Effects of Resampled Data on Time Series Forecasting Accuracy
MetadataShow full item record
This thesis will look at time series forecasting of the air pollutants in Beijing, China and the power consumption of an individual household located in Sceaux, France. The forecast will be taken from two classical methods, the Holt-Winters’ methods and the Seasonal Autoregressive Integrated Moving Average (SARIMA) method. The Holt-Winters’ method will be looked at from the additive, multiplicative and damped methods. The SARIMA model will be looked at as a uni-variate model. In this thesis, it will be shown that less complex algorithms, such as the Holt-Winters’ methods can process larger data sets without the need of resampling, and that they outperform more complex algorithms such as SARIMA with complex seasonal data. It will also show that the Holt-Winters’ forecasting accuracy outperforms the uni-variate SARIMA method at each level of resampling. The thesis will also discuss the computational speed in which the algorithms process the data to complete their forecasts and show that the Holt-Winters’ is significantly faster in almost every case. The main focus of this thesis is to show the effects of resampling on forecasting accuracy. It will be shown to be statistically significant in the forecasting accuracy when measured with the symmetric mean absolute percent error metric.