tslm.R

# Rossmann drugstores competition in Kaggle, 14 October 2015
# This scripts fit a linear model to each store time series (Sales) including trend and seasonality components
# We use tslm() and forecast() in the "forecast" R package 

library(data.table)
library(forecast)
library(plyr)

train <- fread("./train.csv",stringsAsFactors = T)
test  <- fread("./test.csv",stringsAsFactors = T)
store <- fread("./store.csv",stringsAsFactors = T)

train <- merge(train,store,by="Store")
test <- merge(test,store,by="Store")

train[,Date:=as.Date(Date)]
test[,Date:=as.Date(Date)]

train <- train[order(Store,Date)]
test <- test[order(Store,Date)]

ts_fit = function(x) {
  Sales <- ts(x$Sales, frequency = 365)
  DayOfWeek <- x$DayOfWeek
  Open <- x$Open
  Promo <- x$Promo
  StateHoliday <- x$StateHoliday
  SchoolHoliday <- x$SchoolHoliday
  fit <- tslm(Sales ~ trend + season + DayOfWeek + Open + Promo + StateHoliday + SchoolHoliday)
  return(fit)
}

out <- dlply(train, .(Store), ts_fit)

ts_forecast = function(x,y){
  index <- x$Store[1]
  fitt <- y[[index]]
  return(data.frame(forecast(fitt, newdata = data.frame(DayOfWeek = x$DayOfWeek, 
                                                        Open = x$Open, Promo = x$Promo, 
                                                        StateHoliday = x$StateHoliday, 
                                                        SchoolHoliday = x$SchoolHoliday))))
}

predictions <- ddply(test, .(Store), ts_forecast, out)
predictions$Point.Forecast <- ifelse(predictions$Point.Forecast < 0, 0, predictions$Point.Forecast)

Avg_Sales <- train[,.(AS = mean(Sales,na.rm=T)),.(Store,DayOfWeek)]
test <- merge(test,Avg_Sales,by=c("Store","DayOfWeek"))
test <- test[order(Store,Date)]
test[,FPPredictions:=Open * predictions$Point.Forecast]
test[,FPredictions:=ifelse(is.na(predictions$Point.Forecast),AS,predictions$Point.Forecast)]

results <- data.frame(Id=test$Id, Sales=test$FPredictions)
results <- results[order(results$Id),]

write.csv(results, "./Rossmann_TSLM.csv",row.names=F)