- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes PolyCtrl, a long-term forecaster that represents the future as a history-controlled composition of simple polynomial functions. Each channel has a bank of K learnable polynomials of degree P−1 defined over the normalized horizon. A controller shared across channels (Linear–GELU–Linear–tanh) maps each channel's RevIN-normalized history to a K×H control matrix G_c, and the forecast is Σ_k G_{c,k,h}·φ_{c,k}(z_h). A Dual variant applies two such branches to moving-average trend and seasonal components. On eight LTSF benchmarks against five baselines, Dual PolyCtrl reports the best average MSE (0.314) and PolyCtrl the best average MAE (0.347). The authors further report a "capacity saturation" pattern in K and P, and show that forecasts rely on distributed contributions from many functions.
The model is very simple and easy to reproduce, and K and P are fixed across all datasets. The parameter accounting is transparent: Eq. 11 reproduces the reported 9.35M parameters, and the averages in Table 1 and Table 5 agree. The paper also provides extensive visualizations of modes, gates, and contributions for every dataset.
However, the central claim that forecasts are synthesized by controlling polynomial functions is not supported by the formulation or by the experiments. Because the control signal is produced independently for every (k, h), the model is effectively an MLP whose output layer is reparameterized by the polynomial bank, and no ablation separates the polynomial or control components from a plain MLP. The capacity-saturation analysis is a single-run, single-setting test-set sweep that also confounds K with parameter count, and it was used to choose K and P. The interpretability evidence is weak. The empirical comparison is narrow, with small margins, unstated lookback and baseline tuning, and no variance reporting.
Critical strengths
- Function-control view of forecasting: The paper asks where forecasting complexity should sit, and proposes splitting a forecaster into a small bank of explicit functions defined on the future horizon and a history-driven controller that weights each function at each future position (Eq. 4–6). Unlike N-BEATS, where one coefficient per basis applies to the whole horizon, the same function can contribute differently at different positions. Framing a forecaster as "which functions, controlled how" is a clean design principle, and it is a reasonable direction given the recent success of simple linear and MLP forecasters.
- Number versus complexity of functions as an explicit question: Because K (how many functions) and P (how complex each function is) are separate knobs, the paper can pose a concrete question about capacity allocation: is it better to have more simple functions or fewer expressive ones? The early evidence (Fig. 3) that raising P does not make up for a small K is an interesting starting point, even though, as discussed below, the current experiment cannot yet answer the question.
Critical weaknesses
- The method does not leave the representation-and-decoding paradigm it argues against: The abstract and introduction motivate PolyCtrl as letting history "directly determine how the future should be formed" instead of encoding it into a representation that is then decoded (p.1–2), and claim that complexity "no longer needs to be carried by a single complicated future generator" (p.2). In the formulation, however, history is encoded into e(x) = GELU(W_e x), and the forecast is decoded by a linear layer with tanh whose outputs are weighted by fixed horizon profiles. This is a channel-independent MLP with a reparameterized readout.
- From Eq. 4–6, Ŷ_h = Σ_k Φ_{k,h}·tanh(w_{k,h}ᵀe(x) + b_{k,h}), with an independent row of W_g for every (k, h). The control can vary freely along h, so the forecast is not restricted to any polynomial span, and Φ only sets a smooth amplitude envelope (|Ŷ_h| ≤ Σ_k|Φ_{k,h}|).
- The paper's own evidence says the polynomials do little. App. A.2 (Eq. 26–28) shows rank(C) = rank(G) even when all functions are identical, Fig. 3(c) shows a linear basis (P≈2) is only about 0.003 MSE worse than P=8, and the controller holds about 9.28M of the 9.35M parameters while the banks hold 64.2K (Sec. 4.2). The complexity is therefore in the generator, not in the functions.
- No ablation separates the claimed mechanism from a plain MLP (L→D→H with the same RevIN and D), from Φ fixed to a constant or replaced by a free K×H matrix, or from horizon-shared control. Without these, the central claim of the title is untested.
- The two "findings" offered as support do not test the claims they are used for: The abstract presents capacity saturation and distributed contributions as evidence that future generation "may be considerably simpler than commonly assumed," but both are what one would expect from the MLP reading above.
- Capacity saturation: by Eq. 11, increasing K also grows the controller output layer (KH(D+1) parameters), so the K curve is confounded with width and parameter count. Fig. 3 covers only ETTh1 at H=96 with a single run (the P curve moves within 0.376–0.382), and the curves are labeled "Test MSE" and used to choose K=25 and P=8 (Sec. 4.3), which is model selection on the test set.
- Distributed contributions (Table 2, Fig. 4(e)–(f)) are the expected behavior of a K-unit readout without any sparsity pressure. In Fig. 4(g), the controlled trajectory g·φ is a jagged, near-zero signal that does not resemble the smooth φ_k, so the individual "functions" are not what the forecast is built from in any interpretable sense. The reported averages (18.85, 21.43) also cannot be recovered from Table 2 (simple means 17.64, 20.48).
- The competitive results cannot be credited to the formulation: The remaining support for the idea is benchmark accuracy, but the margins are small and the closest comparisons are missing.
- The average MSE gap is 0.314 vs. 0.319 (TimeXer) and 0.320 (SOFTS) with no seeds or standard deviations reported, the lookback length is never stated (it can only be back-solved as L=96 from Eq. 11), and it is unclear whether baselines were rerun or tuned.
- Given the MLP equivalence, the decisive baselines would be RLinear/RMLP, TiDE, and N-BEATS/N-HiTS, none of which are included. The channel-independent design also trails on high-dimensional data (Electricity 0.188 vs. 0.177, Solar-H 0.218 vs. 0.197; Table 1).
Questions for authors
- MLP baseline: How does PolyCtrl compare with a plain channel-independent MLP (Linear L→D, GELU, Linear D→H) using the same RevIN, D=512, and training setup? Is the difference significant across seeds?
- Role of the polynomial bank: What happens when Φ is fixed to a constant, replaced by a free K×H matrix, or shared across channels? What happens when G is made horizon-shared (G_{k,h} = g_k(x)), as in N-BEATS? These results would directly test the "polynomial" and "horizon-wise control" claims.
- Model selection: Were K and P chosen on the validation set or from the test curves in Fig. 3? Does the saturation pattern hold at matched parameter count (e.g., trading K against D) and on other datasets and horizons?
- Protocol: What lookback length was used for PolyCtrl and for each baseline? Were baselines rerun under the Table 4 configuration or tuned individually? How many seeds were used, and what are the standard deviations?
- Table 2: How exactly are 18.85 and 21.43 computed, and what is n in "n-weighted"?
2 comments
Thank you for your detailed follow-up. We would like to clarify these points directly:
-
Linear Basis vs. Higher Degrees (): Concluding from Figure 3(c) that linear bases are inherently sufficient a priori reverses cause and effect. Rather than linearity being universally enough, the plateau after is an emergent outcome of gradient optimization: the network itself discovers that linear capacity suffices on the dataset shown in Figure 3(c) and automatically suppresses unnecessary higher-order terms within the nested polynomial spaces (, Appendix A.1, Eq. 20). Crucially, we cannot assume that unseen datasets beyond the eight standard conference benchmarks will never require higher polynomial orders—retaining provides necessary structural redundancy without manual order tuning.
-
Clarification on Section 4.2 (Shared vs. Channel-Specific Bank): We apologize for the unclear reference: Section 4.2 discusses the parameter scaling of channel-specific vs. shared banks rather than reporting a separate shared-bank ablation. While using a larger shared bank is feasible in principle without theoretical issues, accommodating heterogeneous channels with a single shared bank generally calls for a larger bank width , which increases the controller's complexity and output dimension—under Eq. (11) (), even in adds 369.4K parameters at . By contrast, assigning channel-specific polynomial banks adds only parameters per channel (4.2K on Weather and 64.2K on 321-channel Electricity, of total parameters). Following Occam's razor ("entities should not be multiplied beyond necessity"), we naturally adopt this more economical parameterization.
-
Parameter Savings Against Free Matrix: By the same principle of eliminating unnecessary parameters whenever accuracy is preserved, removing a horizon-scaling redundancy in the target bank is worthwhile even on smaller- datasets—we do not believe modest savings should be dismissed. Moreover, the 3.9% saving only reflects low-dimensional Weather (); under Eq. (11) on higher-dimensional Electricity (), polynomials require only 64.2K bank parameters versus 5.78M for a Free matrix—saving 5.71M parameters (~38% of total model size).
-
On "Structured-Readout MLP" and Escaping the Polynomial Span: Focusing exclusively on raw algebraic equivalence—reducing PolyCtrl to "an MLP with a structured readout" or noting that gating escapes the span of polynomials—overlooks the substantive value brought by this decoupled architecture. First, escaping the static polynomial span via horizon-wise control is not an accidental loophole, but the exact mechanism we explicitly analyzed in Appendix A.2 (Eqs. 26–29) to enable simple substrates to synthesize complex futures. Second, how a model is factorized and decoupled matters: reducing our framework to a flat parameterized MLP misses the explicit control interpretability (inspecting how modulates in Fig. 4 and Appendix C.2), optimization stability (bounded coupling yielding – lower variance in Table 6), and architectural extensibility that this separation enables.
Finally, our core contribution centers on decoupling active generative control from the controlled object rather than polynomial shapes alone; as we have repeatedly emphasized in our rebuttal, polynomials merely serve as a lightweight realization of the controlled object (which, as an indispensable half of the framework, naturally appears in the title). Consistent with our outlook in Section 5, this control paradigm holds strong potential for further development beyond polynomials: for example, by parameterizing a linear forecaster (like DLinear) as low-rank temporal mode matrices (rank ) modulated by a controller, one directly obtains a dynamic DLinear with both process interpretability (via mode control) and outcome interpretability (via the assembled linear mapping). In the final version, we will gladly sharpen our framing around the controller and readout design as you suggested.
Response to the Authors
Thank you for the new experiments with three seeds and for clarifying L=96 and the Table 2 averaging. These address Q4 and Q5.
However, Table 7 supports my reading that PolyCtrl is an MLP with a structured readout:
Correction on my side: In Fig. 3(c), the point about 0.004 worse than P=8 is P=1 (constant), not P~2. A linear basis is already on the plateau, so the polynomial shape beyond a linear term does not matter in this experiment.
I will keep my score and suggest reframing the contribution around the controller and readout design.