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

Daniel Durstewitz; Christoph Jürgen Hemmer; Florian Hess; Charlotte Ricarda Doll; Lukas Eisenmann

cs.LGFeb 18, 2026

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

Authors: Daniel Durstewitz, Christoph Jürgen Hemmer, Florian Hess, Charlotte Ricarda Doll, Lukas Eisenmann

Organizations: Dept. of Theoretical Neuroscience, Central Institute of Mental Health, Mannheim, Germany · Faculty of Physics & Astronomy, Heidelberg Univ., Heidelberg, Germany · Interdisciplinary Center for Scientific Computing, Heidelberg, Germany

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.

Figures & tables

Figure 1: a ) State space of a bistable neuron model with co-existing point (left basin) and limit cycle (right basin) attractor (see Appx. D.2 for details). The two basins are separated by the stable manifold of a saddle node . Trajectories follow the system’s vector field, indicated by arrows with red shading indicating flow velocity (darker = faster). The corresponding TS (right) of model variables $V(t)$ (solid) and $n(t)$ (dashed) converge either to a point attractor (top) or a stable limit cycle (bottom) depending on the basin in which the trajectory was initialized (as indicated by the blue arrows). b ) N-tipping in the true neuron model (gray) and in an AL-RNN (red) trained on trajectories from both basins. Noise was added in both models and eventually drives them across the basin boundary into cyclic activity (cf. Fig. 9 ).

Figure 2: a ) Bifurcation diagram of a spiking neuron model depending on a control parameter $h$ (see Appx. D.2 ; Durstewitz (2009) ). Stable fixed points are indicated by solid black lines, unstable ones by dashed black lines. The gray-shaded area corresponds to a stable limit cycle. Graphs below give snapshots of the state space for different values of $h$ . b ) B-tipping from a ‘bursting’ into a ‘spiking’ regime in the full simulated neuron model (gray) is successfully predicted by an AL-RNN (red) trained only on TS data up to the black dashed line.

Figure 3: DSR measures: Comparison of geometric (dis)agreement ( $D_{\textrm{stsp}}$ ), power spectral distance ( $D_{\textrm{H}}$ ), Kaplan-Yorke fractal dimension ( $D_{\textrm{KY}}$ ), and max. Lyapunov exponent ( $\lambda_{\textrm{max}}$ ) on a ) a poor and b ) a good reconstruction of the chaotic Lorenz-63 system by an AL-RNN. See also Fig. 8 .

Figure 4: The limits of predictability: Three TS from the same Lorenz-63 system (same parameters) started at the same initial condition quickly diverge even with just $1\%$ of noise, while with $10\%$ noise prediction beyond 1 Lyapunov time becomes hopeless.

Figure 5: a ): Long-term geometric ( $D_{\textrm{stsp}}$ ) and temporal ( $D_{\textrm{H}}$ ) forecast accuracy (cf. Appx. B ) as median $\pm$ MAD, comparing DSR and TS models, evaluated on $10,000$ -step roll-outs. Lower = better. Results for both custom-trained (triangles) and foundation models (dots) are shown. b ): Short-term forecasting accuracy ( MASE ); lighter colors = better, green stars = best MASE . c ): Example long-term forecasts of ETTh1 data for DSR and TS models, exposing the failure of TS models to capture long-term behavior. See also Figs. 14 & 15 .

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Normal forms and bifurcation diagrams for saddle-node and supercritical pitchfork and Hopf bifurcations. Shown are the stable and unstable equilibria as solid and dashed black lines, respectively, and the limit cycle as shaded region between its minimum–maximum branches (gray lines), as the control parameter $c$ in the respective normal form is varied.

$I$	$C$	$g_{L}$	$E_{L}$	$g_{\mathrm{Na}}$	$E_{\mathrm{Na}}$	$V_{h,\mathrm{Na}}$	$k_{\mathrm{Na}}$	$g_{K}$	$E_{K}$	$V_{h,K}$	$k_{K}$	$\tau_{n}$	$g_{M}$	$V_{h,M}$	$k_{M}$	$\tau_{h}$	$g_{\mathrm{NMDA}}$	$E_{\mathrm{NMDA}}$
$0$	$6$	$8$	$-80$	$20$	$60$	$-20$	$15$	$10$	$-90$	$-25$	$5$	$1$	$25$	$-15$	$5$	$200$	$10.2$	$0$

Appendix

Neuron model parameter settings

Figure 7: Illustration of DSR process. “ $\approx$ ” is meant here in the sense of topological conjugacy (cf. Def. A.4 ) and (potentially) geometric & temporal invariants (like the Lyapunov spectrum or fractal dimension).

Figure 8: Illustration of DSR measures: Comparison of geometric (dis)agreement ( $D_{\textrm{stsp}}$ ), power spectral distance ( $D_{\textrm{H}}$ ), Kaplan-Yorke fractal dimension ( $D_{\textrm{KY}}$ ), and max. Lyapunov exponent ( $\lambda_{\textrm{max}}$ ) on reconstructions of the chaotic Lorenz-63 system of different quality from very poor ( a ) to excellent ( d ), obtained by an AL-RNN assessed at different training stages.

Figure 9: N-tipping in state space: Illustration of the basin boundary crossing of the noisy trajectory from Fig. 1 b.

Figure 10: A DSR model trained on just one short trajectory (center) from the chaotic Rössler ( a ) or Lorenz ( b ) system (left) correctly reconstructs the full chaotic attractor (red trajectory), generalizes to novel initial conditions (red dots), and correctly infers the location of fixed points not shown in training (yellow dots). Part (a) based on Suppl. Fig. 4 of Durstewitz et al. (2023) .

Figure 11: Although only trained on a short trajectory snippet (gray curve) from the Selkov system (left), the trained DSR model (shallow PLRNN ( Hess et al., 2023 ) ; right) correctly infers properties of the surrounding vector field (darker shading = higher flow velocity), enabling it to generalize to new initial conditions.

Figure 12: Illustration of a DSR training corpus. Time series (center) from different chaotic attractors (left), generated by the equations on the right. The top time series are from non-autonomous systems where a control parameter (in light blue on the right) was systematically varied, while the bottom time series are from autonomous and stationary system simulations ( $135$ such systems are collected in Gilpin (2022) , from which through recombination $\approx 10^{4}$ systems are created in Lai et al. (2025) ).

Figure 13: a ) Example of generic skew-product form of a dynamical system that implements non-autonomous DS by separating a slow base (driver) system from a fast fiber (forced, dependent) system ( Kloeden and Yang, 2020 ) . b ) Specific example (using a piecewise-linear RNN) of how such a form may be enforced in training through a particular regularization (inductive bias) on parameters which encourages the RNN to segregate its latent space into base and fiber system, inspired by Schmidt et al. (2021) .

Figure 14: Long-term forecasts of ETTh1 dataset for different DSR and TS models.

Figure 15: Long-term forecasts of weather temperature dataset for different DSR and TS models.

System	ALRNN			shPLRNN			RC			AutoARIMA			PatchTST			NBEATS
	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE
ETTh1	$0.46\pm 0.34$	$0.16\pm 0.02$	$3.09\pm 0.07$	$0.21\pm 0.13$	$0.18\pm 0.01$	$1.61\pm 0.08$	$0.35\pm 0.01$	$0.17\pm 0.01$	$1.09\pm 0.00$	$3.92$	$0.22$	$1.46$	$3.85\pm 2.66$	$0.16\pm 0.05$	$4.78\pm 1.82$	$3.14\pm 0.57$	$0.38\pm 0.13$	$3.77\pm 0.42$
Traffic	$1.14\pm 0.57$	$0.14\pm 0.03$	$1.90\pm 0.04$	$0.25\pm 0.08$	$0.09\pm 0.01$	$2.35\pm 0.06$	$0.10\pm 0.02$	$0.13\pm 0.02$	$1.26\pm 0.27$	$8.98$	$0.41$	$2.16$	$1.23\pm 0.68$	$0.24\pm 0.10$	$0.98\pm 0.08$	$1.54\pm 0.12$	$0.13\pm 0.03$	$0.92\pm 0.06$
Cloud Requests	$1.28\pm 0.12$	$0.12\pm 0.00$	$3.02\pm 0.05$	$1.14\pm 0.10$	$0.12\pm 0.00$	$2.88\pm 0.03$	$0.73\pm 0.00$	$0.74\pm 0.01$	$23.12\pm 2.06$	$0.61$	$0.10$	$1.90$	$3.21\pm 2.04$	$0.15\pm 0.03$	$2.40\pm 0.37$	$2.10\pm 0.34$	$0.12\pm 0.02$	$4.05\pm 0.79$
Weather (Temp.)	$0.73\pm 0.02$	$0.11\pm 0.00$	$1.80\pm 0.04$	$0.71\pm 0.28$	$0.12\pm 0.00$	$1.62\pm 0.12$	$0.07\pm 0.02$	$0.12\pm 0.03$	$2.53\pm 0.28$	$0.11$	$0.11$	$2.60$	$1.04\pm 0.87$	$0.12\pm 0.03$	$3.19\pm 1.25$	$1.22\pm 0.83$	$0.11\pm 0.02$	$2.09\pm 0.12$
Weather (Press.)	$0.23\pm 0.09$	$0.29\pm 0.02$	$4.66\pm 0.22$	$0.25\pm 0.03$	$0.30\pm 0.02$	$4.80\pm 0.24$	$1.90\pm 1.05$	$0.89\pm 0.05$	$4.68\pm 0.62$	$13.26$	$1.00$	$4.63$	$5.90\pm 0.38$	$0.45\pm 0.01$	$5.63\pm 0.54$	$5.42\pm 2.45$	$0.49\pm 0.18$	$8.47\pm 3.98$

Appendix

Table 1: Performance comparison of custom trained DSR and TS models on empirical time series in terms of geometrical divergence (

D_{\text{stsp}}

), long-term temporal distance (

D_{H}

), and forecast error (MASE). Best in red , second-best in blue .

System	DynaMix			Chronos-t5-base			Chronos-2			TiRex			Panda
	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE	$D_{\text{stsp}}$	$D_{H}$	MASE
ETTh1	$0.60\pm 0.01$	$0.15\pm 0.00$	$1.61\pm 0.41$	$1.58\pm 0.23$	$0.20\pm 0.01$	$1.30\pm 0.03$	$3.44$	$0.16$	$1.71$	$9.20$	$0.29$	$1.46$	$8.18$	$0.69$	$2.86$
Traffic	$0.86\pm 0.02$	$0.18\pm 0.02$	$0.98\pm 0.10$	$0.78\pm 0.24$	$0.11\pm 0.01$	$0.62\pm 0.01$	$0.55$	$0.13$	$0.65$	$3.41$	$0.81$	$0.59$	$4.47$	$0.45$	$0.99$
Cloud Requests	$0.67\pm 0.00$	$0.07\pm 0.01$	$2.62\pm 0.02$	$3.72\pm 1.91$	$0.53\pm 0.08$	$2.43\pm 0.02$	$2.10$	$0.14$	$1.66$	$3.06$	$0.89$	$1.44$	$7.20$	$0.23$	$2.38$
Weather (Temp.)	$0.06\pm 0.00$	$0.07\pm 0.01$	$1.89\pm 0.04$	$8.92\pm 3.75$	$0.44\pm 0.24$	$7.24\pm 0.40$	$2.27$	$0.17$	$1.59$	$3.00$	$0.90$	$1.91$	$1.15$	$0.14$	$1.60$
Weather (Press.)	$0.18\pm 0.01$	$0.26\pm 0.01$	$4.21\pm 0.02$	$3.23\pm 0.67$	$0.54\pm 0.00$	$4.76\pm 0.16$	$11.16$	$0.47$	$4.94$	$7.29$	$0.75$	$4.75$	$10.23$	$0.68$	$4.72$

Appendix

Table 2: Performance comparison of DSR and TS foundation models on empirical time series in terms of geometrical divergence (

D_{\text{stsp}}

), long-term temporal distance (

D_{H}

), and forecast error (MASE). Best in red , second-best in blue .

Figure 16: Example forecast of the chaotic Lorenz-63 system using DynaMix and Chronos-t5-base. Chronos’ tendency to parrot its context input leads to a cyclic repetition of a fixed pattern, and Chronos therefore fundamentally fails to capture the truly aperiodic, chaotic behavior. This manifests as cyclic trajectories in the delay-embedded (cf. Appx. A ) state space (left), the close-to-zero maximum Lyapunov exponent (center) and by a sharply peaked (instead of the truly broad) power spectrum (right). Results reproduced from Hemmer and Durstewitz (2025) .

Figure 17: TS forecasts using custom trained AutoARIMA.

Figure 18: TS forecasts using custom trained shallow PLRNN.

Figure 19: TS forecasts using custom trained AL-RNN.

Figure 20: TS forecasts using custom trained Reservoir Computers.

Figure 21: TS forecasts using custom trained PatchTST.

Figure 22: TS forecasts using custom trained NBEATS.

Figure 23: Zero-shot TS forecasts using DynaMix FM.

Figure 24: Zero-shot TS forecasts using Chronos-t5-base FM.

Figure 25: Zero-shot TS forecasts using Chronos-2 FM.

Figure 26: Zero-shot TS forecasts using TiRex FM.

Figure 27: Zero-shot TS forecasts using Panda FM.

Figure 28: Impact of observability, noise, and DSR-specific training methods on short-term prediction. a ) The number of observed variables has only little influence on reconstruction quality (as measured by $D_{\text{stsp}}$ ), as illustrated here for an ALRNN ( Brenner et al., 2024a ) trained by STF ( Mikhaeil et al., 2022 ) on the $20d$ Lorenz-96 system. b ) Short-term prediction error (MASE on 64 step forecasts) increases with noise level (N/S ratio), but does so more strongly for TS models (blue) than for DSR models (red), for stochastic Lorenz-63. c ) Optimal DSR training improves short-term prediction (tested on stochastic Lorenz-63): Training a shPLRNN by GTF ( Hess et al., 2023 ) or STF ( Mikhaeil et al., 2022 ) with dynamically optimal forcing interval ( $\tau=16$ ) yields lower forecast errors than with suboptimal forcing ( $\tau=100$ ) or without GTF; see Mikhaeil et al. (2022) and Hess et al. (2023) who observed the same for DSR performance. Error bars = MAD.

Figure 29: Predictions from a shallow PLRNN ( Hess et al., 2023 ) trained via sparse teacher forcing ( Mikhaeil et al., 2022 ) on sequential MNIST with the digit label provided as an external input. For each digit, a forecast was produced (top panels) and then reshaped to reconstruct the corresponding image (bottom panels).

Figure 30: Proof-of-concept example that other types of architecture than typically used in the DSR literature could also profit from DSR training techniques, in this case a Mamba-2 block ( Dao and Gu, 2024 ) followed by a 1-hidden-layer MLP trained with (red) or without (blue) GTF ( Hess et al., 2023 ) on human EEG data (gray). Both long-term performance measures ( $D_{\text{stsp}}$ with vs. w/o GTF: $4.86$ vs. $11.45$ , $D_{\text{H}}$ : $0.13$ vs. $0.42$ ) as well as the short-term prediction error ( $\text{MASE}(20)$ : $10.1$ vs. $14.2$ ) are improved by GTF.

Acronym	Definition
AI	Artificial Intelligence
AL-RNN	Almost-Linear Recurrent Neural Network
AR	Auto-Regressive
ARMA	Auto-Regressive Moving-Average
AutoARIMA	Automatic Auto-Regressive Integrated Moving Average
DS	Dynamical System(s)

Appendix

Table 3: Acronyms used throughout the main text.

Time Series Modeling

See all 51 papers

Jun 22, 2026

Topological Out-of-Domain Generalization in Dynamical Systems Reconstruction

Predicting the behavior of dynamical systems (DS) beyond the dynamical and parameter regimes observed in training is a pivotal and essentially unresolved problem in scientific ML. It is central to any good scientific theory, which we expect to be able to make predictions about regimes not covered by currently available data. Recent hierarchical and hyper-network guided approaches for DS reconstruction (DSR) enable training on many DS simultaneously, and revealed that extracted latent features are often related to crucial control parameters of the underlying DS that varied across the training corpus. However, true out-of-domain forecasting abilities of these models, e.g., across tipping points, remain limited, and fine-tuning, or even full model retraining, on time series from the new dynamical regime is usually required. Here, we mathematically analyze the root of these limitations in previous model formulations and identify three core shortcomings rooted in a mismatch between structural assumptions of the reconstruction model and typical properties of physical systems. We propose a combination of remedies for these shortcomings, most importantly feature splitting, and furthermore derive a closed-form bound on the reliable extrapolation range. We demonstrate empirically that our techniques allow for accurate zero-shot prediction into new dynamical regimes, outside the observed training regime, as, e.g., encountered across tipping points.

Georg Trede, Charlotte Ricarda Doll, Elias Weber +1

Dept. of Theoretical Neuroscience, Central Institute of Mental Health, Mannheim, Germany · Faculty of Physics and Astronomy, Heidelberg University, Germany · Interdisciplinary Center for Scientific Computing (IWR), Heidelberg University, Germany

Dynamical Systems Zero-Shot Time Series Forecasting OOD Generalization Nonlinear System Identification

Ask PDF

Sep 28, 2026

Dynamical Parameters: An Interpretability Framework for Time-Series Foundation Models

This work studies a central gap in interpreting time-series foundation models (TSFMs): a dynamical property may be accessible in a hidden state even when the forecast fails to respond correctly as that property changes. We formalize these properties as Dynamical Parameters, including trend slope, oscillation frequency, and autoregressive dependence. We compare their representation accessibility, measured by recovery from hidden states, with their forecast response, measured by agreement with the expected forecast change. Across nine frozen TSFMs and thirteen laws, 42 of 63 model-parameter cells achieve accessibility above 0.95, whereas their median reference-aligned response relative to the conditional reference is only 0.46. To explain this gap, causal geometry compares the hidden-state change required to produce the reference response with the change induced by the parameter intervention. Directly modifying the hidden state recovers the reference response, but the parameter intervention often moves the state in a different direction. These results show that accessible parameter information need not be expressed in forecasts when input changes miss the required hidden-state direction.

Kang Yang, Gaofeng Dong, Liying Han +1

Department of Electrical and Computer Engineering University of California, Los Angeles

Explainable Time Series Forecasting Representation Probing Time Series Representation Learning Time-Series Foundation Models Time Series Modeling

Ask PDF

Sep 27, 2026

Correct then Forecast: Observer State-Space Models for Time Series Forecasting

Time series forecasting requires extrapolating the dynamics of an observed process beyond the last available measurement. Yet recurrent forecasting models typically treat observations as inputs that directly control their latent dynamics. It leads to a regime change when these observations become unavailable at prediction time. Following a state-estimation perspective, we introduce Observer State-Space Models (OSSMs), a class of recurrent models that interprets the observed input time series as measurements of an underlying autonomous dynamical system. OSSMs explicitly separate latent-state propagation from measurement assimilation: a single transition governs the dynamics across both context and forecasting intervals, while available observations correct the estimated state through an observer. This formulation naturally exposes classical control-theoretic properties, including observability and convergence of the state estimation error. We further show that conventional and recent SSMs can be recovered as particular instances of our OSSM framework, thereby providing a unified interpretation of their recurrent dynamics and revealing modeling inconsistencies. We perform experiments across several benchmarks showing that OSSM achieves substantial improvements while maintaining the same parameter count and training setup as the corresponding SSM baseline. These results support a simple principle for recurrent forecasting: observations should correct the estimated latent state, rather than control the dynamics used to propagate it.

Alexis-Raja Brachet, Guillaume Clavier--Frémond, Abdelhakim Ziani +2

MICS, CentraleSup´elec, Universit´e Paris-Saclay, France · CentraleSup´elec, IETR UMR CNRS 6164, France · Universit`a di Torino, Torino, Italy.

Latent Dynamics Modeling Time Series Forecasting State Space Models Time Series Modeling

Ask PDF

Abstract

Time Series Modeling

Explore similar work

Topological Out-of-Domain Generalization in Dynamical Systems Reconstruction

Dynamical Parameters: An Interpretability Framework for Time-Series Foundation Models

Correct then Forecast: Observer State-Space Models for Time Series Forecasting