Causality-Informed Large-Small Model Collaborative Framework for Time Series Forecasting: Dataset

We propose a causality-informed collaborative framework that integrates large pre-trained models with small auxiliary models for time series forecasting. The architecture leverages the strong baseline predictions of large models while introducing two complementary enhancements: (i) residual correction guided by domain-specific signals, and (ii) causal denoising with external instrumental variables to mitigate spurious correlations. Comprehensive experiments on real-world financial and environmental datasets demonstrate that the framework consistently improves out-of-sample forecasting accuracy over large model baselines. Notably, incorporating instrumental-variable\u2013based denoising yields additional gains, showing that exogenous signals can further enhance predictive performance. Moreover, the improvements scale with model size, as larger pre-trained models benefit more from the collaborative design. Overall, the study demonstrates that causality-aware auxiliary modules can systematically boost large pre-trained models, delivering interpretable and scalable improvements for time-series forecasting in noisy real-world environments.