A new model, PaGNet, is proposed for forecasting corporate tax avoidance proxies from firm-year panel data. It combines a LightGBM branch with a Panel-MLP branch, using attention-pooled temporal aggregation and shared-trunk multi-task learning. The model is evaluated on a dataset of Korean listed firms, showing improved performance over six baselines. The model's branch-reliance diagnostic provides additional insights into the forecasting process.