将外部回归量纳入分层/分组时间序列

Incorporating external regressor in a hierarchical/ grouped time series

各位贡献者,

我一直在处理分层时间序列,它与许多商店中的一组相同产品有关。为此,在我的例子中,当我们基于“商店”和“product_type”等 2 个属性聚合数据集时,我们应该聚合目标变量,即每个组或层次结构的每个产品的“需求” . 我想做的是在我的模型中添加另一个分类变量,让我们说“动态谐波回归”,因为我正在使用每周时间序列。但是,当我的外部变量是具有 4 个级别的分类变量时,我不知道应该如何包含它。我想知道我如何汇总这个或者我是否可以做些什么。 在这里您可以找到一个可重现的小例子:

library(tidyverse)
library(tsibble)
library(tsibbledata)
library(fable)
library(fabletools)
library(fpp3)
library(readxl)
library(fable.prophet)
library(feasts)

store <- c(rep('st1', 8), rep('st2', 8))
product_type <- c(rep('type1', 4), rep('type2', 4), rep('type1', 4), rep('type2', 4))
products <- c(rep('A', 2), rep('B', 2), rep('C', 2), rep('D', 2), 
              rep('A', 2), rep('B', 2), rep('C', 2), rep('D', 2))

demands <- c(round(sample(c(1:100), 16, replace = TRUE)))
external_reg <- c(sample(c('red', 'green', 'blue'), 16, replace = TRUE))
date_week <- rep(1:4, 4)
date_year <- rep(2019:2022, 4)

my_data <- tibble(date_year, date_week, store, product_type, products, demands, external_reg)

my_data %>%
  mutate(Date = ymd(paste0(date_year, "-01-01")) + weeks(date_week - 1)) %>%
  mutate(Week = yearweek(Date)) %>%
  as_tsibble(key = c(store, product_type), index = Week) %>%
  aggregate_key(store * product_type, Demand_Agg = sum(demands))

很明显,外部回归器应该是我的 tsibble:

中的一列
# A tsibble: 36 x 4 [53W]
# Key:       store, product_type [9]
       Week store        product_type Demand_Agg
     <week> <chr*>       <chr*>            <dbl>
 1 2019 W01 <aggregated> <aggregated>        188
 2 2020 W02 <aggregated> <aggregated>        142
 3 2021 W02 <aggregated> <aggregated>        259
 4 2022 W03 <aggregated> <aggregated>        186
 5 2019 W01 st1          <aggregated>         89
 6 2019 W01 st2          <aggregated>         99
 7 2020 W02 st1          <aggregated>         52
 8 2020 W02 st2          <aggregated>         90
 9 2021 W02 st1          <aggregated>         95
10 2021 W02 st2          <aggregated>        164
# … with 26 more rows

非常感谢您。

外部回归量列 (external_reg) 已从您的输出中删除,因为您尚未指定应如何聚合它。鉴于它是一个离散变量,以保留此信息的方式聚合数据可能很棘手。您选择如何聚合这取决于您,并且可能取决于您要使用的模型。如果您有像温度这样的连续变量,您可能需要计算平均温度。

例如,如果您想保留 external_reg 的第一个值,您可以将它与 aggregate_key(<tsibble>, store * product_type, Demand_Agg = sum(demands), external_reg = first(external_reg))

聚合
library(dplyr)
library(fable)
library(tsibble)
library(lubridate)

my_data <- structure(list(date_year = c(2019L, 2020L, 2021L, 2022L, 2019L, 
2020L, 2021L, 2022L, 2019L, 2020L, 2021L, 2022L, 2019L, 2020L, 
2021L, 2022L), date_week = c(1L, 2L, 3L, 4L, 1L, 2L, 3L, 4L, 
1L, 2L, 3L, 4L, 1L, 2L, 3L, 4L), store = c("st1", "st1", "st1", 
"st1", "st1", "st1", "st1", "st1", "st2", "st2", "st2", "st2", 
"st2", "st2", "st2", "st2"), product_type = c("type1", "type1", 
"type1", "type1", "type2", "type2", "type2", "type2", "type1", 
"type1", "type1", "type1", "type2", "type2", "type2", "type2"
), products = c("A", "A", "B", "B", "C", "C", "D", "D", "A", 
"A", "B", "B", "C", "C", "D", "D"), demands = c(45, 12, 70, 66, 
77, 6, 27, 52, 8, 73, 70, 27, 84, 100, 79, 51), external_reg = c("blue", 
"green", "red", "blue", "green", "blue", "red", "green", "blue", 
"blue", "green", "green", "red", "green", "blue", "green")), class = c("tbl_df", 
"tbl", "data.frame"), row.names = c(NA, -16L))

my_data %>%
  mutate(Date = ymd(paste0(date_year, "-01-01")) + weeks(date_week - 1)) %>%
  mutate(Week = yearweek(Date)) %>%
  as_tsibble(key = c(store, product_type), index = Week) %>%
  aggregate_key(store * product_type, Demand_Agg = sum(demands), external_reg = first(external_reg))
#> # A tsibble: 36 x 5 [53W]
#> # Key:       store, product_type [9]
#>        Week store        product_type Demand_Agg external_reg
#>      <week> <chr*>       <chr*>            <dbl> <chr>       
#>  1 2019 W01 <aggregated> <aggregated>        214 blue        
#>  2 2020 W02 <aggregated> <aggregated>        191 green       
#>  3 2021 W02 <aggregated> <aggregated>        246 red         
#>  4 2022 W03 <aggregated> <aggregated>        196 blue        
#>  5 2019 W01 st1          <aggregated>        122 blue        
#>  6 2019 W01 st2          <aggregated>         92 blue        
#>  7 2020 W02 st1          <aggregated>         18 green       
#>  8 2020 W02 st2          <aggregated>        173 blue        
#>  9 2021 W02 st1          <aggregated>         97 red         
#> 10 2021 W02 st2          <aggregated>        149 green       
#> # … with 26 more rows

reprex package (v2.0.1)

于 2022-05-08 创建