RAMuST: A Regime-Aware Multiscale and Mixed-Frequency Transformer for Industry-Level Corporate Income Tax Forecasting

We present the Regime-Aware Multiscale Transformer (RAMuST), a mixed- frequency Transformer-based architecture that couples a multi-scale monthly en- coder with a yearly regime/shock head and a Transformer decoder for industry- level corporate income tax forecasting. In mixed-frequency data, monthly macro–economic and financial indicators are patchified at several temporal scales and fused through a soft scale gate integrated with economic cycle, trends, volatil- ity, and seasonality, while yearly regime embeddings and a shock scalar summa- rize within-year dynamics and transient disturbances. To analyze the model, we designate the 2021–2024 window—covering the post COVID-19 period—as the validation interval and focus on the two most volatile Korean industries, Human health and Real estate activities, for corporate income tax forecasting. As a result of experiments, the model consistently surpasses recent strong time-series baselines by at least threefold and up to tenfold.