To model and forecast monthly tax revenue collection, and propose policy recommendations for stabilization of revenue collection volatility are our objectives in this study. Data analysis of the evolution of tax revenues shows a slight growth and the Box.com test gives |λ|=0.07 which leads to a check logarithmic transformation. The study revealed a positive trend in data with an amplitude of variations increasing slightly over time. It showed also a quarterly seasonality, highlighting a peak in each third month of each year and a relative increase in the first two quarters. The correlogram and partial correlogram show respectively the lag of orders 3 and 12 and 3 and 9 which was significant and lead to conclude the seasonal autoregressive coefficient. Then, the Dickey-Fuller test (p-value < 0.01) and Phillips Perron test (p-value < 0.01) come up to confirm the stationarity of the series under study. ARIMA (0,0,0) (2,1,2) [4] and ARIMA (0,0,0) (2,1,1) [4] models have an AIC of 359.44 and 359.43 respectively, which is smaller than the other models and have been selected for analyzing and predicting the tax revenues of Burundi. But specifically, the study used ARIMA (0,0,0) (2,1,2) [4] which is better than ARIMA (0,0,0) (2,1,1) [4] because of its least parameters. The study thought that this model is recommendable for this institution, which supports the governmental constraints.