TabPFN-TSRA: Retrieval-Augmented TabPFN for Time-Series Forecasting

TabPFN-based time-series forecasting enables zero-shot prediction by reformulating forecasting as tabular regression, but existing methods still rely on explicit temporal features and global historical summarization, which can be insufficient for non-stationary, event-driven dynamics and may dilute informative local signals. We propose TabPFN-TSRA, a retrieval-augmented framework for TabPFN-based forecasting. TabPFN-TSRA retrieves analogous historical patterns using level-shift-corrected correlation matching and organizes the retrieved segments into temporally aligned feature matrices for TabPFN. This design provides compact and relevant local evidence beyond fixed features while preserving temporal structure. Experiments on 11 benchmark datasets show that TabPFN-TSRA outperforms TabPFN-TS and TabPFN-TSP, achieving an 82% win rate in terms of MAE and enhancing the ability to forecast trend evolution. These results demonstrate the effectiveness of retrieval augmentation for TabPFN-based forecasting. Code will be available at: xxx.