Using R Prophet across multiple groups

Question

I've used Prophet previously to forecast individual metrics, but I now need to forecast into 2022, using around 3 years of historic data and including multiple groups. I'd rather not create 1000+ forecasts for each possibility, so hoping I can do this within Prophet.

I've read through a couple of solutions on Stackoverflow which work for a singular group, but I have 3 different groups to forecast.

Below I've created a sample table, with dummy Y data showing how the data is formatted: (Note: There can and will be zero values, but I can exclude them if it'd negatively affect the outputs)

ds	establishment	category	channel	y
2020-01-01	High School	Stationery	Direct	27
2020-01-01	High School	Stationery	Paid Search	31
2020-01-01	High School	Arts	Direct	47
2020-01-01	High School	Arts	Paid Search	0
2020-01-01	College	Stationery	Direct	60
2020-01-01	College	Stationery	Paid Search	35
2020-01-01	College	Arts	Direct	54
2020-01-01	College	Arts	Paid Search	15
2020-01-02	High School	Stationery	Direct	27
2020-01-02	High School	Stationery	Paid Search	31
2020-01-02	High School	Arts	Direct	47
2020-01-02	High School	Arts	Paid Search	0
2020-01-02	College	Stationery	Direct	60
2020-01-02	College	Stationery	Paid Search	35
2020-01-02	College	Arts	Direct	54
2020-01-02	College	Arts	Paid Search	15
2020-01-03	High School	Stationery	Direct	27
2020-01-03	High School	Stationery	Paid Search	31
2020-01-03	High School	Arts	Direct	47
2020-01-03	High School	Arts	Paid Search	0
2020-01-03	College	Stationery	Direct	60
2020-01-03	College	Stationery	Paid Search	35
2020-01-03	College	Arts	Direct	54
2020-01-03	College	Arts	Paid Search	15
...	...	...	...	...

The following code works for a single group, but I'd like to expand this to include all groups.

d1 <- df %>%
  nest(-establishment) %>%
  mutate(m = map(data, prophet)) %>%
  mutate(future = map(m, make_future_dataframe, period = 730)) %>%
  mutate(forecast = map2(m, future, predict))

d <- d1 %>%
  unnest(forecast) %>%
  select(ds, establishmentShortcut, yhat)

If anybody can recommend a solution, I'd appreciate it.

Just thought I'd drop a word of caution about using prophet, as described nicely in [this article](https://medium.com/geekculture/is-facebooks-prophet-the-time-series-messiah-or-just-a-very-naughty-boy-8b71b136bc8c). If you're aware of this already, I apologize and feel free to ignore! — Daniel Molitor, Aug 06 '21 at 15:16

score 1 · Answer 1 · answered Aug 07 '21 at 01:16

This is what the fable package is designed to do. There is a fable extension for handling prophet models called fable.prophet. Here is an example using the same structure as your data.

library(fable.prophet)
library(dplyr)
library(tibble)
library(tsibble)

df <- tibble(
    ds = rep(seq(as.Date("2020-01-01"), by="1 day", length=50), rep(8,50)),
    establishment = rep(rep(c("High School","College"), c(4,4)), 50),
    category = rep(rep(c("Stationery","Arts"), c(2,2)), 100),
    channel = rep(c("Direct","Paid Search"), 200),
    y = sample(0:100, 400, replace=TRUE)
  ) %>%
  as_tsibble(index=ds, key=c("establishment","category","channel"))

fc <- df %>%
  model(
    prophet = prophet(y),
  ) %>%
  forecast(h=10)

^{Created on 2021-08-07 by the reprex package (v2.0.1)}

You can swap out prophet() for any of the other models in the fable package. See https://otexts.com/fpp3/ for details.

This looks very promising Rob, but I'm struggling with the example given. I'm getting the error: "Error in prophet(y) : object 'y' not found." I've tried directly referencing the column, but get an invalid 'pos' argument. — Ed Cunningham, Aug 09 '21 at 13:25
Use the most recent versions of packages and run the code in a new session. Possibly you have old packages, or perhaps you have clashes with other loaded packages. — Rob Hyndman, Aug 09 '21 at 23:13

Using R Prophet across multiple groups

1 Answers1