A Comparative Analysis Of Holts-Winters’ And Neural Network Prediction Models On Annual Bloemfontein’S Precipitation: Risk Aversion

Neural Networks are a series of algorithms that mimic the human brain’s operation to recognise various relationships amongst vast amounts of data. These algorithms are better in predictive analytics than linear regression models due to the use of the hidden layers of neutrons. Failure to accurately predict precipitation pattern may lead major risks agriculture and other heavy water-using economic sectors. The present study’s aim was to aid water resources management sector with accurate annual predicted precipitation values for informed decision making. Prior to predictions, data quality controls were conducted, where outliers in the dataset were detected, removed and replaced by expectation maximum algorithm aided by SPSS computer software. Holt-Winter and Neural Network models were comparatively deployed in the predictions of the annual precipitation in the study area. Evident from these results is that NN provided more accurate results in prediction than HW where MSE and MAE for NN were relatively smaller. It can therefore be concluded that NN produces relatively more accurate results than linear models. The study therefore recommends that researchers in meteorology, agrometeorology, water resources management sectors deploy NN for forecasting climate variables for better informed decision making for protection and mitigation of climate-related risks.