Dynamic Modular Time Series Models: On-the-Fly Reconfiguration for Domain Adaptation

TL;DR: What if your time series model could rewire itself like LEGO blocks depending on the data domain? Let's build modular small hybrid models, where convolutional, RNN, and transformer layers are dynamically composed at inference time for domain-specific adaptation. To test this, we’ll create a model zoo of lightweight modules, a controller that selects modules per input, and benchmark performance on heterogeneous zero-shot forecasting tasks.

Research Question: Can dynamically reconfigurable modular hybrid models—assembled from small convolutional, RNN, and transformer blocks—match or exceed the performance of monolithic models like Reverso in zero-shot forecasting across diverse domains?

Hypothesis: A modular, dynamically reconfigurable architecture will outperform static small hybrid models (e.g., Reverso) in cross-domain zero-shot forecasting, by tailoring model structure to the unique characteristics of each time series input.

Experiment Plan: Design a library of lightweight modules (convolution, RNN, transformer, etc.). Develop a controller (possibly a reinforcement learning agent) that assembles modules at inference time based on input characteristics (e.g., periodicity, noise). Train on a diverse multitask dataset, testing on held-out domains. Compare with Reverso and similar fixed-architecture models on performance, efficiency, and adaptability. Analyze module selection patterns for interpretability.

References:

• Fu, X., Li, Y., Papaioannou, G., & Kim, Y. (2026). Reverso: Efficient Time Series Foundation Models for Zero-shot Forecasting.
• Shi, H.-N., Huang, T.-J., Han, L., Zhan, D.-C., & Ye, H.-J. (2025). One-Embedding-Fits-All: Efficient Zero-Shot Time Series Forecasting by a Model Zoo. arXiv.org.