plyr + forecasting multiple regressors - r

Question

Taking the content in this thread a bit further: I've gone as far as I can, but finally hit a wall. I'm looking to use PLYR to create some ARIMA models with exogenous regressors at scale. A high-level overview of the process I've been using (code with example data follows)
1) I have a dataframe with businesses, regions, revenue and orders, all by date
2) For each combination of business + region, I want to create a forecast for revenue based on previous values of revenue + previous values of orders.
3) I want to use an ARIMA model (using auto.arima() ) to figure out optimal orders for both revenue and orders, then apply that information to a forecast function
4) The problem I run into seems to boil down to not being able to pass multiple lists to a PLYR argument to operate on, which most likely in turn boils down to my not fully understanding how llply works (so hopefully this is an easy task)
Here's some sample data I'm working off:
library(plyr)
library(xts)
library(forecast)
data <- data.frame(
biz = sample(c("telco","shipping","tech"), 100, replace = TRUE),
region = sample(c("mideast","americas","asia"), 100, replace = TRUE),
date = rep(seq(as.Date("2010-02-01"), length=10, by = "1 day"),10),
revenue = sample(1:100),
orders = sample(1:100)
)
Edit: First, reorganize data through ddply to get rid of duplicate entries:
dataframe <- ddply(data, c("biz","region","date"), function(df) {
c(revenue = sum(df[,4]),
orders = sum(df[,5]))
})
Step 1: Create a list that contains the time series info for each combination of business + region:
list1 <- dlply(dataframe, .(biz,region), identity)
Step 2: Turn that list into an XTS object so we can use it for time-series analysis:
xtsobject <- llply(list1, function(list) {
xts(x=list[,c("revenue","orders")], order.by=list[,"date"])
})
Here's where I run into trouble. I want to make a list of orders from the auto.arima() function to pass into a forecast.Arima() function. This would be straightforward if I were just doing one variable with no exogenous regressors:
arimamodel1 <- llply(xtsobject, function(list) {
fity <- auto.arima(list$revenue)
})
And then I would apply that list to the forecast.Arima() function:
forecast1 <- llply(arimamodel1, function(model) {
forecast.Arima(model, h=2)
})
That comes out fine. I've tried changing the argument to include some room for the extra regressors, but I'm not sure the forecasts are actually pulling in the x values:
arimamodel2 <- llply(xtstest, function(list) {
fity <- auto.arima(list$revenue, xreg=list$orders)
fitx <- auto.arima(list$orders)
})
and the forecasts:
forecast2 <- llply(arimamodel2, function(model) {
forecast.Arima(model, h=2)
})
... But it seems like in the forecast function, I should be doing something to account for the x regressor model in the way I normally use forecast.Arima() with multiple regressors; something like:
forecast.Arima(model,h=2, xreg=forecast(model,h=2)$mean)
But this doesn't work. Does anybody have any insight into how to use PLYR to make forecasts based on auto.arima() for multiple regressors?

Answer 1

I'm pretty sure I figured this out, in case anybody stumbles on to this question. It's just a matter of making a function that passes through all these arguments, then passing that function through lapply or llply (the data in the question won't work for auto.arima because of the way it was created, but it works on the actual data I'm using):
arimafunc <- function(list) {
fity <- auto.arima(list$revenue, xreg=list$orders)
fitx <- auto.arima(list$orders)
forecast <- forecast.Arima(fity,h=2,xreg=forecast(fitx,h=2)$mean)
return(forecast)
}
then pass through the list apply:
forecasts <- lapply(xtsobject,FUN=arimafunc)
I'm sure there's a way to do this using built-in functionality of something like llply or from one of the base commands, mapply, but this works for now...