initial_timeseries_models.Rmd

---
title: "Time Series Analysis of the Online Retail Dataset"
author: "Mick Cooney <mickcooney@gmail.com>"
date: "Last updated: `r format(Sys.time(), '%B %d, %Y')`"
output:
  rmdformats::readthedown:
    toc_depth: 3
    use_bookdown: TRUE
    code_folding: hide
    fig_caption: TRUE

  html_document:
    fig_caption: yes
    theme: spacelab #sandstone #spacelab #flatly
    highlight: pygments
    number_sections: TRUE
    toc: TRUE
    toc_depth: 2
    toc_float:
      smooth_scroll: FALSE

  pdf_document: default
---


```{r import_libraries, echo=FALSE, message=FALSE}
library(conflicted)
library(tidyverse)
library(scales)
library(cowplot)
library(magrittr)
library(glue)
library(stringr)
library(rlang)
library(purrr)
library(furrr)
library(DT)
library(tidygraph)
library(tidyquant)


source("lib_utils.R")

conflict_lst <- resolve_conflicts(
  c("magrittr", "rlang", "dplyr", "readr", "purrr", "ggplot2", "DT",
    "tidyquant")
  )


knitr::opts_chunk$set(
  tidy       = FALSE,
  cache      = FALSE,
  warning    = FALSE,
  message    = FALSE,
  fig.height =     8,
  fig.width  =    11
  )

options(
  width = 80L,
  warn  = 1,
  mc.cores = parallel::detectCores()
  )

theme_set(theme_cowplot())

set.seed(42)

plan(multisession)
```

# Load Data

We first want to load our datasets and prepare them for some simple association
rules mining.

```{r load_transaction_data, echo=TRUE}
tnx_data_tbl <- read_rds("data/retail_data_cleaned_tbl.rds")

tnx_data_tbl %>% glimpse()
```

To use our rules mining we just need the invoice data and the stock code, so
we can ignore the rest. Also, we ignore the issue of returns and just look at
purchases.

```{r prepare_data_arules, echo=TRUE}
tnx_purchase_tbl <- tnx_data_tbl %>%
  filter(
    quantity > 0,
    exclude == FALSE
    ) %>%
  select(
    invoice_date, invoice_id, stock_code, description, customer_id,
    quantity, price, stock_value
    )

tnx_purchase_tbl %>% glimpse()
```


# Construct Basic Time-Series

There are a number of different perspective on this data - we can look at
customer-based time-series or individual items or just total revenue.

In all these cases we aggregate daily, weekly and monthly, so rather than
repeating our code over and over, we construct a function to do this.

```{r construct_aggregation_time_series, echo=TRUE}
construct_time_series_data <- function(data_tbl) {
  daily_tbl <- data_tbl %>%
    tq_transmute(
      mutate_fun = apply.daily,
      FUN        = sum,
      na.rm      = TRUE,
      col_rename = "amount"
      )
  
  weekly_tbl <- data_tbl %>%
    tq_transmute(
      mutate_fun = apply.weekly,
      FUN        = sum,
      na.rm      = TRUE,
      col_rename = "amount"
      )
    
  monthly_tbl <- data_tbl %>%
    tq_transmute(
      mutate_fun = apply.monthly,
      FUN        = sum,
      na.rm      = TRUE,
      col_rename = "amount"
      )

  agg_tbl <- list(
      daily   = daily_tbl,
      weekly  = weekly_tbl,
      monthly = monthly_tbl
      ) %>%
    bind_rows(.id = "period")
  
  return(agg_tbl)  
}
```




## Total Revenue

A revenue-based approach is the simplest, so we start there.

```{r construct_revenue_time_series, echo=TRUE}
use_data_tbl <- tnx_purchase_tbl %>%
  select(date = invoice_date, stock_value)

ts_revenue_data_tbl <- use_data_tbl %>%
  construct_time_series_data() %>%
  mutate(
    label = "total_revenue",
    
    .after = "period"
  )

ts_revenue_data_tbl %>% glimpse()
```

Now that we have constructed this simple time-series, we now want to plot them.

```{r plot_totalrevenue_time_series, echo=TRUE}
ggplot(ts_revenue_data_tbl) +
  geom_line(aes(x = date, y = amount)) +
  scale_y_continuous(labels = label_comma()) +
  facet_wrap(vars(period), scales = "free_y", ncol = 2) +
  xlab("Date") +
  ylab("Amount") +
  ggtitle("Time Series Plots of Total Revenue")

```


## Frequent Purchases

Time series plots of different item types may be useful, but only for items
that have regular transactions, as otherwise the data will be mostly zero
entries. Instead, we want to find items that are purchased on a regular basis
and inspect those.

```{r plot_histogram_day_count, echo=TRUE}
stock_daycount_tbl <- tnx_purchase_tbl %>%
  select(invoice_date, stock_code) %>%
  distinct() %>%
  count(stock_code, name = "day_count")

ggplot(stock_daycount_tbl) +
  geom_histogram(aes(x = day_count), bins = 50) +
  xlab("Days in Transaction Log") +
  ylab("Count of Items") +
  ggtitle("Histogram of Date Appearances for Stock Items")
```

Rather than focus on all these items, we choose the arbitrary cut-off of the
top 12 items, and construct time series of the daily sales of each of these
items.


```{r construct_item_sales_time_series, echo=TRUE}
use_data_tbl <- tnx_purchase_tbl %>%
  select(date = invoice_date, stock_code, stock_value)

use_stock_tbl <- stock_daycount_tbl %>%
  arrange(desc(day_count)) %>%
  slice_head(n = 12)

ts_itemsales_data_tbl <- use_data_tbl %>%
  semi_join(use_stock_tbl, by = "stock_code") %>%
  group_by(stock_code) %>%
  construct_time_series_data() %>%
  mutate(
    label = "item_sales",
    
    .after = "period"
  ) %>%
  ungroup()

ts_itemsales_data_tbl %>% glimpse()
```


```{r plot_itemsales_time_series, echo=TRUE}
ggplot(ts_itemsales_data_tbl %>% filter(period == "daily")) +
  geom_line(aes(x = date, y = amount)) +
  scale_y_continuous(labels = label_comma()) +
  facet_wrap(vars(stock_code), scales = "free_y") +
  xlab("Date") +
  ylab("Sales") +
  ggtitle("Time Series Plots of Item Sales (daily)") +
  theme(axis.text.x = element_text(angle = 30, vjust = 0.5))


ggplot(ts_itemsales_data_tbl %>% filter(period == "weekly")) +
  geom_line(aes(x = date, y = amount)) +
  scale_y_continuous(labels = label_comma()) +
  facet_wrap(vars(stock_code), scales = "free_y") +
  xlab("Date") +
  ylab("Sales") +
  ggtitle("Time Series Plots of Item Sales (weekly)") +
  theme(axis.text.x = element_text(angle = 30, vjust = 0.5))


ggplot(ts_itemsales_data_tbl %>% filter(period == "monthly")) +
  geom_line(aes(x = date, y = amount)) +
  scale_y_continuous(labels = label_comma()) +
  facet_wrap(vars(stock_code), scales = "free_y") +
  xlab("Date") +
  ylab("Sales") +
  ggtitle("Time Series Plots of Item Sales (monthly)") +
  theme(axis.text.x = element_text(angle = 30, vjust = 0.5))

```








# R Environment

```{r show_session_info, echo=TRUE, message=TRUE}
sessioninfo::session_info()
```