Correcting outliers that cause wide uncertainty intervals

In the first type of outlier we looked at, the problem was that the seasonality was affected and forever changed yhat in the forecast (if you remember from Chapter 2, Getting Started with Facebook Prophet, yhat is the predicted value for future dates contained in Prophet's forecast DataFrame). In this second problem, yhat is minimally affected but the uncertainty intervals widen dramatically.

To simulate this issue, we need to modify our NatGeo data a bit. Let's say that Instagram introduced a bug in their code that capped likes to 100,000 per post. It somehow went unnoticed for a year before being fixed, but unfortunately, all likes above 100,000 were lost. Such an error would look like this:

Figure 9.6 – Capped likes on National Geographic's Instagram account

You can simulate this new dataset yourself with the following code:

df3 = df.copy()
df3.loc[df3['ds'].dt...