Enhancing the Forecasting Accuracy of Direct Tax using a Hybrid Approach and Machine Learning Models

Abstract

Direct taxes have a significant impact on an economy, both in terms of income collection for the government as well as their consequences on individuals, enterprises, and overall economic behavior. Accurate forecasting of direct tax revenue is vital for effective fiscal planning and policy formulation. Traditional forecasting approaches frequently fail to reflect the complex patterns and financial datasets. We applied machine learning and hybrid methodologies to improve the prediction accuracy of direct tax revenue estimates. Univariate time series models, which include the Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing State Space (ETS), and Radial Basis Function Neural Network (RBFNN) model, have been used. Two-stage hybrid models WA-ARIMA, WA-ETS, WA RBFNN, EMD-ARIMA, EMD-ETS, EMD-RBFNN, which are based on denoising and decomposition techniques, i.e., Wavelet Analysis (WA) and Empirical Mode Decomposition (EMD). Three-stage hybrid models including WA-CEEMDAN-ARIMA, WA-CEEMDAN ETS, and WA-CEEMDAN-RBFNN are created based on the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) technique, and these models are applied using both WA and EMD. Comparison is made on the basis of Mean Square Error (MSE). Hence, in terms of prediction accuracy of direct tax, the WA-CEEMDAN-ARIMA has the lowest MSE value as compared to all other existing models. Hence, results show that WA-CEEMDAN-ARIMA is proved to be appropriate for this time series data