2

I am using an additive Facebook Prophet model for predicting forecasts. On my test data it seems to be understanding the spikes well, but the overall trend is completely wrong, based on the last few data points going down:

enter image description here

Is there a way to level the trend line? I have adjusted changepoint_prior_scale many times, the current settings are:

changepoint_prior_scale = 0.25, changepoint_range=0.85

I am surprised that it continues to project a negative trend rather than a level trend, since the forecast window is quite long compared to the regularity of the major changepoints. Does Prophet always forecast trend so linearly? Is there a way to adjust this?

Darcey BM
  • 301
  • 1
  • 5
  • 20
  • Are you dealing with seasonal patterns? Also I see some big outliers. You could use the holidays functionality to account for them. – tvgriek Aug 23 '20 at 16:42
  • Hi there, I am using the holiday function to account for relevant holidays, but I have added those particular outlier dates as regressors (they are from promotions etc), and I want the model to be able to distinguish between the impact of diff types of promotions, hence why I have these one hot encoded as separate columns. Would you mind expanding on what you mean by "dealing with seasonal patterns"? Currently I am not but would gladly add this in – Darcey BM Aug 23 '20 at 18:54

0 Answers0