Facebook Prophet Tool: Hyperparameter Tuning on Monthly Data

[sheidari]

I am using the Prophet tool to forecast revenue for my company and one of the challenges i currently face is being able to modify the code in order to leverage the hyperparameter tuning features for monthly data. The tool has the option to select auto tuning (HPO) but it doesn't work with monthly data.  However, I have read somewhere (can't seem to find the post) where it can be tweaked for monthly data.

Has anyone been able to figure this out? Would love some help! I'm not a programmer and have been leveraging Alteryx to build this out so would really appreciate a fellow coder's help in solving this issue!

Here's the code from the macro that leverages Bayesian optimization for the auto tuning:

# Conditional Install
cond.install <- function(package.name){
options(repos = "http://cran.rstudio.com") #set repo
#check for package in library, if package is missing install
if(package.name%in%rownames(installed.packages())==FALSE) {
install.packages(package.name, .libPaths()[2])}else{require(package.name, character.only = TRUE)}}

# conditionally install package
cond.install('forecast')
cond.install('prophet')
cond.install('rBayesianOptimization')
cond.install('dplyr')
cond.install('lubridate')


library(dplyr)
library(lubridate)
library(forecast)
library(prophet)
library(rBayesianOptimization)



#reading data
cv_set <- read.Alteryx("#1", mode="data.frame")
valid <- read.Alteryx("#2", mode="data.frame")

#make sure the date format is defined

cv_set$ds <- as.Date(cv_set$ds)
date_seq <- as.Date(valid$ds)

#define hyper search parameter
rand_search_grid =  data.frame( 
  changepoint_prior_scale = sort(runif(10, 0.01, 20)),
  seasonality_prior_scale = c(sort(sample(c(runif(5, 0.01, 0.05), runif(5, 1, 20)), 5, replace = F)),
                              sort(sample(c(runif(5, 0.01, 0.05), runif(5, 1, 20)), 5, replace = F))),
  n_changepoints          = sample(5:50, 10, replace = F)

)

#Define deafult function for prophet. Change Linear to Logistic cap setting
prophet_fit_bayes = function(changepoint_prior_scale, seasonality_prior_scale, n_changepoints) {
    
    error = c()
    for (d in date_seq) {
      train = subset(cv_set, ds < d)
      test  = subset(cv_set, ds == d)
    
      m = prophet(train, growth = 'linear',
                 seasonality.prior.scale = seasonality_prior_scale, 
                 changepoint.prior.scale = changepoint_prior_scale,
                 n.changepoints = n_changepoints,
                 weekly.seasonality = F,
                 daily.seasonality = F)
    
      future = make_future_dataframe(m, periods = 1)
    
    # NOTE: There's a problem in function names with library(caret)
      forecast = predict(m, future)
      forecast$ds = as.Date(forecast$ds)
    
      error_d = forecast::accuracy(forecast[forecast$ds %in% test$ds, 'yhat'], test$y)[ , 'MAPE']
      error = c(error, error_d)
  }
  
    ## The function wants to _maximize_ the outcome so we return 
    ## the negative of the resampled MAPE value. `Pred` can be used
    ## to return predicted values but we'll avoid that and use zero
    list(Score = -mean(error), Pred = 0)
}
  
changepoint_bounds    = range(rand_search_grid$changepoint_prior_scale)
n_changepoint_bounds  = as.integer(range(rand_search_grid$n_changepoints))
seasonality_bounds    = range(rand_search_grid$seasonality_prior_scale)

bayesian_search_bounds = list(changepoint_prior_scale = changepoint_bounds,
                              seasonality_prior_scale = seasonality_bounds,
                              n_changepoints = as.integer(n_changepoint_bounds))


#rBayesian parameters. Assume n_iteration is 1 for demo purpose
ba_search = BayesianOptimization(prophet_fit_bayes,
                                    bounds = bayesian_search_bounds,
                                    init_grid_dt = rand_search_grid, 
                                    init_points = 1, 
                                    n_iter = %Question.iteration.var%,
                                    acq = 'ucb', 
                                    kappa = 1, 
                                    eps = 0,
                                    verbose = TRUE)


best_params_ba  = c(ba_search$Best_Par)



#Start Prophet 

# Holiday Setting
custom1 <- data_frame(
  holiday = 'custom1',
  ds = as.Date(c('1991-12-31')))

custom2 <- data_frame(
  holiday = 'custom2',
  ds = as.Date(c('1992-12-31', '1993-01-01')))
holidays <- bind_rows(custom1, custom2)


if ('%Question.noholiday.var%' == "True") {
    m = prophet(cv_set, growth = 'linear',
                 seasonality.prior.scale = best_params_ba[['seasonality_prior_scale']], 
                 changepoint.prior.scale = best_params_ba[['changepoint_prior_scale']],
                 n.changepoints = best_params_ba[['n_changepoints']])
}
if ('%Question.holiday.var%' == "True") {
    m <- prophet(holidays = holidays, growth = 'linear',
                 seasonality.prior.scale = best_params_ba[['seasonality_prior_scale']], 
                 changepoint.prior.scale = best_params_ba[['changepoint_prior_scale']],
                 n.changepoints = best_params_ba[['n_changepoints']])
    m <- add_country_holidays(m, country_name = '%Question.country.var%')
    m <- fit.prophet(m, cv_set)
}

future <- make_future_dataframe(m, periods = %Question.forecast.var%)
forecast <- predict(m, future)

yhat <- as.data.frame(forecast$yhat)
yhat_l <- as.data.frame(forecast$yhat_lower)
yhat_u <-as.data.frame(forecast$yhat_upper)
trend <- as.data.frame(forecast$trend)

df1 <- cbind(yhat, yhat_l, yhat_u, trend)

write.Alteryx(df1, 1)

AlteryxGraph(3, width=576, height=576)
plot(m, forecast) + add_changepoints_to_plot(m)
invisible(dev.off())


AlteryxGraph(4, width=576, height=576)
prophet_plot_components(m, forecast)
invisible(dev.off())


#Output best params for reference
df5 <- best_params_ba

write.Alteryx(df5, 5)