acceptodds
Under review as a conference paper at ICLR 2027

Time Series Forecasting May Be as Simple as Controlling Polynomials

Abstract

Time series forecasting is commonly formulated as a representation-and-decoding problem, where historical observations are encoded into latent representations and future sequences are subsequently generated from them. In this paradigm, history is primarily treated as a source of information to be encoded and extracted. Modern architectures have made historical representation increasingly adaptive, selective, and expressive. Yet history is still primarily processed to obtain a representation from which the future is subsequently predicted. We ask whether history can play a more direct role: instead of merely being represented, can history actively govern how the future is formed? Based on this perspective, we learn a bank of polynomial predictive functions directly over the future horizon. A lightweight controller converts history into horizon-wise control signals that dynamically modulate their contributions. Complex future dynamics are then synthesized through the history-conditioned coordination of these simple functions. We refer to this model as PolyCtrl, reflecting its history-driven control of polynomial predictive functions. Extensive experiments on eight real-world datasets across multiple forecasting horizons demonstrate that this simple formulation achieves highly competitive forecasting performance. Further analysis reveals a consistent capacity-saturation phenomenon. Once the polynomial order and the size of the predictive-function bank provide sufficient representational capacity, further increasing either of them yields only marginal improvements. Accurate forecasts also rarely depend on only a few dominant functions. Instead, they typically emerge from the coordinated contributions of multiple simple polynomial predictive functions. These findings suggest that the future-generation mechanism required for strong time series forecasting may be considerably simpler than commonly assumed. With a sufficiently rich collection of simple predictive functions and history-driven dynamic control, complex future dynamics can already be effectively modeled.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Polynomial Predictive Functions, History-Driven Control, Dynamic Function Composition, Predictive Function Control

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_igadOpenReview ↗
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
Reviewer_igadOpenReview ↗

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:

  • Free KxH without tanh is mathematically a channel-specific linear readout on the MLP encoder. Relative to Plain MLP (0.3557), the gain comes from the channel-specific readout (0.3493) and the tanh hidden layer (Free KxH, 0.3440). Adding the polynomial bank gives 0.3450, so it adds no measurable gain on top of Free KxH. On Weather, the best variant is Free KxH without tanh (0.2462).
  • Horizon-shared control, which keeps predictions in the polynomial span, collapses (0.4480). The model works because per-(k,h) control escapes the polynomial span, not because the future is a combination of polynomials. The response itself says polynomials were chosen "solely" for parameter cost, which saves about 3.9% at Weather H=720.
  • Plain MLP has 13-22x fewer parameters and a channel-shared head, so it is not an equal-capacity comparison. The response cites Sec. 4.2 for shared banks achieving similar results, but the revised Sec. 4.2 contains no such result.

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.

Thank you for your detailed follow-up. We would like to clarify these points directly:

  1. Linear Basis vs. Higher Degrees (PP): 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 P=2P=2 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 (Π0⊂Π1⊂⋯⊂ΠP−1\Pi_0 \subset \Pi_1 \subset \dots \subset \Pi_{P-1}, 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 P=8P=8 provides necessary structural redundancy without manual order tuning.

  2. 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 KK, which increases the controller's complexity and output dimension—under Eq. (11) (NPolyCtrl=CKP+D(L+1)+KH(D+1)N_{\text{PolyCtrl}} = CKP + D(L+1) + KH(D+1)), even +1+1 in KK adds 369.4K parameters at H=720,D=512H=720, D=512. By contrast, assigning channel-specific polynomial banks adds only KP=200KP=200 parameters per channel (4.2K on Weather and 64.2K on 321-channel Electricity, <1%<1\% of total parameters). Following Occam's razor ("entities should not be multiplied beyond necessity"), we naturally adopt this more economical parameterization.

  3. Parameter Savings Against Free K×HK \times H Matrix: By the same principle of eliminating unnecessary parameters whenever accuracy is preserved, removing a 90×90\times horizon-scaling redundancy in the target bank is worthwhile even on smaller-CC datasets—we do not believe modest savings should be dismissed. Moreover, the 3.9% saving only reflects low-dimensional Weather (C=21C=21); under Eq. (11) on higher-dimensional Electricity (C=321,H=720C=321, H=720), polynomials require only 64.2K bank parameters versus 5.78M for a Free K×HK \times H matrix—saving 5.71M parameters (~38% of total model size).

  4. 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 K×HK \times H gating escapes the span of polynomials—overlooks the substantive value brought by this decoupled architecture. First, escaping the static polynomial span ΠP−1\Pi_{P-1} 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 gc,k,hg_{c,k,h} modulates ϕc,k(zh)\phi_{c,k}(z_h) in Fig. 4 and Appendix C.2), optimization stability (bounded (−1,1)(-1,1) coupling yielding 10×10\times–50×50\times 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 KK low-rank temporal mode matrices (rank R∈{1,2,3}R \in \{1,2,3\}) modulated by a K×H×RK \times H \times R 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.

Reviewer_DB4rOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the problem of future sequence generation in time-series forecasting. The authors argue that existing methods primarily rely on historical information encoding and predictive decoding while overlooking the direct role of historical information in controlling future predictive functions. To address this issue, the paper proposes PolyCtrl, which constructs a learnable bank of polynomial predictive functions and uses historical information to generate dynamic control weights. These weights are then used to combine different polynomial functions to produce future predictions.

Critical strengths

This paper attempts to rethink time-series forecasting from the perspective of future sequence generation, shifting the research focus from complex historical representation learning to the dynamic control of predictive functions using historical information. This idea has certain research significance and offers an alternative perspective for developing simple yet effective forecasting models.

Critical weaknesses

W1: According to the formulation in Section 3, PolyCtrl uses an MLP controller to generate a K×H-dimensional weight matrix, which is then multiplied element-wise with the polynomial function bank and summed to produce the forecasts. However, the paper lacks sufficient comparisons with directly predicting future values using an MLP, removing the polynomial function bank, or replacing it with simple constant basis functions. Since the controller itself can already generate independent outputs for different future positions, these comparisons are essential to demonstrate whether the polynomial function bank provides additional predictive capability. Currently, it remains unclear whether the performance improvements mainly arise from polynomial function modeling or the nonlinear mapping capability of the controller itself.

W2: Figure 3 shows that increasing the number of predictive functions K significantly improves forecasting performance, which the authors use to emphasize the importance of coordinating multiple predictive functions. However, according to Equation (11), the controller's parameter count includes the term KH(D+1). Therefore, increasing K not only increases the number of controllable functions but also directly increases the model capacity. The current experiments lack parameter-matched comparisons, making it difficult to rule out the possibility that the performance improvements result from additional parameters. Consequently, the necessity of the multi-function coordination mechanism has not been sufficiently supported by rigorous experimental evidence.

W3: The paper lacks direct comparisons with Fourier basis functions, fixed basis functions, and other simple function families. Therefore, the rationale for choosing polynomials as the core predictive functions has not been sufficiently validated.

W4: The paper mainly compares the proposed method with five baselines: TimeXer, DLinear, iTransformer, CMoS, and SOFTS. However, it does not include N-BEATS, N-HiTS, or other methods directly related to basis expansion and structured future forecasting. In particular, N-BEATS already employs polynomial basis functions for trend forecasting, making it highly relevant to the proposed approach. Furthermore, the paper lacks comparisons with more recent forecasting methods published in 2025–2026. Therefore, the current experiments are insufficient to convincingly demonstrate the performance advantages of PolyCtrl over existing basis-function-based forecasting approaches and recent state-of-the-art models.

W5: The experimental figures and tables in the paper still have considerable room for improvement in terms of formatting and visual presentation. I suggest that the authors increase the font sizes in the figures and use fonts consistent with the main text to improve the overall visual consistency, readability, and professionalism of the paper.

Questions for authors

Could the authors provide the complete source code, experimental scripts, and relevant hyperparameter configurations? These materials would help improve the transparency and reproducibility of the experiments and allow researchers to further verify the effectiveness of the proposed method.

Reviewer_qmoaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PolyCtrl for time series forecasting. It uses a bank of learnable polynomial functions and an MLP controller that assigns history-dependent weights to each function at every future time step. The authors also introduce Dual PolyCtrl, which models trend and seasonal components separately. Experiments on eight datasets show competitive results against existing methods. The paper further studies model efficiency, the effects of polynomial order and function-bank size, and the contributions of individual functions to the predictions.

Critical strengths
  1. Simple and clear model design. PolyCtrl combines learnable polynomial functions with history-dependent control signals. The formulation is straightforward, and the contribution of each function can be examined separately.

  2. Competitive forecasting results. The authors evaluate the method on eight datasets and four prediction horizons. Both PolyCtrl variants achieve competitive performance against several established forecasting models, particularly on the ETT benchmarks.

Critical weaknesses
  1. Limited novelty over existing basis-based forecasting methods. N-BEATS already uses polynomial bases for forecasting, while BasisFormer learns predictive bases and combines them using history-dependent coefficients. PolyCtrl introduces horizon-wise control, but its distinction from these earlier approaches is not sufficiently established. The paper lacks direct comparisons and controlled experiments showing the specific benefit of this design. This makes the main methodological contribution less convincing.

  2. Insufficient evidence for the proposed mechanism. The MLP controller already produces separate control signals for each future time step. Moreover, increasing the number of polynomial functions K also substantially increases the controller's parameter count, as shown in Eq. (11). Therefore, the sensitivity analysis cannot distinguish the benefit of polynomial functions from increased model capacity. The paper lacks parameter-matched MLP comparisons, while the Top-m analysis only measures reconstruction of the model's own predictions rather than forecasting accuracy. These limitations weaken the paper's central conclusions.

Questions for authors
  1. What is the main methodological difference between PolyCtrl and existing basis-based forecasting methods such as N-BEATS and BasisFormer? Can the authors provide direct comparisons that isolate the benefit of horizon-wise control?

  2. According to Eq. (11), increasing K also increases the controller's parameter count. Can the authors compare PolyCtrl with a parameter-matched MLP prediction head to show that the improvement comes from the polynomial structure rather than increased model capacity?

Comments on the paper

1 comment

Thank you for your thoughtful follow-up. Here are our concise clarifications on the remaining points:

1. Methodological Novelty: Decoupling the controller from the controlled object conceptually and architecturally makes the prediction generation process transparent and traceable. For example, when combined with DLinear, this separation yields a dynamic linear model with higher interpretability, which we believe offers valuable inspiration for future designs.

2. Experimental Evidence & Role of the Baselines:

  • Plain MLP vs. Capacity-Matched Comparison: We would like to clarify that our Plain MLP (L→D→HL \to D \to H) serves as a standard black-box direct-decoding baseline and indeed has fewer parameters than PolyCtrl's L→D→KHL \to D \to KH controller. However, from the perspective of our proposed conceptual separation (how to control vs. what to control), simply widening a black-box MLP to match PolyCtrl's parameter count still yields an unanalyzable direct mapping and is not the most meaningful control for isolating our decoupled design. Instead, the true capacity-matched ablation that shares the exact same L→D→KHL \to D \to KH controller budget is Free K×HK \times H in Table 7.
  • Polynomial vs. Free K×HK \times H Matrix (0.3440 vs. 0.3450): Achieving virtually identical accuracy to an unconstrained Free K×HK \times H matrix highlights the superior parameter efficiency of polynomials—eliminating horizon-scaling redundancy in the target bank (e.g., saving 373.8K parameters on Weather at H=720H=720, and 5.71M / ~38% parameters on 321-channel Electricity under Eq. 11) while allowing flexible per-channel bases.

3. Overall Positioning & Contribution: Finally, we do not claim SOTA performance, nor do we view raw accuracy as the sole measure of a model's contribution; our primary goal is to share this interesting and insightful behavior. Combined with our supplementary experiments and our parameter-efficiency rationale for polynomials, we believe the current evidence sufficiently demonstrates our method's simplicity, efficiency, and competitive performance, providing a new perspective for interpretability analysis and validating our contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.