University of Wisconsin Extended Campus is now Wisconsin Online Collaboratives! This name reflects the partnerships of the 13 universities within the Universities of Wisconsin–our state's premier system of public higher education. Through these partnerships we will continue to support online degrees, certificates and courses–along with support services to you.

Capstone Projects

Comparison of Recent Univariate Time Series Models Released by Big Tech Companies

Program: Data Science Master's
Location: Not Specified (remote)
Student: Nicole Temple

Recently, several big tech companies have open-sourced new univariate timeseries methods. Although these newer methods have quickly risen in popularity and use, a review of the literature shows no instance where the accuracy of the following univariate time series methods were compared:  DeepAR from Amazon’s GluonTS package, Prophet from Facebook’s Kats package, and Silverkite from LinkedIn’s Greykite package. Comparing these new methods with other well-established methods is critical so analysts can make data-driven decisions when choosing the best method for performing time series forecasts. This study compares the accuracy of the three previously mentioned forecasting methods along with three well-established methods, including, Auto Regressive Integrated Moving Average (ARIMA), Holt-Winters Exponential Smoothing and XGBoost with the purpose of providing suggestion on which methods to use based on trend and seasonality of the data. For this study, each of the six methods was performed on four different synthetically created datasets varying on seasonality and/or trend. Performance was evaluated using root mean square error (RMSE). This study found that XGBoost performed the best on datasets without a trend component, while Holt-Winters performed the best on datasets with a trend component. Silverkite was the most consistent top performer and did well with volatility, seasonality, and trend components. Despite Prophet’s and DeepAR’s popularity, both methods consistently failed to perform well.