Genetic Algorithm-Optimized Neural Networks: An Intelligent Solution for Tax Revenue Prediction

The open-source project introduced in this article was published by Aman-K-Mishra on GitHub (project name: Tax-Revenue-Prediction-GA-NN). Its core is combining genetic algorithms (GA) with neural networks to automatically search for optimal network architectures, solving the problem of tax revenue prediction based on macroeconomic indicators—especially suitable for small-sample nonlinear data scenarios, providing a practical example for machine learning modeling.

Tax revenue prediction is an important foundation for fiscal planning and policy formulation. Traditional econometric methods assume linear relationships, but the relationships between tax revenue and factors like GDP, inflation rate, population, import-export trade, and corporate tax rates are complex and nonlinear. Although neural networks can capture nonlinearity, they face small-sample challenges: only 129 annual observation samples and 6 macroeconomic features—complex architectures are prone to overfitting, requiring intelligent methods to automatically select optimal architectures.

Reasons for choosing genetic algorithms: suitable for scenarios with large search spaces and discrete parameter optimization. The workflow includes: 1. Feature standardization using StandardScaler; 2. GA evolution of candidate architectures; 3. Fast evaluation using validation set MSE; 4. Retraining of the best architecture; 5. Saving the model and scaler; 6. CLI prediction interface. The input features are 6 macroeconomic indicators: GDP, inflation rate, population, import-export, and corporate tax rate.

After GA search optimization, the final architecture is a feedforward neural network with two hidden layers. The performance metrics are as follows: R²=0.827, RMSE=69.5k, MAE=55.4k, MSE=4.8 billion—indicating that the model captures the relationships between variables well and reaches a practical prediction level.

The project structure includes folders like data/, models/ (for saving models and scalers), and files like predict.py, train.py, ga_search.py, etc. Core components: ga_search.py implements GA search logic, train.py handles the training process, predict.py provides a CLI prediction interface. Tech stack: Python, PyTorch, NumPy, Pandas, scikit-learn.

Value of automated architecture design: reduces manual intervention, avoids local optima, adapts to small data. Fiscal domain applications: budget planning, policy simulation, risk early warning. Educational and research value: demonstrates the combination of evolutionary algorithms and deep learning, provides a complete workflow, suitable for machine learning practice projects.

Current limitations: small dataset size, only annual data used, limited feature dimensions. Future improvement directions: use real government data, add cross-validation, try LSTM/Transformer time-series prediction, build a web dashboard, support CSV batch prediction.