Monday, April 25 2022
2:00pm - 4:00pm
Masters Presentation
A Brief Comparison of Time Series Forecasting Methods: From Classical to Deep Learning Models

Abstract: "Time series forecasting has historically involved classical techniques that use traditional statistical methods to gain insights. These techniques persist in the modern age because of their robustness, accuracy, and computationally cheap implementation. Moreover, the increasing size and complexity of data has brought about the need to explore information using alternate methods, namely those of machine learning and deep learning. The heart of the debate between these methods is whether or not the machine and deep learning methods can outperform the classical statistical ones. The newer deep learning methods offer the promise of higher accuracy despite their higher computational complexity. However, research shows that the machine and deep learning models fail to outperform the traditional statistical techniques. The motivation of this project is to empirically replicate these results to either support or weaken the claim that classical statistical methods outperform machine learning and deep learning methods. In this paper, select time series forecasting models from three major classes of models are compared, including classical statistical techniques (exponential smoothing, ARIMA), basic machine learning methods (support vector regression, classification and regression trees, K-nearest neighbors, Gaussian process, multilayer perceptron), and deep learning neural networks (recurrent neural network, long short-term memory, gated recurrent unit). Two datasets of time series are used to test the forecasting performance of each model. The performance is measured in terms of prediction accuracy and computational efficiency, and the models are compared both among and within the sets of method classes. Additionally, as a secondary objective, the obtained performance metrics are compared between the two datasets to observe how certain models perform for different types of data (e.g., traditional seasonal data vs. nontraditional volatile data). The datasets include average hourly air temperature data obtained from the National Oceanic and Atmospheric Administration (NOAA) and hourly cryptocurrency data obtained from bitcoin-to-dollar (BTC/USD) spot prices. "
Speaker:Jadon Costa
Location:SCB 4119

Download as iCalendar