Forecasting of Rainfall in Kota Bharu Station Using Holt-Winters, SARIMA and Time Series Regression Model

Abstract

Rain plays a vital role as a component of the hydrologic cycle in various aspects of life. However, continuous heavy rainfall can exacerbate other natural disaster problems, and it is essential to accurately forecast rainfall amounts, especially in areas that consistently experience heavy rain. This study focuses on comparing the forecast accuracy of three methods such as Holt Winters Exponential Smoothing (HWES), Seasonal Autoregressive Integrated Moving Average (SARIMA) and Time Series Regression (TSR) models, for monthly rainfall amount data at Kota Bharu Station. Based on the results, the SARIMA model was chosen as the best model and the most suitable method for forecasting rainfall amount compared to the HW and TSR models, as it presents lower values of Mean Absolute Deviation (MAD), Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE). The inconsistent trend of the data made it difficult for HW to adjust the level, trend and seasonal components effectively. Even though the TSR model was able to capture the seasonality patterns by using dummy variables, the linearity of the model and the inconsistent fluctuations of the data make the model less flexible in capturing sudden changes or extreme rainfall values. By incorporating seasonal differencing, seasonal Autoregressive (AR) and seasonal Moving Average (MA) components, the SARIMA model can capture both the seasonal patterns and fluctuations in rainfall amount over time. With a MAPE value of 40.5, the SARIMA model also exhibits reasonable forecasting performance in terms of MAPE interpretation and outperforms other methods. Given the limitations of this study, more robust error metrics, such as Mean Arctangent Absolute Percentage Error (MAAPE) or Mean Absolute Scaled Error (MASE) are recommended for future research especially when rainfall data contain minimal values near 0. Additionally, including other climate-related factors such as temperature, humidity, and wind speed, can further improve forecast accuracy.