the scalar part of the geometric product between a token state and its kk-step context. In the feature branch, the cyclic shift is an orthogonal map. By the Cartan–Dieudonné theorem it can be written as a versor action VkCV~kV_k\mathbf C\tilde V_k, so the summed entries give the scalar part of S(VkCV~k)\mathbf S(V_k\mathbf C\tilde V_k). We deliberately keep these terms per channel instead of collapsing them into one number per shift. The projection P\mathcal P recovers the scalar exactly with a row of ones, so the representation contains the inner product while retaining more information, much as several attention heads carry more than one. Together with the exact bivector coefficients from the wedge branch, the pre-activation features therefore contain the grade-0 part and the sampled grade-2 part of the product. We now state this as a proposition in App. C.

Applying a nonlinearity after the geometric product is standard in this line of work. Clifford neural layers, CGENN, GATr, and CliffordNet all do so, and GATr gates only the scalar grade so that the other grades keep their transformation behavior. TS-CAN follows the same principle. SiLU acts only on the in-phase grade, which carries no orientation. The bivector grade stays linear, so it keeps its exact spectral meaning (Prop. 2) and its sign flip under time reversal. The asymmetry noted in the manuscript refers to exchanging state and context, which play different roles in our model. The geometric product is itself non-commutative, and only its scalar part is symmetric. Likewise, the co-spectral correspondence holds exactly for the pre-activation in-phase features, and the gate is applied afterwards as a design choice.

To verify that the results do not depend on the gate, we ran controls on ETTh1-96 with three seeds. All variants have the same parameters and FLOPs as TS-CAN. Here A=S⊙TkC~A=\mathbf S\odot\mathcal T_k\tilde{\mathbf C} and B=C~⊙TkSB=\tilde{\mathbf C}\odot\mathcal T_k\mathbf S.

VariantInput to P\mathcal PMSEMAEΔMSE
TS-CAN[A−B, SiLU(A)][A-B,\,\mathrm{SiLU}(A)]0.3614 ± 0.00140.3935 ± 0.0011–
No gate[A−B, A][A-B,\,A]0.3622 ± 0.00200.3940 ± 0.0014+0.24%
Symmetric in-phase[A−B, SiLU(12(A+B))][A-B,\,\mathrm{SiLU}(\tfrac12(A+B))]0.3712 ± 0.01770.3981 ± 0.0082+2.71%
Gated wedge[SiLU(A−B), SiLU(A)][\mathrm{SiLU}(A-B),\,\mathrm{SiLU}(A)]0.3723 ± 0.01610.3983 ± 0.0078+3.02%

On ETTh1-96, removing the gate matches TS-CAN within seed variance, so the literal sampled geometric-product layer is already a strong model. Across all eight datasets and four horizons, however, the grade-wise design is clearly better. Removing the gate from the scalar grade increases MSE by 6.4% on average, and moving it onto the bivector grade increases MSE by 7.5% (see our response to W3). This supports the principle above: the nonlinearity belongs on the scalar grade, and the bivector grade should stay linear. A SiLU on the wedge would add an even, reversal-invariant component and mix orientation with magnitude.

AuthorsOpenReview ↗

Response to Reviewer aB3n [3/4]

W3: The key ablation does not isolate antisymmetry

We thank the reviewer for this suggestion. To attribute the improvement specifically to antisymmetry, we designed controls that match TS-CAN in parameters, FLOPs, and activation, and ran them on eight datasets at all four horizons. Let A=S⊙TkC~A=\mathbf S\odot\mathcal T_k\tilde{\mathbf C} and B=C~⊙TkSB=\tilde{\mathbf C}\odot\mathcal T_k\mathbf S. TS-CAN feeds [A−B, SiLU(A)][A-B,\ \mathrm{SiLU}(A)] to the projection P\mathcal P. The controls are:

  • V1 (sign flip): [A+B, SiLU(A)][A+B,\ \mathrm{SiLU}(A)]
  • V2 (activation-matched unstructured): [SiLU(A), SiLU(B)][\mathrm{SiLU}(A),\ \mathrm{SiLU}(B)]
  • V3 (no gate): [A−B, A][A-B,\ A]
  • V4 (symmetric in-phase): [A−B, SiLU(12(A+B))][A-B,\ \mathrm{SiLU}(\tfrac12(A+B))]
  • V5 (gated wedge): [SiLU(A−B), SiLU(A)][\mathrm{SiLU}(A-B),\ \mathrm{SiLU}(A)]
  • V6 (strict superset): [A, B, SiLU(A)][A,\ B,\ \mathrm{SiLU}(A)]

V1–V5 have exactly the same parameters and FLOPs as TS-CAN; V6 has 26% more parameters. All runs use the same per-dataset configuration and training protocol as the main experiments, with 3 seeds. The table reports MSE / MAE for TS-CAN and the relative MSE change of each control.

DatasetFFTS-CANV1V2V3V4V5V6
ETTh1960.361/0.3930.366/0.3960.371/0.3980.362/0.3940.371/0.3980.372/0.3980.377/0.399
1920.409/0.4250.417/0.4230.413/0.4220.414/0.4220.412/0.4220.413/0.4220.414/0.424
3360.429/0.4340.463/0.4450.459/0.4430.476/0.4500.465/0.4460.451/0.4410.452/0.444
7200.449/0.4670.500/0.4820.552/0.5000.519/0.4900.544/0.5000.519/0.4890.523/0.492
ETTh2960.280/0.3340.300/0.3480.301/0.3480.298/0.3470.304/0.3490.303/0.3500.295/0.347
1920.345/0.3830.363/0.3910.369/0.3940.366/0.3960.365/0.3920.370/0.3970.361/0.396
3360.347/0.3930.397/0.4260.400/0.4260.397/0.4260.398/0.4260.399/0.4260.394/0.420
7200.392/0.4260.423/0.4490.426/0.4520.421/0.4470.422/0.4490.422/0.4490.427/0.452
ETTm1960.286/0.3450.299/0.3530.303/0.3520.298/0.3500.299/0.3500.296/0.3480.298/0.349
1920.332/0.3700.337/0.3740.343/0.3760.346/0.3780.345/0.3780.344/0.3790.343/0.377
3360.369/0.3950.369/0.3970.370/0.3960.370/0.3960.370/0.3960.367/0.3950.370/0.393
7200.422/0.4260.435/0.4320.425/0.4290.434/0.4330.431/0.4320.426/0.4290.431/0.431
ETTm2960.168/0.2580.171/0.2550.172/0.2560.170/0.2540.171/0.2550.171/0.2560.170/0.256
1920.228/0.2980.230/0.2960.231/0.2970.228/0.2950.229/0.2950.229/0.2950.234/0.297
3360.280/0.3300.288/0.3350.288/0.3340.287/0.3340.284/0.3310.285/0.3320.287/0.334
7200.357/0.3880.383/0.3930.382/0.3920.383/0.3940.382/0.3920.382/0.3910.380/0.388
Electricity960.135/0.2310.145/0.2460.145/0.2460.144/0.2440.143/0.2440.144/0.2450.142/0.241
1920.147/0.2430.160/0.2600.161/0.2610.160/0.2590.159/0.2580.160/0.2590.160/0.258
3360.164/0.2660.178/0.2770.178/0.2770.177/0.2750.177/0.2750.177/0.2760.177/0.277
7200.208/0.2960.215/0.3060.215/0.3060.214/0.3040.214/0.3040.215/0.3060.213/0.304
Weather960.151/0.1980.163/0.2090.162/0.2070.161/0.2070.161/0.2070.161/0.2060.161/0.206
1920.196/0.2410.205/0.2470.205/0.2460.203/0.2450.203/0.2450.201/0.2440.206/0.247
3360.237/0.2750.262/0.2900.260/0.2890.259/0.2870.260/0.2870.259/0.2880.261/0.290
7200.311/0.3280.336/0.3410.336/0.3400.336/0.3420.336/0.3400.335/0.3390.337/0.341
Traffic960.361/0.2610.404/0.2840.403/0.2840.400/0.2800.399/0.2810.401/0.2820.403/0.288
1920.370/0.2710.425/0.2910.422/0.2890.418/0.2850.416/0.2850.419/0.2850.418/0.289
3360.388/0.2790.443/0.2990.448/0.3040.440/0.2980.437/0.2960.445/0.3000.436/0.296
7200.427/0.2920.472/0.3180.474/0.3220.475/0.3200.473/0.3210.475/0.3230.466/0.315
Exchange960.082/0.2030.083/0.2020.084/0.2030.084/0.2020.082/0.2030.085/0.2030.103/0.223
1920.172/0.2960.184/0.3070.182/0.3050.184/0.3070.183/0.3060.185/0.3090.184/0.306
3360.307/0.4030.329/0.4190.332/0.4210.325/0.4170.329/0.4200.328/0.4190.365/0.440
7200.792/0.6660.805/0.6700.892/0.7090.789/0.6620.915/0.7160.902/0.7110.943/0.727
Average0.309/0.3380.330/0.3490.334/0.3510.329/0.3480.334/0.3500.333/0.3500.335/0.351

TS-CAN has lower MSE than the control in 189 of the 192 comparisons. The three exceptions are all within 0.6%. Averaged over the 32 settings, the controls are 6.4%–8.3% worse in MSE and 3.0%–4.0% worse in MAE.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank all the reviewers for their careful reviews and constructive comments. We have revised the paper accordingly, and all updates are highlighted in blue

The reviewers generally held positive opinions of our work. They found the idea "novel" and "interesting", noting that "the symmetric and antisymmetric terms offer a clear way to think about directional information" and that "the links to Lévy area and spectral analysis help motivate the design". They described the results as "strong" and the evaluation as "solid", with particularly substantial gains on Traffic and Exchange, and found the paper "well-written". They also found the attention-free design and the reported inference efficiency "appealing", and acknowledged that our component ablations, parameter-matched mixers, surrogate-data tests, and direct delay-detection experiments "go beyond reporting benchmark accuracy".

To address the reviewers' concerns, we made a substantial effort during this rebuttal period and ran 744 supplementary experiments in total. These include four new baselines (CliffordNet-1D, a Clifford Transformer, ARMD, and NsDiff), each evaluated on all 32 long-term settings. They also include six activation- and capacity-matched antisymmetry controls over 8 datasets × 4 horizons × 3 seeds, separate temporal- and channel-wedge ablations on all 32 settings, and a new efficiency comparison including TimeMixer++. All experiments follow the same protocol as the main results, and every new result is reported in full in the individual responses.

The reviewers also raised constructive concerns. The major revisions are summarized below, and details are given in the individual responses.

  • Additional baselines: We added CliffordNet-1D, a Clifford Transformer, and two diffusion-based models (ARMD and NsDiff), all evaluated on the 32 long-term settings.
  • Isolating antisymmetry: We added controls matched to TS-CAN in activation, context, and, where applicable, parameters and FLOPs, evaluated on 8 datasets and 4 horizons.
  • Broader ablations: We extended the temporal- and channel-wedge ablations to all 8 datasets and 4 horizons, including Traffic and Exchange.
  • Theory and reproducibility: We clarified the algebraic properties of the layer, strengthened the theoretical justification (App. C.6), and added a complete forward-pass algorithm (App. D).
  • Efficiency and presentation: We added TimeMixer++ to the efficiency comparison, revised the abstract, and updated the references.

We would be happy to answer any further questions and look forward to the reviewers' feedback.

Area_Chair_2LUmOpenReview ↗

Invitation to Discuss the Authors’ Response

Dear Reviewers,

The authors have now provided their response. I encourage all reviewers to read it and participate in the discussion with the authors to clarify any remaining questions or concerns.

Thank you to both the authors and reviewers for your time and effort throughout the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.