An application of exponential smoothing methods to weather related data

Abstract

Exponential smoothing is a recursive time series technique whereby forecasts are updated for each new incoming data values. The technique has been widely used in forecasting, particularly in business and inventory modelling. Up until the early 2000s, exponential smoothing methods were often criticized by statisticians for lacking an objective statistical basis for model selection and modelling errors. Despite this, exponential smoothing methods appealed to forecasters due to their forecasting performance and relative ease of use. In this research report, we apply three commonly used exponential smoothing methods to two datasets which exhibit both trend and seasonality. We apply the method directly on the data without de-seasonalizing the data first. We also apply a seasonal naive method for benchmarking the performance of exponential smoothing methods. We compare both in-sample and out-of-sample forecasting performance of the methods. The performance of the methods is assessed using forecast accuracy measures. Results show that the Holt-Winters exponential smoothing method with additive seasonality performed best for forecasting monthly rainfall data. The simple exponential smoothing method outperformed the Holt’s and Holt-Winters methods for forecasting daily temperature data.