Modelling and Forecasting Volatility of Value Added Tax Revenue in Kenya

Taxation is one of the means by which governments finance their expenditure by imposing charges on citizens and corporate entities. Kenya Revenue Authority (KRA) is the agency responsible for the assessment, collection and accounting for of all revenues that are due to government. Volatile government revenue is a challenge for fiscal policy makers since it creates risks to government service provision and can make planning difficult, as revenue falls short of expenditure needs both frequently and unexpectedly. The main objective of this study was to model and forecast the volatility of VAT revenue collected in Kenya as well as computing its value at risk and the expected shortfall. Secondary data on daily VAT revenue collections for a period of 3 years was analyzed. The first step was to model the mean equation of the return series using the ARIMA model and ARIMA(3,0,3) was identified to be the most suitable since it had the least values of AIC and BIC. The Lagrange Multiplier test confirmed the presence of ARCH effects using the residuals of the mean equation. A number of heteroscedastic models were fitted and the TGARCH family (ARIMA(3,0,3)/TGARCH(1,2)) was preferred to fit the volatility of the returns. One step ahead forecasting of volatility of the returns was done using the model which gave a value of 7.212. Estimation of value at risk and expected shortfall involved use of POT method by fitting a GPD function to the return data series. The first step was determination of threshold by use of MRL plot and later fitting a GPD function to the return data series using the threshold. The shape, location and scale parameters were estimated using MLE and they were later used to compute the VaR loss and ES at 95% and 99% confidence intervals. The VaR at 95% and 99% was 1.45% and 1.49% respectively while the ES at both the intervals was 0.04% and 0.1% respectively. This study concluded that volatility is persistent in the daily VAT revenue collections and it can easily be modelled using conditional heteroscedastic models.