A Machine learning model for revenue forecasting at local government authorities in Kenya - a case of Nairobi City County

Abstract

Revenue forecasting is a key component of the fiscal cycle in local governments. It enables officials to estimate how much revenue they can expect to raise from various revenue sources under their mandate and therefore, what potential they have to meet their annual expenditure requirements. Budgets that are drawn up with poorly forecasted revenue figures are prone to inevitable shortfalls, and consequently, the inability to honor obligations to ongoing development projects, recurrent expenses and service delivery to citizens. Expert judgement is commonly used in forecasting revenue due to its simplistic nature. However, this is not always a reliable method, nor is it backed by factual evidence that may be used to incorporate various other aspects that also influence changes in revenue potential. The objective of this research was to develop a machine-learning (ML) model for revenue forecasting at local government authorities in Kenya. The study adopted an experimental research design to analyze the change in revenue estimates over time. Non-probability sampling was used to review primary data from Nairobi City County, and thereafter, one statistical and six different machine-learning models were evaluated on the same data set. The statistical model used was SARIMAX, which was informed by literature and was used as a baseline to measure the efficacy of an AI approach. The ML models comprised of regressors, tree-based algorithms and a recurrent neural network. They were trained on 14 different revenue streams spanning over 36 months, validated on 9 months of collections and tested over a 6-month spread. The statistical model registered a forecast accuracy of 53.9% on the total revenue collected, while the highest accuracy score on a single revenue stream was attained by the Random Forest model at 79.8% on Land Rates collections. The best overall ML model across all revenue streams was found to be the Ridge Regressor. These ML results outperformed those of the benchmark model and showed the potential that machine learning has in budgeting, revenue forecasting and decision-making processes at the county level. Keywords: Machine Learning, Revenue Forecasting, Local Government, Time Series Analysis