Time Series Forecasting
File size: 2,915 Bytes
bbf1ad1
 
 
 
 
 
 
 
 
e5aac22
bbf1ad1
 
 
 
 
 
 
 
ef7222a
bbf1ad1
 
 
 
 
 
 
 
 
 
 
 
 
ef7222a
bbf1ad1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
---
license: cc-by-4.0
pipeline_tag: time-series-forecasting
datasets:
- williamgilpin/dysts
---

# DynaMix

[![arXiv](https://img.shields.io/badge/arXiv-2505.13192-b31b1b.svg)](https://arxiv.org/abs/2505.13192) (accepted NeurIPS 2025 paper)

DynaMix is a foundation model for zero-shot inference of dynamical systems that preserves long-term statistics. Unlike traditional approaches that require retraining for each new system, DynaMix provides context driven generalization to unseen dynamical systems.

- **Accurate Zero-Shot Dynamical Systems Reconstruction**: DynaMix generalizes across diverse dynamical systems without fine-tuning, accurately capturing attractor geometry and long-term statistics.
- **Context Felxible Dynamics Modeling**: The multivariate architecture captures dependencies across system dimensions and adapts to different dimensionalities and context lengths.
- **Efficient and Lightweight**: Designed to be efficient with a few thousand parameters, DynaMix can also run on CPU for inference, and is enabling orders-of-magnitude faster inference than traditional foundation models.
- **General Time Series Forecasting**: Extends beyond DSR to general time series forecasting using adaptable embedding techniques.

For complete documentation and code, visit the [DynaMix repository](https://github.com/DurstewitzLab/DynaMix-python).

## Model Description

DynaMix is based on a mixture of experts (MoE) architecture operating in latent space:

1. **Expert Networks**: Each expert is a specialized dynamical model, given trhough RNN based architectures

2. **Gating Network**: Selects experts based on the provided context and current latent representation of the dynamics

By aggregating the expert weighting with the expert prediction the next state is predicted.

## Usage

To load the model in python using the corresponding codebase [DynaMix repository](https://github.com/DurstewitzLab/DynaMix-python), use:

```python
from src.utilities.utilities import load_hf_model

# Initialize model with architecture
model = load_hf_model(model_name="dynamix-3d-alrnn-v1.0")
```

Given context data from the target system with shape (`T_C`, `S`, `N`) (where `T_C` is the context length, `S` the number of sequences that should get processed and `N` the data dimensionality), generate forecasts by passing the data through the `DynaMixForecaster` along with the loaded model. Further details can be found in the GitHub repository [DynaMix repository](https://github.com/yourusername/zero-shot-DSR).


## Citation

If you use DynaMix in your research, please cite our paper:

```
@misc{hemmer2025truezeroshotinferencedynamical,
title={True Zero-Shot Inference of Dynamical Systems Preserving Long-Term Statistics},
author={Christoph Jürgen Hemmer and Daniel Durstewitz},
year={2025},
eprint={2505.13192},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2505.13192},
}
```