A Multi-Factor Time Series Forecasting Model for Tax Revenue Based on Hybrid Deep Neural Networks

Accurate tax revenue forecasting is essential for formulating sound budgets, optimizing resource allocation, promoting economic growth, and maintaining the stability of public finances. However, numerous influencing factors and complex patterns make tax revenue forecasting a challenging task. In this study, we develop a novel data-driven workflow for forecasting tax revenue in China. Additionally, we construct a multi-factor hybrid deep neural network forecasting model that combines a residual Long Short-Term Memory (LSTM) network with a Backpropagation (BP) neural network. Furthermore, we investigate five major macroeconomic indicators that affect China’s tax revenue using adaptive lasso: the value added of the tertiary industry, the value added of the industrial sector, the value added of industry as a percentage of GDP, total social fixed asset investment, and the total profit of industrial enterprises above the designated size. These indicators can serve as relevant evidence for further tax analyses. Comparative experiments demonstrate that the proposed LSTM-BP forecasting model yields more accurate tax revenue forecasting results.