Predicting Activity Levels in Virtual Communities

Abstract

Data was collected from two communities: a smaller community with approximately 200 participants, but where the number of participants is precisely known; and a very large community, with thousands of participants, where the number of participants can only be estimated from the number of different nicknames used within a given time interval. Data from certain days when there were documented technical issues that may have affected activity has been removed from the sample. In both cases, we have taken one geographically centric data series and one global series, covering users in multiple areas and time zones. We use the number of messages sent in three-hour intervals as a proxy for the activity level in a community, as accurate figures regarding the number of messages viewed by unique persons are difficult to establish. This results in a data set of approximately 9,500 samples from each community, collected over a period of just less than four years. When fitting the models, we used accepted back-testing standards, relying on a fixed interval (one year) when fitting parameters and forecasting activity for any given point in time. One exception to this is seasonality adjustment, where we used the entire data set—this should not have a significant effect, as we made the same seasonal adjustment to the input for all models. Empirical results provided throughout the article are based on data from the larger community.