# This is a variation of DSEverything's ETS Starter Script located at: https://www.kaggle.com/dongxu027/time-series-ets-starter-lb-0-590
# Most of the credit goes to him. I added Box-Cox pre-processing of the data (the lambda parameter of the ets() function) and used data
# back to 2016 rather than just 2017. I also made some of the processing steps more data.table idiomatic (and hopefully faster as a result). 
# Version 1: LB 0.564


library(dplyr)
library(forecast)
library(data.table)
library(foreach)
library(lubridate)
library(doMC)

train <- fread('../input/train.csv')
test <-fread('../input/test.csv')

train[, date := as.Date(fast_strptime(date, format = "%Y-%m-%d"))]
test[, date := as.Date(fast_strptime(date, format = "%Y-%m-%d"))]

train[, store_item_nbr := paste(store_nbr, item_nbr, sep = "_")]
test[, store_item_nbr := paste(store_nbr, item_nbr, sep = "_")]

# train_sub -  use 2016 data to start with the forecast
train_sub <- train[date >= as.Date("2016-01-01"), c('date','store_item_nbr', 'unit_sales')]

# dcast the data from long to wide format for time series forecasting
train_sub_wide <- dcast(train_sub, store_item_nbr ~ date, value.var = "unit_sales", fill = 0)
train_ts <- ts(train_sub_wide[, 2:ncol(train_sub_wide)], frequency = 7) # considering one week as a short shopping cycle


fcst_intv = 16  # 16 days of forecast interval (Aug 16 ~ 31) per the submission requirement
fcst_matrix <- matrix(NA,nrow=nrow(train_ts),ncol=fcst_intv)

# register 15 cores for parallel processing in ETS forecasting
registerDoMC(detectCores())
fcst_matrix <- foreach(i=1:nrow(train_ts),.combine=rbind, .packages=c("forecast")) %dopar% { 
    fcst_matrix <- forecast(ets(train_ts[i,], lambda = 0.5),h=fcst_intv)$mean
}

# post-processing the forecast table
fcst_matrix[fcst_matrix < 0] <- 0

colnames(fcst_matrix) <- as.character(seq(from = as.Date("2017-08-16"), 
                                          to = as.Date("2017-08-31"), 
                                          by = 'day'))
fcst_df <- as.data.frame(cbind(train_sub_wide[, 1], fcst_matrix)) 
colnames(fcst_df)[1] <- "store_item_nbr"

# melt the forecast data frame from wide to long format for final submission
fcst_df_long <- melt(fcst_df, id = 'store_item_nbr', 
                     variable.name = "fcst_date", 
                     value.name = 'unit_sales')
fcst_df_long$store_item_nbr <- as.character(fcst_df_long$store_item_nbr)
fcst_df_long$fcst_date <- as.Date(parse_date_time(fcst_df_long$fcst_date,'%y-%m-%d'))
fcst_df_long$unit_sales <- as.numeric(fcst_df_long$unit_sales)

# generate the final submission file
submission <- left_join(test, fcst_df_long, 
                        c("store_item_nbr" = "store_item_nbr", 'date' = 'fcst_date'))
submission$unit_sales[is.na(submission$unit_sales)] <- 0
submission <- select(submission, c('id', 'unit_sales'))
write.csv(submission, "submission_v0.csv", row.names = FALSE)