• Time series analyses apply to endless job types and sectors

  • Economic forecasting

  • Climate change planning

  • Political polling and predictions

• Why learn it in R?

  • R is largely considered a gold standard for open-source statistical analysis

  • One of the ten most popular programming languages world wide

• We will look at Walmart stock price data
  • Sourced via Kaggle

• Scope: Jan 2014 to Jan 2024

• Focus: will look at average opening price on a monthly level

• Goal: Determine which time series model is most effective / applicable for this situation

# read in data, scan check data types and affirm no missing values
full_data = read_csv("wmt_data.csv")
glimpse(full_data)
sum(is.na(full_data)) # no missing values

# select columns we need, renaming it to lowercase for easier coding
data = select(full_data, c(Date, Open))
colnames(data) = c("date", "open")

# look at 10 year span, January of 2014 to January 2024
data = filter(data, date >= "2014-01-01" & date <= "2024-01-01") 

## create single column for month/year, drop original full date column
data = mutate(data, month_year = floor_date(date, unit = "month"))
data = select(data, c(open, month_year))

# getting average open by month
data = group_by(data, month_year) %>% 
  summarize(
    avg_open = mean(open),
    #min_open = min(open),
    #max_open = max(open)
  )

# round to nearest cent
data$avg_open = round(data$avg_open,2)

• Overall trend is positive and linear (additive)

• Throughout whole time duration, no obvious pattern of consistent and noticeable seasonal oscillation
  • Aug 2017 to Feb 2019 large difference between Q2 and Q4
  • Feb 2022 to Feb 2023; large difference between Q1 and Q2
• Will perform formal decomposition before making final model choice.

# convert data frame to time series object
data_ts = ts(data$avg_open,
             start = c(2014,1),
             end = c(2023, 12),
             frequency = 12)

# create with decompose() function
help("decompose")
decomp_add = decompose(x = data_ts,
                       type = "additive")

# create with stl() function
help(stl)
decomp_stl = stl(x = data_ts,
    s.window = "periodic") 
    # s.window function signals that sesonality does, at least in theory, exist

After looking at initial visualizations and decompisitions:

• Positive trend that is fairly linear

• Constant seasonal oscillation

• Given this, will test the HW additive and HW damped additive models

• Will use the last 12 months (10% of the data) as testing split

hw_add = hw(training_ts,
            h = 12,
            damped = FALSE,
            seasonal = "additive")

hw_add_damped = hw(training_ts,
            h = 12,
            damped = TRUE,
            seasonal = "additive")

hw_add_predictions = data.frame(round(hw_add$mean,2))
hw_add_damped_predictions = data.frame(round(hw_add_damped$mean,2))

predictions = cbind(data_testing$month_year,
                    hw_add_predictions,
                    hw_add_damped_predictions)

colnames(predictions) = c("month_year", "hw_add", "hw_add_damped")

• Ultimately, HW Add was the ideal choice for this scenario.

• Original dataset (Kaggle)

• “Statistical language R is making a comeback against Python”

• “Why R is the Best Language for Statistical Analysis”

• “20 Exciting Use Cases for Time Series”