PaGNet Hybrid Model for Corporate Tax Avoidance Forecasting
September 18, 2026
PaGNet combines a LightGBM branch for panel-temporal summaries with a Panel-MLP branch using attention-pooled temporal aggregation. Tested on a dataset of 1,754 Korean firms, the hybrid architecture uses a non-trainable blender to provide both predictions and branch-reliance diagnostics for transparency.
HOW THIS AFFECTS YOU
●
builderThis hybrid approach provides a template for building transparent, multi-target forecasting systems on panel data.
●
researcherThe architecture demonstrates how to combine gradient boosting and neural networks for structured temporal data.