Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Bootstrapping Neural Networks for Time Series Forecasting

Löfwander, Simon LU (2018) STAN40 20172
Department of Statistics
Abstract
In this study artificial neural networks (ANNs) are utilized for time series forecasting. Analyzed time series includes two Google AdWords metrics and revenue for an e-commerce. Classes of ANNs included in the analysis are feedforward multilayer perceptron’s and two types of recurrent neural networks (RNNs), specifically Jordan- and Elman neural networks. Since ANNs lack econometric interpretation and yield single-point predictions, focus is put on implementing a parametric bootstrapping to create prediction intervals for the forecasts. The constructed prediction intervals were successful in containing all observations in validation data sets for each of the studied time series. The analyzed time series exhibited similar characteristics,... (More)
In this study artificial neural networks (ANNs) are utilized for time series forecasting. Analyzed time series includes two Google AdWords metrics and revenue for an e-commerce. Classes of ANNs included in the analysis are feedforward multilayer perceptron’s and two types of recurrent neural networks (RNNs), specifically Jordan- and Elman neural networks. Since ANNs lack econometric interpretation and yield single-point predictions, focus is put on implementing a parametric bootstrapping to create prediction intervals for the forecasts. The constructed prediction intervals were successful in containing all observations in validation data sets for each of the studied time series. The analyzed time series exhibited similar characteristics, which resulted in a comparable set of predictor variables in the final models, while they differed individually in terms of best performing ANN-class. (Less)
Please use this url to cite or link to this publication:
author
Löfwander, Simon LU
supervisor
organization
course
STAN40 20172
year
type
H1 - Master's Degree (One Year)
subject
keywords
Artificial neural networks, time series forecasting, feedforward neural networks, recurrent neural networks, Jordan, Elman, Google AdWords, conversions, clicks, bootstrapping, prediction intervals.
language
English
id
8939439
date added to LUP
2018-05-23 13:22:11
date last changed
2018-05-23 13:22:11
@misc{8939439,
  abstract     = {{In this study artificial neural networks (ANNs) are utilized for time series forecasting. Analyzed time series includes two Google AdWords metrics and revenue for an e-commerce. Classes of ANNs included in the analysis are feedforward multilayer perceptron’s and two types of recurrent neural networks (RNNs), specifically Jordan- and Elman neural networks. Since ANNs lack econometric interpretation and yield single-point predictions, focus is put on implementing a parametric bootstrapping to create prediction intervals for the forecasts. The constructed prediction intervals were successful in containing all observations in validation data sets for each of the studied time series. The analyzed time series exhibited similar characteristics, which resulted in a comparable set of predictor variables in the final models, while they differed individually in terms of best performing ANN-class.}},
  author       = {{Löfwander, Simon}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Bootstrapping Neural Networks for Time Series Forecasting}},
  year         = {{2018}},
}