README.md

In alexhallam/cfsales: Data from the open Kaggle dataset Corporacion Favorita Grocery Sales

cfsales

The goal of cfsales is to make the Corporación Favorita Grocery Sales Forecasting data easy to access. The intention is to have a realistic dataset for proving the value of various forecast methods, pipelines, data-cleaning, etc. in retail-like settings.

Due to the size of the data exceding the size limits set by CRAN, a data package was made and stored as a drat repository. It is neccesary that cfsalesdata1 and cfsalesdata2 are installed.

Installation

Due to the data size this repository could not be stores as a single github repository. Files had to be broken in to two data packages as drat repositories. The data package cfsalesdata1 and cfsalesdata2 that is available and easy to install through drat repositories on GitHub as shown below. The cfsalesdata1 package contains the majority of the data tables. If it is required that the full training set be loaded then the cfsalesdata2 packages will also be required.

You can install the required packages with the following command.

install.packages("cfsalesdata1",repos = "https://alexhallam.github.io/drat/", type = "source")
install.packages("cfsalesdata2",repos = "https://alexhallam.github.io/drat/", type = "source")

The cfsalesdata* packages provides the following data tables.

• cfsalesdata1::holidays_events: holidays and events, with metadata.
• cfsalesdata1::items: item metadata
• cfsalesdata1::transactions: count of sales transactions for each store_nbr/date combination. Only included for the training data time frame.
• cfsalesdata1::oil: daily oil price
• cfsalesdata1::stores: store metadata
• train*: training set for the Kaggle competition. The target variable is unit_sales. The granularity is store-unit-day, where each store day has a count of many various units. Note: The training set had to be split into 4 sets to aid in the file compression process. These sets are cfsalesdata1::train1 (45.7 MB),cfsalesdata1::train2(36.4 MB), cfsalesdata2::train3(42.7 MB),cfsalesdata2::train4(28.9 MB). They may be bound together with rbind().
• cfsalesdata1::test: test set for the Kaggle competition. It should be noted that some items ids are present in the test set which are not present in the training set. It is expected that the forecaster should generate predictions of unit sales for new items.
• cfsalesdata1::store_day_train: one possible aggregation of raw data to simulate a data set much like one would see in the wild. The target variable in this data table is transactions. This is different from the target variable in the train* data table. The granularity is store-day.
• cfsalesdata1::store_day_test: the test set for final validation of forecast methods.

Once the cfsalesdata1 and cfsalesdata2 package has been installed and loaded data may be called with the data() function.

library(cfsalesdata1)
data("store_day_train") 
str(store_day_train)

Since these data sources are timeseries one may also be interested in converting a data source into a tsibble for functions that return information on time gaps, easy calculation of sliding window functions, and other usefull time series calculations.

> library(tsibble)
> tsibble::as_tsibble(x = store_day_test, index = date, key = store_nbr)

# A tsibble: 864 x 47 [1D]
# Key:       store_nbr [54]
# Groups:    @ date [16]
   date       store_nbr dcoilwtico is_navidad is_carnaval is_new_year is_black_firday is_cyber_monday is_futbol
   <date>         <int>      <dbl>      <dbl>       <dbl>       <dbl>           <dbl>           <dbl>     <dbl>
 1 2017-08-16         1       46.8          0           0           0               0               0         0
 2 2017-08-17         1       47.1          0           0           0               0               0         0
 3 2017-08-18         1       48.6          0           0           0               0               0         0
 4 2017-08-19         1       NA            0           0           0               0               0         0
 5 2017-08-20         1       NA            0           0           0               0               0         0
 6 2017-08-21         1       47.4          0           0           0               0               0         0
 7 2017-08-22         1       47.6          0           0           0               0               0         0
 8 2017-08-23         1       48.4          0           0           0               0               0         0
 9 2017-08-24         1       47.2          0           0           0               0               0         0
10 2017-08-25         1       47.6          0           0           0               0               0         0
# ........

To use the full training data use rbind() to join all of the training sets

library(cfsalesdata1)
library(cfsalesdata2)
library(lobstr)
train <- rbind(train1, train2, train3, train4)

# be aware of the size of this object
lobstr::obj_size(train)
# 4,015,907,776 B (aka 4 gigabytes)

alexhallam/cfsales documentation built on Oct. 1, 2019, 2:50 a.m.