在R中将周度数据转换为时间序列以进行预测

Got it, let's work through this problem together. Your dataset has quarters, weeks within each quarter, and values for variables A and B, but no actual dates—we'll turn this into a proper 52-week frequency time series that's ready for forecasting.

First, let's start with your original data (formatted as an R code block):

df <- structure(
  list(
    Variable = c("A", "B", "A", "B", "A", "B", "A", "B", "A"),
    Quantity = c("1", "100", "2", "5", "6", "30", "8", "15", "133"),
    YearQuarter = c("2017Q2", "2017Q2", "2017Q3", "2017Q3", "2017Q4", "2017Q4", "2018Q1", "2018Q1", "2018Q2"),
    Week = c("1", "10", "1", "2", "1", "6", "2", "9", "13")
  ),
  class = "data.frame",
  row.names = c(NA, -9L)
)

Step 1: Clean and Transform the Data #

First, we need to fix data types (Quantity is currently character) and calculate the yearly week number (since each quarter maps to 13 weeks: Q1=1-13, Q2=14-26, Q3=27-39, Q4=40-52). We'll use dplyr and lubridate for this:

library(dplyr)
library(lubridate)

# Clean data types and extract year/quarter
df_clean <- df %>%
  mutate(
    Quantity = as.numeric(Quantity),
    Week = as.integer(Week),
    # Extract year and quarter number from YearQuarter string
    Year = as.integer(substr(YearQuarter, 1, 4)),
    Quarter = as.integer(substr(YearQuarter, 5, 5))
  ) %>%
  # Calculate the position of the week within the full year
  mutate(YearlyWeek = Week + (Quarter - 1)*13) %>%
  select(Variable, Year, YearlyWeek, Quantity)

Step 2: Create a Complete 52-Week Time Frame #

Your data has gaps (not every week has observations for A or B), so we'll create a full grid of all years and 52 weeks, then merge your data into it. This ensures our time series is continuous:

# Get all unique years from your dataset
years <- unique(df_clean$Year)

# Create a full grid of variables, years, and all 52 weeks
full_time_grid <- expand.grid(
  Variable = c("A", "B"),
  Year = years,
  YearlyWeek = 1:52
)

# Merge with cleaned data, fill missing values (adjust NA handling based on your use case)
ts_data <- full_time_grid %>%
  left_join(df_clean, by = c("Variable", "Year", "YearlyWeek")) %>%
  # Optional: Replace NAs with 0, or use interpolation later if needed
  mutate(Quantity = tidyr::replace_na(Quantity, 0))

Step 3: Convert to a Time Series Object #

For forecasting, you can use either base R's ts object or a more modern tidy format like tsibble (ideal for the fable forecasting package).

Option 1: Base R ts Object (split by Variable)

Since we have two variables, we'll create separate time series for A and B:

# Create time series for Variable A
ts_A <- ts_data %>%
  filter(Variable == "A") %>%
  pull(Quantity) %>%
  ts(start = c(min(years), 1), frequency = 52)

# Create time series for Variable B
ts_B <- ts_data %>%
  filter(Variable == "B") %>%
  pull(Quantity) %>%
  ts(start = c(min(years), 1), frequency = 52)

# Check the structure of the time series
ts_A

Option 2: tsibble for Tidy Forecasting

If you prefer a tidy approach (recommended for flexible forecasting), use the tsibble package:

library(tsibble)

# Create a tsibble with a year-week index
tsibble_data <- ts_data %>%
  # Combine Year and YearlyWeek into a standard year-week index
  mutate(WeekIndex = yearweek(paste(Year, YearlyWeek, sep = "-W"))) %>%
  select(Variable, WeekIndex, Quantity) %>%
  as_tsibble(key = Variable, index = WeekIndex)

# View the tidy time series
tsibble_data

Step 4: Forecast Future Weeks #

With your time series ready, you can use packages like forecast (for base R ts objects) or fable (for tsibble). Here's an example using fable to forecast the next 52 weeks:

library(fable)
library(feasts)

# Fit an ARIMA model and forecast 52 weeks ahead
forecast_result <- tsibble_data %>%
  model(ARIMA(Quantity)) %>%
  forecast(h = 52)

# Plot the forecast alongside historical data
forecast_result %>%
  autoplot(tsibble_data)

This setup gives you a continuous 52-week frequency time series, perfect for forecasting future weeks. Adjust the missing value handling (e.g., use linear interpolation instead of 0) based on your specific data context!

内容的提问来源于stack exchange，提问作者flightless13wings