2

我想用时间序列 T 做这两件事(结合):

  1. 预测 T 的季节性调整分量(用于分解的 STL)并“加回”季节性(我假设季节性分量不变,所以我对季节性分量使用天真的方法)
  2. 拟合带有 ARIMA 误差的回归模型(公式中包含外生回归量)

换句话说,我想使用 T 的季节性调整组件集成外部预测器并“加回”季节性来获得预测。

我可以分别做这两个操作,但我不能让它们结合起来工作

以下是一些玩具示例:

首先,加载库和数据:

library(forecast)
library(tsibble)
library(tibble)
library(tidyverse)
library(fable)
library(feasts)
library(fabletools)


us_change <- readr::read_csv("https://otexts.com/fpp3/extrafiles/us_change.csv") %>%
  mutate(Time = yearquarter(Time)) %>%
  as_tsibble(index = Time)

带有 T 的季节性调整分量的拟合和预测示例:

model_def = decomposition_model(STL,
                                Consumption  ~ season(window = 'periodic') + trend(window = 13),
                                ARIMA(season_adjust ~ PDQ(0,0,0)),
                                SNAIVE(season_year),
                                dcmp_args = list(robust=TRUE)) 

fit <- us_change %>% model(model_def)

report(fit)

forecast(fit, h=8) %>% autoplot(us_change)

具有 ARIMA 误差的回归模型示例(收入作为预测变量):

model_def = ARIMA(Consumption ~ Income + PDQ(0,0,0))

fit <- us_change %>% model(model_def)

report(fit)

us_change_future <- new_data(us_change, 8) %>% mutate(Income = mean(us_change$Income))

forecast(fit, new_data = us_change_future) %>% autoplot(us_change)

这些示例有效,但我想做这样的事情:

model_def = decomposition_model(STL,
                                Consumption  ~ season(window = 'periodic') + trend(window = 13),
                                ARIMA(season_adjust ~ Income + PDQ(0,0,0)),
                                SNAIVE(season_year),
                                dcmp_args = list(robust=TRUE))


fit <- us_change %>% model(model_def)

report(fit)

us_change_future <- new_data(us_change, 8) %>% mutate(Income = mean(us_change$Income))

forecast(fit, new_data = us_change_future) %>% autoplot(us_change)

我在控制台中得到这个输出:

> fit <- us_change %>% model(model_def)
Warning message:
1 error encountered for model_def
[1] object 'Income' not found

> 
> report(fit)
Series: Consumption 
Model: NULL model 
NULL model> 

所以我尝试在分解模型中这样做:

model_def = decomposition_model(STL,
                                Consumption  ~ season(window = 'periodic') + trend(window = 13),
                                ARIMA(season_adjust ~ us_change$Income + PDQ(0,0,0)),
                                SNAIVE(season_year),
                                dcmp_args = list(robust=TRUE))

合身没问题,但现在我在预测中遇到错误:

> forecast(fit, new_data = us_change_future) %>% autoplot(us_change)
Error in args_recycle(.l) : all(lengths == 1L | lengths == n) is not TRUE
In addition: Warning messages:
1: In cbind(xreg, intercept = intercept) :
  number of rows of result is not a multiple of vector length (arg 2)
2: In z[[1L]] + xm :
  longer object length is not a multiple of shorter object length

我究竟做错了什么?

4

1 回答 1

3

您的代码在这里没有任何问题,只是我没有考虑过人们在制作decomposition_model(). 我更新了分解建模方法以包含外生回归量,以便它们可以在组件模型中使用(https://github.com/tidyverts/fabletools/commit/8dd505f6378327b8e93b8440ec17ecf9badf2561)。如果您更新软件包,您的第一次建模尝试应该可以正常工作。

至于为什么第二次尝试没有成功,预测方法是找到 us_change$Income 并将其用作未来预测的外生回归量。该值的长度为us_change,与 的长度不匹配us_change_future,从而导致(混淆)错误。


代表:

library(tidyverse)
library(tsibble)
library(fable)
library(feasts)

us_change <- readr::read_csv("https://otexts.com/fpp3/extrafiles/us_change.csv") %>%
  mutate(Time = yearquarter(Time)) %>%
  as_tsibble(index = Time)

model_def = decomposition_model(STL,
                                Consumption  ~ season(window = 'periodic') + trend(window = 13),
                                ARIMA(season_adjust ~ Income + PDQ(0,0,0)),
                                SNAIVE(season_year),
                                dcmp_args = list(robust=TRUE))


fit <- us_change %>% model(model_def)

report(fit)
#> Series: Consumption 
#> Model: STL decomposition model 
#> Combination: season_adjust + season_year
#> 
#> ========================================
#> 
#> Series: season_adjust 
#> Model: LM w/ ARIMA(1,0,2) errors 
#> 
#> Coefficients:
#>          ar1      ma1     ma2  Income  intercept
#>       0.6922  -0.5777  0.1975  0.2035     0.5993
#> s.e.  0.1163   0.1305  0.0755  0.0462     0.0883
#> 
#> sigma^2 estimated as 0.3234:  log likelihood=-157.39
#> AIC=326.77   AICc=327.24   BIC=346.16
#> 
#> Series: season_year 
#> Model: SNAIVE 
#> 
#> sigma^2: 0

us_change_future <- new_data(us_change, 8) %>% mutate(Income = mean(us_change$Income))

forecast(fit, new_data = us_change_future) %>% autoplot(us_change)

reprex 包(v0.2.1)于 2019 年 10 月 9 日创建

于 2019-10-08T22:51:22.787 回答