Interpreting Prophet Output for Weekly and Yearly Seasonality

Question

I am going through the tutorial for using prophet in R.

You can find the dataset here: https://github.com/facebookincubator/prophet/blob/master/examples/example_wp_peyton_manning.csv

# R
library(prophet)
library(dplyr)

df <- read.csv('peyton.csv') %>%
  mutate(y = log(y))

head(df)

          ds        y
1 2007-12-10 9.590761
2 2007-12-11 8.519590
3 2007-12-12 8.183677
4 2007-12-13 8.072467
5 2007-12-14 7.893572
6 2007-12-15 7.783641

df$ds<-as.Date(df$ds,'%m/%d/%Y')

m <- prophet(df)

future <- make_future_dataframe(m, periods = 365)

forecast <- predict(m, future)

plot(m, forecast)

prophet_plot_components(m, forecast)

The output for prophet_plot_components(m, forecast) is below:

Do I interpret this plot for the yearly seasonality portion as:

Whatever your prediction is for a specific date, increase or decrease it by so-and-so amount to account for yearly seasonality? For example, it looks like on April 1st, y is expected to be -0.5. How do I use this result? Do I take the average y for the year and subtract it by -0.5 to account for seasonality? Alittle confused.

Any help would be great!

Answer 1

@Nick Prophet predict() function has following syntax:

temp_dataframe_to_store_prediction <- predict(model_name, new_dataframe_created)

In your case:

forecast <- predict(m, future)

The predictions or forecasts from the predict() function includes trend and seasonality for each row i.e. for each time_stamp. Therefore, you need not include any separate line of code for including or excluding seasonality in the final forecast values. Your interpretation is correct i.e. final forecast values is attained by adding(+/-) values of seasonality and trend to the stationary "y" value. Please note, the "y" value used by prophet in your code is log(original_y)

 mutate(y = log(y))

This line makes the original_y value stationary. Therefore, in order to interpret the final results you need to take exponential of forecasted "y" value. This brings the forecast to original scale. No need to take exponential if original_y was not transformed.