Position: Why a Dynamical Systems Perspective is Needed to Advance Time Series Modeling

Abstract: Time series (TS) modeling has come a long way from early statistical, mainly linear, approaches to the current trend in TS foundation models. With a lot of hype and industrial demand in this field, it is not always clear how much progress there really is. To advance TS forecasting and analysis to the next level, here we argue that the field needs a dynamical systems (DS) perspective. TS of observations from natural or engineered systems almost always originate from some underlying DS, and arguably access to its governing equations would yield theoretically optimal forecasts. This is the promise of DS reconstruction (DSR), a class of ML/AI approaches that aim to infer surrogate models of the underlying DS from data. But models based on DS principles offer other profound advantages: Beyond short-term forecasts, they enable to predict the long-term statistics of an observed system, which in many practical scenarios may be the more relevant quantities. DS theory furthermore provides domain-independent theoretical insight into mechanisms underlying TS generation, and thereby will inform us, e.g., about upper bounds on performance of any TS model, generalization into unseen regimes as in tipping points, or potential control strategies. After reviewing some of the central concepts, methods, measures, and models in DS theory and DSR, we will discuss how insights from this field can advance TS modeling in crucial ways, enabling better forecasting with much lower computational and memory footprints. We conclude with a number of specific suggestions for translating insights from DSR into TS modeling.

• The paper demonstrates that integrating dynamical systems theory with time series modeling uncovers underlying attractor geometries and fundamental predictability limits.
• It details advanced DSR methodologies such as control-theoretic teacher forcing and invariant measures, enabling robust capture of chaotic dynamics.
• It provides empirical evidence that DSR models outperform traditional time series models in long-term forecasting while using fewer parameters.

The Imperative of a Dynamical Systems Perspective in Time Series Modeling

Foundations of Dynamical Systems (DS) Theory in Time Series Analysis (TSA)

The paper advances a position that time series modeling should be firmly rooted in dynamical systems theory, arguing that virtually all real-world time series data arise from an underlying DS—the temporal evolution governed by differential equations or recursive maps. The authors thoroughly dissect fundamental DS notions: state spaces, trajectories, attractors (point, limit cycle, chaotic), basins of attraction, bifurcations, and tipping points, emphasizing the geometric and topological structure that DS theory provides to understanding TS phenomena (Figure 1).

Figure 1: State space of a bistable neuron model with co-existing point and limit cycle attractors, demonstrating multistability and basin boundaries.

A critical insight emphasized is multistability: the coexistence of multiple attractors for a fixed set of parameters, typical in complex systems (e.g., climate, neural circuits), and the role of noise or parameter shifts in initiating tipping points—a regime transition either via noise-induced crossing (N-tipping) or bifurcation (B-tipping). The authors argue that such features are not mere mathematical curiosities but frequently govern empirical time series, with significant implications for predictability and control.

The formalism extends to rigorous definitions: Lyapunov exponents quantifying stability and chaos, delay embedding theorems for recovering attractor geometry from partial observations, and topological conjugacy for surrogate model equivalence. The presentation is mathematically rigorous, aligning with the expectations of researchers seeking precise characterizations.

Dynamical Systems Reconstruction (DSR): Methodology and Evaluation

DSR seeks to infer a surrogate generative model from time series data that approximates the true DS, ideally establishing topological conjugacy between model and data-generating system. The review synthesizes a wide spectrum of DSR architectures: RNNs (especially piecewise-linear variants for tractability), reservoir computers, SINDy, Koopman operator theory, and continuous-time models like Neural ODEs. The absence of transformer-based architectures in DSR is attributed to their lack of intrinsic temporal recursion, an essential property for modeling DS flow maps.

The authors delineate the central requirements of DSR: reproduction of attractor geometry and invariant measures, preservation of long-term statistics, and fidelity to temporal structure. Empirical evaluation employs state space distribution metrics (e.g., Kullback-Leibler divergence, Wasserstein distance), fractal dimensions, Lyapunov spectra, and power spectral distances (Figure 2).

Figure 2: DSR measures—comparison of geometrical and spectral agreement, fractal dimension, and Lyapunov exponent, highlighting reconstruction quality in chaotic systems.

Training methodologies are scrutinized, particularly the efficacy of control-theoretic teacher forcing variants (sparse teacher forcing, generalized teacher forcing) in overcoming chaos-induced exploding gradients, enabling stable long rollouts. Alternative loss functions incorporating dynamical invariants (e.g., Lyapunov exponents) and optimal transport measures are recommended for direct alignment with DS objectives.

Interpretability is underscored as a hallmark of DSR: models may not only forecast but also yield mechanistic insight, reconstructing the vector field and enabling topological analysis—critical for scientific inference and real-world interventions.

Insights and Empirical Comparisons: DSR vs. TS Models

Strong empirical evidence is provided for DSR superiority in both short- and long-term forecasting. Model comparisons demonstrate that DSR trained with control-theoretic methods (AL-RNN, shPLRNN, RC) outperform conventional TS models (ARIMA, PatchTST, NBEATS) in preserving long-term attractor statistics and geometric properties (Figure 3). Notably, DSR achieves this with orders-of-magnitude fewer trainable parameters, challenging assumptions about model scale versus complexity.

Figure 3: Long-term geometrical and spectral forecast accuracy for DSR and TS models, with DSR dominating both custom and foundation architectures.

Additionally, DSR models frequently match or exceed TS models even in short-term quantitative prediction (MASE), despite their simpler architecture and reduced resource demands. Furthermore, transformer-based TS foundation models (Chronos, Chronos-2, Panda, TiRex) are empirically shown to fail at reconstructing chaotic attractors and exhibit context parroting, converging to fixed points or limit cycles and misrepresenting fundamentally aperiodic dynamics (Figure 4).

Figure 4: Zero-shot forecast of Lorenz-63—Chronos fails to capture chaotic behavior, parroting context and yielding cyclic trajectories; DynaMix correctly reproduces chaos.

The paper posits that the training protocol and corpus―particularly simulation data from DS, including chaotic regimes―are more consequential for forecasting and representation than architectural sophistication. Foundation models trained on simulated DS (e.g., DynaMix) generalize robustly and exhibit tractable, integrated representations.

Implications: Limits of Predictability, Tipping Points, and OOD Generalization

A DS perspective elucidates the principle limits of forecast horizons imposed by chaos—positive Lyapunov exponents guarantee exponential divergence, rendering long-term prediction statistically futile, especially in the presence of noise (Figure 5). The paper strongly cautions against claims of prolonged prediction in noisy chaotic domains, highlighting the necessity of invariant measure evaluation over MSE or MAPE.

Figure 5: Three trajectories from Lorenz-63 with noise—initial conditions quickly diverge, illustrating the fundamental limit to predictability.

DSR enables detection and forecast of tipping points—abrupt regime shifts due to noise (N-tipping), parameter drift (B-tipping), or rate-induced changes (R-tipping). High-fidelity DSR models offer the prospect of forecasting and simulating post-tipping dynamics and regime transitions, critical for climate, neurology, and socio-economic systems.

A detailed treatment of the challenge of OOD generalization distinguishes topological shifts in DS (bifurcations, regime transitions) from mere distribution shifts, establishing that DSR's mechanistic approach can, in principle, extrapolate into unseen regimes if control parameters and context can be inferred or included.

Alternative Perspectives and Integration with Statistical Methods

The authors acknowledge limitations of classical DS theory (low-dimensional, deterministic, autonomous, fully observed). They argue, however, that modern DSR models and control-theoretic training techniques scale indifferently to dimensionality and noise and are empirically validated in high-dimensional noisy datasets (e.g., fMRI, EEG, temperature records). The DS framework interfaces naturally with ergodic theory, statistical measures, and may be integrated with LLMs to leverage external context and world knowledge, especially in regime forecasting.

Recommendations and Future Directions

The paper concludes with concrete recommendations:

1. Adopt DSR training techniques—sparse and generalized teacher forcing, invariant-measure regularization, and rollouts—in TS modeling to improve both short- and long-term prediction with reduced complexity.
2. Pre-train foundation models on DS simulations—especially chaotic DS—to better capture variability and generalization in real-world TS data.
3. Refocus on RNNs and continuous-time models—in light of their dynamical universality and tractable time representation, moving away from transformer-centric TS modeling.
4. Address hard OOD problems via DS methods—explicitly simulate and infer regime transitions and control parameters to forecast and extrapolate through topological shifts.
5. Emphasize mechanistic understanding—DSR models provide insight independent of specific domain expertise, informing robust predictions and interventions.

Conclusion

This position paper delivers a rigorous, empirically substantiated argument that advancing time series modeling necessitates a dynamical systems perspective. By foregrounding geometric, topological, and ergodic properties, DS theory and DSR establish fundamental constraints and capabilities in forecasting, regime detection, and generalization. The integration of DSR methodologies—both architectural and algorithmic—enables more accurate, interpretable, and efficient models, with broad applicability across scientific, medical, and industrial domains. The recommendations delineate actionable pathways for AI researchers to enhance both the practical fidelity and theoretical robustness of time series forecasting.

Figure 6: DSR model trained on a single short chaotic trajectory generalizes to reconstruct the full attractor and unseen initial conditions, demonstrating the power of mechanistic DS reconstruction.