What are events in time-series data and how do we distinguish them from outliers? We can define an event as any incident that disrupts the normal process that generates a time series. This disruption creates volatility in the data beyond the “normal” pattern or variation. An event can have many different forms including holidays, sales and promotions, natural disasters, public health emergencies, sudden economic and geopolitical changes, etc. The difference between events and outliers in historical data is that events are comprehensible and explainable phenomena that we believe have an impact on the variation of our time series.
In forecasting, we use events to improve the quality and accuracy of forecasting models. To incorporate events in the forecasting process, we are creating independent variables in the background, which can be binary or represent a shape. We include them in our algorithms to more accurately predict that pattern or variation into future time periods.
In SAS Visual Forecasting, we recently introduced new capabilities for interactive event management, including creating and managing custom events via an intuitive UI. Forecasters need to specify the type (pulse, level shift, ramp, or temporary change), the occurrence, the duration, and the recurrence of the event. The rest is handled automatically by SAS' forecasting system when generating the forecasts.
During SAS Explore, our forecasting product manager, Joe Katz, presented these new capabilities in detail. Luckily, we made the recording available below to watch on demand. Happy forecasting!
