search

facebook / prophet

Tool for producing high quality forecasts for time series data that has multiple seasonality with linear or non-linear growth.
https://facebook.github.io/prophet
MIT License
18.26k stars 4.51k forks source link

Robustness of facebook prophet #1969

Open AliWaheed opened 3 years ago

AliWaheed commented 3 years ago

Background I trained/tested a timeseries of 36 months in which I use 6 months for testing thus, training period is 32 months.

Issue I performed two experiments on the same Data Frame with one single change i.e. I added - 0.0000000001 - randomly to the training months. To my surprise the forecasts came out very different. I did this with a small sample set and achieved an average of 23% differences in the sum of the forecasted values (take this number with a grain of salt). I've also seen cases with almost 50% difference.

Question Why such a minuscule number when added to a training data value. Changes the forecast drastically?

AliWaheed commented 2 years ago

Is this an example of chaos theory? If so, does that mean that time series solutions are in general susceptible to Chaos, or is it just fbprophet's method that is susceptible to Chaos or, is it just a bug and I am way over my head?