Accounting for Outliers and Special Events
An outlier is any data point that lies significantly away from other data points along one or multiple different axes. Outliers may be incorrect data, resulting from a miscalibrated sensor producing invalid data, or even a finger slip on the keyboard during data entry, or they can be accurately recorded data that happens to wildly miss historical trends for various reasons, such as whether a tornado passed over a wind speed sensor.
These uncharacteristic measurements will sway any statistical or machine learning model, so correcting outliers is a challenge throughout data science and statistics. Fortunately, Prophet is generally robust at handling mild outliers. With extreme outliers though, there are two problems Prophet can experience – one problem with seasonality and another with uncertainty intervals.
In this chapter, you’ll see examples of both of these problems and learn how to alleviate their effects on your forecast...
