STATE TAX FORECAST WITH USE OF COMBINED UNISORTED MODELS APPLIED TO THE STATE OF PARAÍBA

Abstract

The article aims to present alternative models for efficient forecasting of state taxes. Nine univariate models were tested for tax collection in the state of Paraíba, with data from 1997 to 2020. The models were subject to combinations, including those proposed in Figueiredo (2019), Hyndman and Athanasopoulos (2018) and Shaub and Ellis (2020). The combinations were generated using the following models: Holt-Winters with exponential smoothing by additive treatment; SARIMA (Seasonal Autoregressive Integrated Moving Average); X-13-ARIMA-SEATS; o ETS (Error Trend Seasonal); NNAR (Neural Network Autoregression); TBATS (Exponential Smoothing Method + Box-Cox Transformation + ARMA model for residuals + Trigonometric Seasonal); model STLM (Seasonal and Trend decomposition using Loess THETAM (Theta method of Assimakopoulos and Nikopoulos, 2000); and, SNAIVE (Seasonal Naive). The forecast results were compared using the performance indicators RMSE (Root Mean Squared Error); MAE (Mean Absolute Error); MPE (Mean Percentual Error); and MAPE (Mean Absolute Percentual Error). The findings point to the need for improvement in the preparation of tax forecasts in the state of Paraíba. The incorporation of the use of singular forecast models or combined, can offer gains to forecast formulations and state budget planning and execution.