I am using Forecast Studio to create monthly forecasts. I am encountering a problem where the forecast model selection creates an exponential increase for the last three months when the actuals show a basically constant history. Graphically it looks like a 'hockey stick'.
I then drill into that specific series and select another forecast model that better reflects the data. It may have a greater MAPE but it makes much more sense for the long term projection.
Has anyone ever encountered this situation? If so, is there a programmatic or better way to deal with this? My datasets are quite large with many levels of hierarchy.
It seems to me that a trend model is picked by SAS Forecast Studio - which apparently seems to be not appropriate for your data.
Here is what you might want to consider:
a) use a hold-out sample to select the model (see: Project/Forecasting Settings/Model selection)
b) use a selection statistic different to MAPE like SBC or AIC (see: Project/Forecasting Settings/Model selection)
c) run "Level Model Forecast Quality Extreme Table" report from "Reports and Stored Processes" as a post-process and deal with the flagged models interactively.