Forcasting the Daily Air Temperature in Uppsala Using Univariate Time Series

• Main Author: Aggeborn Leander, Noah
• Format: Others
• Language: English
• Published: Uppsala universitet, Statistiska institutionen 2020
• Subjects: ARIMA; naïve; neural network; SMHI; walk forward validation; Diebold-Mariano test.; Social Sciences; Samhällsvetenskap
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-412996

This study is a comparison of forecasting methods for predicting the daily maximum air temperatures in Uppsala using real data from the Swedish Meteorological and Hydrological Institute. The methods for comparison are univariate time series approaches suitable for the data and represent both standard and more recently developed methods. Specifically, three methods are included in the thesis: neural network, ARIMA, and naïve. The dataset is split into a training set and a pseudo out of sample test set. The assessment of which method best forecast the daily temperature in Uppsala is done by comparing the accuracy of the models when doing walk forward validation on the test set. Results show that the neural network is most accurate for the used dataset for both one-step and all multi-step forecasts. Further, the only same-step forecasts from different models that have a statically significant difference are from the neural network and naïve for one- and two-step forecasts, in favor of the neural network.