Explaining Small Hybrid Models: Interpretable Attribution for Efficient Zero-Shot Forecasters

TL;DR: Small models like Reverso are fast and accurate, but can we also make them explainable? We’ll develop new attribution methods tailored for convolution-RNN hybrids and show how they can reveal the “reasoning steps” behind zero-shot forecasts, providing confidence and insight for practitioners.

Research Question: Can we develop interpretable attribution techniques specifically for small hybrid time series models (convolution + RNN) to better understand and trust their zero-shot forecasts?

Hypothesis: Customized attribution methods (e.g., hybrid integrated gradients, temporal occlusion) will yield more faithful and actionable explanations for small hybrid models than generic explainability tools, improving user trust and facilitating model debugging.

Experiment Plan: Develop hybrid attribution methods that respect the unique structure of convolution-RNN interleaved models. Apply these to Reverso and baseline transformer models on standard benchmarks. Conduct user studies with practitioners to evaluate the usefulness and clarity of the explanations. Analyze how explanations vary between architectures and what insights they provide.

References:

• Fu, X., Li, Y., Papaioannou, G., & Kim, Y. (2026). Reverso: Efficient Time Series Foundation Models for Zero-shot Forecasting.
• Chakraborty, S., Delibasoglu, I., & Heintz, F. (2025). Scaling transformers for time series forecasting: do pretrained large models outperform small-scale alternatives? Artificial Intelligence Review.