Abstract
Revenue forecasting accuracy is critical to governmental operations. This paper addresses the question: What is the best technique for forecasting sales tax revenue? Prior studies in this area have focused on the differences between machine learning techniques and traditional approaches and neglected to consider how differences in pre-processing steps for the data before the forecasting model is applied are important. Here, we show that machine learning techniques do not always provide increased forecasting accuracy. Instead, the modeling choices matter, but less than the prior literature and practice suggested. Rather, pre-processing makes the most significant difference in forecasting accuracy, and forecasters need to understand the unique characteristics of time series data to improve forecasting performance. The immediate implications of these findings are that the focus of practitioners of in sales tax revenue forecasting should shift from prioritizing model choice towards data pre-processing.