Forecasting Tax Revenue with Machine Learning and Granger Causality

This paper examines various tax elements that give Nigeria's government a collection basket. This work examines how taxes provide an adequate revenue mechanism and their implications for fiscal policy and macroeconomic stability. To do this, FIRS (Federal Inland Revenue Service) information on various tax components was collected monthly from 2010 to 2021. Our final model that was the best predictor of revenue was chosen after weighing models: multivariate extended short-term memory networks (MLSTM), seasonal autoregressive integrated moving average (SARIMA), multivariate linear regression (MLR) and a hybrid model (MLSTM & XGBoost). Each model was rated on predicted value and accuracy. This multiple independence of prediction factor was considered in the study. Granger Causal relationship between the Dependent and Independent was conducted for hypothesis testing. With increasing numbers of independent variables, the hybrid, MLSTM, and MLR models were better than one another. The hybrid model achieved an R2 value of 99.7%.