A Comparative Analysis of Univariate Time Series Methods for Estimating and Forecasting Daily Spam in United States

Spam has been one of the most difficult problems to be addressed since the invention of Internet. Outbound spam can reflect the information security level of an organization as most spam emails are generated by compromised computers. Understanding the trend of outbound spam can help organizations adopt proactive policies and measures toward a more informed decision on resource allocation in terms of security investment. This can also facilitate governments in designing incentive mechanisms such as cyber insurance. Using three years of daily spam volume from the United States, this research investigates the distribution of daily volume and compare univariate time series methods for forecasting. The methods explored include various exponential smoothing methods, auto-regressive integrated moving average (ARIMA), and neutral network auto-regressive (NNAR) modeling. Our evaluation of multiple univariate time series models suggests that Holt’s linear trend models based on exponential smoothing and neural network modeling can lead to more accurate spam prediction. The conclusion and implications are discussed.