Monday, April 13 2020
2:30pm - 5:00pm
Masters Presentation
Comparing the Predictive Accuracy of Auto Regressive Integrated Moving Average Models and Long Short-Term Memory Neural Nets

Committee: Santorico (Advisor/Chair), Borgwardt, Austin

series forecasting methods are often compared on empirical datasets, either in forecasting competitions or domain specific studies. Forecasting competitions work well for showing a modelís general performance on data from some domain: financial time series, social demography time series, electricity load in power networks, etc. However, these competitions do not effectively describe data types and attributes that perform well for each method. To do so, requires simulation studies. Here, we use simulated data in order to examine the relative forecasting performance of an Auto Regressive Integrated Moving Average model (ARIMA) and a Long Short-Term Memory model (LSTM) under various conditions. The ARIMA is a widely used classical forecasting model. The LSTM neural net is a newer supervised machine learning algorithm. We examine the effect of sample size, forecast horizon, and linearity of generated time series on the two modelsí forecast performance. For our simulated data, sample size does not seem to affect forecast accuracy. However, for 1-step ahead forecasts, the LSTM neural net performs better on nonlinear data in four out of six cases considered, while the ARIMA model performs better for the linear data in all cases. The difference between the LSTM and ARIMA modelsí performances seems to decrease as the forecast horizon becomes larger.
Speaker:Eric Olberding
Location:See Zoom link from email

Download as iCalendar