Calibrating the Rainfall Forecast of the HyBMG Outputs Using Bayesian Model Averaging : A Case Study

Indonesia which is located in tropical region has a unique weather characteristic with rainfall happening seasonally. However, the climate change happened during the last decade has introduced a difficulty in predicting the rainfall events in Indonesia, which then influences the agricultural sector in Indonesia. Jember is one of the districts in Indonesia which has a relatively big contribution to the rice production in Indonesia especially East Java. Therefore, forecasting rainfall in Jember is a crucial work. Several ways have been carried out by BMKG (Agency for Meteorology, Cimatology and Geophysics) Indonesia in order to have a reliable rainfall forecast in Indonesia. One of the approaches applied by BMKG is by using the outputs of HyBMG consisting of four statistical models, where each output is treated as a single forecast. This paper discusses the result of forecasting rainfall in the area of study using the combination of the HyBMG outputs. The combined forecast is also calibrated using two different estimation procedures of Bayesian Model Averaging (BMA) i.e. Expectation Maximization (EM) and Markov Chain Monte Carlo (MCMC). The results show that both approaches have their own strength. BMA-EM is quiet simple to be applied but the forecast performance really depends on the choice of the prior distribution of the rainfall event while the MCMC is quiet flexible but it needs a longer computation time. This paper compares the result of both approaches and the performance of is very competitive. However, both calibration procedures outperform the original forecast generated by HyBMG.