ConFlux: Multivariate Time Series in Flux, One Unified Forecast in Confluence

Summary

ConFlux proposes a foundation model for multivariate time-series forecasting whose central idea is to reorganize variables rather than time. For every batch the model computes five pairwise association matrices (linear, non-linear, trend, periodic, anomaly co-occurrence), aggregates them into a weighted graph, and uses the sign pattern of the Fiedler vector of that graph's Laplacian to sort variables so that highly related channels sit next to each other. Adjacent variables are then averaged into "variate patches," passed through a standard pre-LN Transformer encoder, and decoded to future values through a Fourier-basis reconstructor that in principle supports arbitrary output horizons without retraining. The conceptual move — treating the variate axis, rather than the time axis, as the axis for tokenization and patching — is the paper's main methodological contribution and is well motivated: attention scales as $O(N^2)$ with $N = \lceil C/P \rceil$ patches, so a multivariate foundation model can accept many more channels than a per-variable-token architecture. The empirical program is broad: zero-shot, in-domain fine-tuning, and from-scratch protocols across standard ETT/Weather/Solar/Electricity benchmarks, PEMS traffic, and 14 high-dimensional datasets. Reported MSE/MAE consistently improves over Chronos-2, Timer-XL, and specialized baselines, and inference-efficiency curves versus attention-based rivals are strong.

The main limitation is that the manuscript's evidence outruns its specification. Load-bearing details (the dimensionality at which self-attention actually operates, whether the five association matrices are summed or weight-summed, how a fixed-frequency DFT is combined with "learnable" $\omega_k$, how "unifying to exactly 1000 variables" is implemented, the fine-tune and from-scratch protocols) are inconsistent within the paper or omitted. Headline percentages (12.5%, 29.7%, 30.7%) are computed under an unstated averaging rule and are heavily influenced by single outlier datasets (Solar, in particular). Several baselines are mischaracterized, multiple figure/text cross-references are wrong, and no seed variance is reported anywhere. Crucially, the pretraining corpus (LargeST) overlaps in provenance with the PEMS "zero-shot" benchmarks; this potential contamination is not audited, and it materially affects how the empirical case should be read.

Strengths

• Conceptual contribution. Reframing the tokenization axis from time to variable, and then patching along that axis, is a genuinely new lens on foundation-model architecture for multivariate forecasting. The framing generalizes cleanly to arbitrary $C$ via variate patching.
• Principled ordering with a theoretical hook. Deriving the variate ordering as the optimal rounding of a spectral relaxation of a weighted-position-cost problem (Appendix B, Theorem B.1) gives the sorting step a defensible interpretation rather than being a heuristic.
• Breadth of empirical evaluation. Zero-shot, fine-tuning, from-scratch, and high-dimensional protocols are all reported, with per-horizon breakdowns in the appendix. The paper compares against a wide range of foundation and specialized baselines.
• Efficiency story is architecturally motivated. Reducing attention from $O(C^2)$ over $C$ variate tokens to $O(N^2)$ over $N = \lceil C/P \rceil$ patches, then decoupling forecast length from the decoder via a Fourier-basis reconstructor, is a clean design that scales to thousands of variables without OOM (Table 8).
• Ablation isolates the two mechanisms. Table 3 separately removes sorting, training-time patching, and inference-time patching, giving evidence that both mechanisms contribute rather than only one.

Weaknesses

• Attention dimensions contradict the complexity claim. In Section 4.3.1 the attention weights and encoder output are labelled $R^{C\times C}$ and $R^{C\times d}$, but the very next paragraph claims $O(N^2)$ complexity over $N \ll C$ patches. The reader cannot tell at which granularity attention actually runs; this is a central architectural claim, so the notation must be fixed.
• Fourier-basis decoder is under-specified. The reconstruction equation combining $\text{Re}(F_{c,k})$ and $\text{Im}(F_{c,k})$ is not rendered, $\omega_k$ is described as both "the discrete angular frequency of bin $k$" and "learnable" (which are incompatible), and the DFT is said to be applied "along the temporal dimension" of $\tilde Z \in R^{C\times d}$ — a representation whose second axis is a latent dimension, not time. The arbitrary-horizon story rests on this decoder; the derivation should be spelled out.
• Training-corpus overlap with "zero-shot" PEMS. The pretraining corpus includes LargeST, a large California PeMS traffic dataset; the "zero-shot" short-term evaluation uses PEMS03/04/07/08, which are also PeMS subsets. The paper does not audit sensor or time-window overlap, so the strong PEMS gains cannot be cleanly attributed to zero-shot transfer.
• Headline percentages driven by an unstated averaging rule and outliers. The 30.7% Timer-XL reduction is a ratio of averaged MSE, heavily inflated by Solar (0.148 vs 0.601). The 27.1% ETTm2 "gain over its performance" does not match either the zero-shot→fine-tune delta (~15%) or the ConFlux-vs-Chronos-2 delta (~32%). Averaging conventions need to be stated and the outlier disclosed.
• No seed variance or significance testing. Every reported number is a single-run point estimate, with margins often only a few thousandths of MSE. The reader cannot judge whether the reported gaps are meaningful.
• Missing decoder ablation. The three named contributions are sorting, patching, and the Fourier-basis decoder, but only the first two are ablated. Whether the sinusoidal-basis design is doing real work versus simply providing a large output head is unresolved.
• Quantile-trained but not quantile-evaluated. The loss is a 21-quantile pinball loss, yet every reported metric is a point-forecast metric. CRPS, per-quantile loss, or coverage of the reported quantiles are absent.
• Baseline descriptions are inaccurate. Crossformer is described as "built on a auto-correlation mechanism" (Autoformer's mechanism, not Crossformer's) and TimesNet as a "frequency-domain … model," which mischaracterizes an FFT-driven period-detector plus 2D time-domain convolutions.
• Figure/text/label inconsistencies in the analysis appendix. Appendix J's text describes Figure 7a and 7b as ETTm2 patch-size sweeps, but the panel labels read "Atec" and "Solar"; the recommended patch size (16) contradicts the reported optimum at 8; and the inference efficiency section references Figure 5 when the intended figure is Figure 8, quotes latency in "µs" against an axis labelled "seconds," and reports memory "in MB" against an axis labelled "GB".
• Internal claim/data mismatches. In Table 4 Chronos beats ConFlux on Electricity even though the text calls ConFlux "the top performer" after fine-tuning. In Table 9 ConFlux is not best at horizon 720 on ETTh2, contra "the advantage of ConFlux is particularly pronounced in long-term forecasting." In Table 13, PEMS04/07 Avg entries appear lower than the mean of the per-horizon rows.
• Attribution error under from-scratch. In Appendix O, the from-scratch high-dimensional gains are credited to "our choice to include high-dimensional series in the training corpus," but that corpus is not seen in the from-scratch regime.
• Under-specified pretraining scope. For a foundation-model claim, the corpus (six real datasets + KernelSynth) is unusually small and token counts are not reported, making it hard to judge whether "zero-shot" performance is truly cross-domain transfer or partial in-family transfer.
• Notation collisions. $A$ is reused for the aggregated association matrix and for the attention matrix, both in $R^{C\times C}$; $L$ is reused for the context length and for the graph Laplacian within the same paragraph.
• Duplicate PatchTST reference. The bibliography lists PatchTST twice (single-author 2022 arXiv, four-author 2023 ICLR), and in-text citations flip between the two forms.

Reproducibility & code

Veritas inspected the released materials in "paper + code (read, not executed)" mode. The manuscript contains no code-availability statement, and the released material shipped with the submission does not include a runnable codebase, so reproducibility currently rests entirely on the paper text. Read against that text, the following gaps stand out:

• Aggregation of the five association matrices. Section 4.1 defines $A$ as a plain sum, Appendix A instead calls it "weighted aggregation." If per-criterion weights exist, their values (and whether they are learned) are never given; if they don't, the language should be reconciled. Either way, an independent implementation cannot reproduce the Solar heat-map in Figure 6 (or the Fiedler-vector ordering) numerically.
• Association-measure hyperparameters. The MI binning strategy, anomaly $z$-threshold, trend moving-average window, and the periodic-similarity kernel bandwidth $\sigma_p$ / "dominant period" selection rule are not stated. These directly drive the sort.
• The "unify to exactly 1000 variables" preprocessing. The training pipeline forces every corpus dataset to 1000 channels via "a variable patching step," but padding/truncation/replication behaviour for datasets with fewer (ETTm2, 7 channels) or many more (Wiki-20k, 20,000 channels) variables is not documented. This is central to the pretraining recipe.
• Fine-tune protocol. Section 5.3 refers only to "a consistent downstream protocol." Learning rate, epochs, batch size, and which layers are unfrozen are not stated. The 30.7% Timer-XL reduction and the "top performer after fine-tuning" claim rest on this unspecified protocol.
• From-scratch protocol. Optimizer, LR schedule, epochs, and batch size for the target-dataset training runs (Tables 6, 7, 12, 13) are not stated. Applying the pretraining recipe verbatim to small benchmarks like ETTh1 will not work.
• Efficiency benchmark harness. The 6-unit latency and <1400 MB memory readings at 1024 variables (Appendix L / Figure 8) depend on batch size, context length, warmup, precision, and hardware, none of which are stated; the axis units disagree with the text ("MB" vs the plot's "GB"; "microseconds" vs the plot's "seconds").
• No seeds, no error bars. Twenty or more model configurations (four sizes × five data fractions in Figure 4; three ablations in Table 3; five training-time patch sizes in Figure 7a; six lookback lengths in Figure 7c) are reported as single points, which precludes any variance-based reproducibility check.

None of the veritas assessments could confirm any specific numeric value from execution (code was not run). What we can say is that even a well-resourced re-implementation is unlikely to match the reported per-cell values within a few thousandths of MSE without a released model or a complete pretraining recipe.

Recommended Changes

Essential

• Fix the attention-dimension notation in Section 4.3.1 so that self-attention operates on $N$ patch tokens ($N\times N$ attention weights, $N\times d$ output) and the $O(N^2)$ complexity claim is consistent.
• Give the Fourier-basis decoder equation explicitly and reconcile "learnable $\omega_k$" with "the discrete angular frequency of bin $k$"; describe over which axis of $\tilde Z \in R^{C\times d}$ the DFT is actually taken.
• Audit the LargeST / PEMS overlap and either provide sensor- and time-window-level evidence that the "zero-shot" PEMS setting is out-of-distribution or move PEMS out of the zero-shot claim.
• State the exact aggregation rule for $A$, including the weights on each of the five association matrices and the normalization ("weighted aggregation" in Appendix A vs summation in the main text).
• Specify the "1000-variable" preprocessing. How are datasets with fewer or more channels reshaped, padded, replicated, or truncated?
• Add the fine-tune and from-scratch protocols (LR, epochs, batch size, layers unfrozen, optimizer) as an appendix subsection.
• Report seed variance (at least mean ± std over 3 seeds) for the Base configuration on the headline zero-shot and fine-tune tables; state the averaging rule for the 12.5%, 29.7%, and 30.7% figures and disclose the Solar outlier in the 30.7% Timer-XL number.
• Release code / model weights or state the release plan; otherwise the paper's reproducibility rests entirely on prose that currently has gaps.

Suggested

• Add a Fourier-basis-decoder ablation (e.g. against a linear head or a shared MLP at fixed horizons) to isolate the decoder's contribution from raw output-layer capacity.
• Compare Fiedler-vector sorting to simpler orderings (Pearson-only, hierarchical-clustering leaf order, random) so the ~2–5% "w/o Sort → full" gap can be attributed to the specific spectral construction.
• Add probabilistic evaluation (CRPS, quantile pinball loss, calibration) commensurate with the 21-quantile training objective.
• Clarify the "27.1% gain on ETTm2" by naming the baseline it is computed against, or drop the number.
• Soften "top performer after fine-tuning" to reflect the Electricity row in Table 4, and soften "advantage particularly pronounced in long-term forecasting" to reflect the ETTh2 720 row in Table 9.
• Correct the Crossformer and TimesNet descriptions in Appendix E.
• Fix Figure 7 panel labels (Atec/Solar/ETTm2) so they agree with the prose in Appendix J, resolve the P=8 optimum vs P=16 recommendation, and reconcile the µs/s and MB/GB units and the Figure 5/Figure 8 cross-reference in Appendix L.
• Report token counts and per-domain sizes for the pretraining corpus so readers can compare it to Chronos, MOIRAI, and Timer-XL, and add a footnote or reason for the Time-MoE "—" entries on Electricity.
• Merge the duplicate PatchTST bibliography entry and use a single in-text citation form.