PaGNet: A Panel-Aware GBDT--Neural Network for Mult... | AI Research

Forecasting corporate tax avoidance is a critical task for regulators and auditors, but it is notoriously difficult because predictive signals are often hidden across multiple years of financial data and related tax metrics. Furthermore, because these models are used for high-stakes screening, they must be transparent and explainable rather than acting as "black boxes." This paper introduces PaGNet (Panel-Aware GBDT–Neural Network), a hybrid model designed to forecast tax avoidance proxies while providing clear insights into how the model reaches its conclusions.

A Hybrid Approach to Forecasting

PaGNet processes firm-year data using two distinct branches that work in parallel. The first is a LightGBM branch, which excels at analyzing panel-temporal summaries—essentially looking at trends over a three-year history. The second is a Panel-MLP branch, which uses attention-based pooling to aggregate temporal data and multi-task learning to understand the relationships between different tax proxies. By combining these two approaches, the model leverages the strengths of both tree-based ensembles and neural networks. The ai search story also surfaces in Stanford AI discovery identifies natural weight..., adding another angle.

Transparency Through Branch Reliance

A key innovation of PaGNet is its "per-target validation-optimal blender." Instead of using a complex, hidden fusion process, the model uses a simple, non-trainable method to combine the outputs of the two branches. This produces a "branch-reliance diagnostic," which acts as a report card for each tax proxy. It tells users whether the model relied more on the tree-based branch or the neural branch to make its prediction. This allows regulators to see exactly which part of the model is driving the forecast for specific tax metrics, satisfying the need for auditable and explainable AI in financial oversight.

Performance and Insights

When tested on the KoTaP dataset—which covers 1,754 Korean firms from 2011 to 2024—PaGNet demonstrated significant improvements in forecasting accuracy for accrual-based tax proxies (TSTA and TSDA), outperforming six baseline models by a notable margin. The researchers also conducted a "panel-flatten" control experiment to determine where these gains originated. They found that while access to multi-year historical data provided the bulk of the predictive power, PaGNet’s specialized architecture provided a consistent, additional refinement that improved performance further. The ai search story also surfaces in Google AI Releases TimesFM 3 for..., adding another angle.

Limitations and Considerations

The authors emphasize that PaGNet is not intended to be a "universally superior" model for all tabular data. Instead, it is a specialized tool for panel-based financial forecasting. The study highlights that while the model’s branch-reliance diagnostic is very stable for accrual-based proxies, it can be more sensitive to specific data splits when predicting noisier metrics like effective tax rates. Consequently, the authors suggest that the model’s diagnostic reports are most reliable when applied to targets with strong, persistent signals, and they caution against treating the model as a general-purpose solution for all financial prediction tasks. The ai search story also surfaces in Qwen Developers Open-Source Local-First Search Layer..., adding another angle. as detailed in the full paper on Arxiv