Not logged in.

Contribution Details

Type Bachelor's Thesis
Scope Discipline-based scholarship
Title Implementation and Evaluation of Forecasting using Variants of Exponential Smoothing
Organization Unit
Authors
  • Piero Neri
Supervisors
  • Michael Hanspeter Böhlen
  • Jamal Mohammed
Language
  • English
Institution University of Zurich
Faculty Faculty of Business, Economics and Informatics
Date 2021
Abstract Text A time series is a sequence of measurements, typically occurring at equispaced time intervals. Two time series were studied in this thesis consisting of temperature and humidity values in the city of Zurich in 2019. Decomposition methods are used to break down a time series into suitable components for exploring the characteristics of the time series which give us an insight on selecting an appropriate predictive model or whether model assumptions are met. A classical decomposition method was implemented and its application to the data sets revealed the presence of a double seasonal component (i.e., a repetitive oscillatory pattern), corresponding to the alternation of nights and days embedded within a yearly seasonal component. This thesis focused on the implementation of various exponential smoothing methods in order to make meteorological predictions based on past values. The obtained performance of the forecasts of the exponential smoothing methods were compared to the performance of a commercial package of ARIMA models. The time series of temperature and humidity values were subdivided into 49 four-week time intervals, each shifted by one week in order to cover the whole 52 weeks of 2019. Within each four-week interval, the first three weeks were used as "training set", in order to identify the exponential smoothing methods and ARIMA models that best fitted the data, respectively. The training sets were also used to optimize parameters and to obtain the best fit for each method. The remaining last week of each four-week interval was used as "test set" for predictions, based on the selected exponential smoothing methods and ARIMA models. The accuracy of forecasts was measured by comparing predicted temperature and humidity values against the values of the test set. Forecasts were performed for future values with horizons ranging between 1 and 24 hours. In order to cover the whole test set with predictions, the first hour of the test set was iteratively transferred to the training set, allowing to cover the whole one-week interval with forecasts. Suitable metrics were used in order to describe the accuracy of predictions for different time horizons within each interval. Eventually, all metrics for all intervals were averaged, in order to facilitate a more global analysis of performance. Predictions were particularly good at short term horizons (i.e., few hours into the future). The approach of identifying the best fitting parameters and the best fitting methods described in the thesis could be applied to many univariate time series of interest. Moreover, the optimization strategy which was used, featuring a mix of grid search and random search approach, reduced the computation time without sacrificing prediction performance quality: an important feature for large time series.
PDF File Download
Export BibTeX