Nelson Kirimi Micheni2025-12-012025-12-012025-10-01http://repository.embuni.ac.ke/handle/123456789/4501Many countries rely on tax revenues to finance their expenditures; thus, forecasting revenue is important in fiscal planning, policy formulation, and fiscal decision-making. There have been ongoing discussions about Kenya’s budget-making process and whether its revenue estimates are realistic. Over the last 13 fiscal years, revenue collections in Kenya have increased tremendously by 263% (from KES 0.707 trillion in 2011/2012 to KES 2.571 trillion in 2024/25). This growth has been attributed to numerous government efforts, including expanding the tax base, changing tax rates, increasing voluntary compliance, and enhancing revenue mobilization. Despite this remarkable revenue growth, the set revenue targets have seldom been met. Both underestimation and overestimation of tax revenue have led to economic instability. For this reason, it is prudent for the country to explore scientific forecasting methods, such as time series analysis, since tax revenue is collected over time. The failure to meet targets by revenue collectors is often due to the setting of higher targets, driven by the ambition to reach a certain percentage of Gross Domestic Product (GDP), as well as inefficiencies in tax administration and inaccurate forecasts. Failure to meet the revenue targets has often led to unmet expenditure commitments, which have led to increased domestic and foreign borrowing in Kenya. The primary aim is to fit a suitable model that can be used in forecasting domestic revenues in Kenya using the SARIMA and HW time series methods, compare their performances, and use them to create a 1-year forecast. The forecasting process involved model identification, model estimation, adequacy testing, and modelling. Secondary data on the domestic taxes collected in Kenya between July 2014 to December 2020 across the various tax heads was used. Based on its minimal AIC=1358.68, BIC=1363.03, and the least forecasting errors (MAPE=7.67, MASE=0.38, and MAE=4,998.15), the SARIMA (0,1,1)(0,1,0)[12] model was preferred compared to the Additive and Multiplicative HW methods. The predictive abilities of the three models were measured using the Diebold-Mariano test and were found to be significantly different. The Model Confidence Set procedure reaffirmed this by eliminating the Additive and Multiplicative HW models and retaining the SARIMA model as the most suitable, hence recommended for forecasting domestic tax revenues. In most months, SARIMA provided more conservative forecasts that were generally closer to actual figures, but sometimes underestimated revenue. Incorporating macroeconomic variables and advanced machine learning techniques could increase the accuracy levels and the model’s reliability, hence can be explored in future studies.enThesis