Summary

Outliers are a fact of any data analysis, but they do not always have to cause headaches. Prophet is very robust at handling most outliers without any special consideration, but sometimes problems can arise. In this chapter, you learned about the two problems most common with outliers in Prophet: uncontrolled seasonality and exploding uncertainty intervals.

In both cases, simply removing the data is the best approach to solving the problem. As long as data exists in other periods of the seasonality cycles for those gaps where data was removed, Prophet has no problem finding a good fit.

You also learned several automated outlier detection techniques, from the basic techniques of Winsorization and trimming, which tend not to work well on time series exhibiting a trend, to the more advanced technique of stacking forecasts and using errors in the first model to remove outliers for the second model.

Finally, you learned how to model outliers as special events, which has...