2

我将这个脚本作为学校预测项目的一部分运行，但我得到了一些奇怪的结果，尤其是 MAPE 值。它应该做的是预测未来 12 个月的国际恐怖主义事件。谁能告诉我这份报告是否准确，或者我是否遗漏了什么？我试图将这些图表包括在内，但我认为它们不能在这里发布。

谢谢

library(ggplot2)
library(forecast)
library(tseries)
library(reprex)

terror <- tibble::tribble(
  ~imonth, ~iyear, ~monthly,
  1,   2015,     1534,
  2,   2015,     1295,
  3,   2015,     1183,
  4,   2015,     1277,
  5,   2015,     1316,
  6,   2015,     1168,
  7,   2015,     1263,
  8,   2015,     1290,
  9,   2015,     1107,
  10,   2015,     1269,
  11,   2015,     1172,
  12,   2015,     1091,
  1,   2016,     1162,
  2,   2016,     1153,
  3,   2016,     1145,
  4,   2016,     1120,
  5,   2016,     1353,
  6,   2016,     1156,
  7,   2016,     1114,
  8,   2016,     1162,
  9,   2016,     1045,
  10,   2016,     1140,
  11,   2016,     1114,
  12,   2016,      923,
  1,   2017,      879,
  2,   2017,      879,
  3,   2017,      961,
  4,   2017,      856,
  5,   2017,     1081,
  6,   2017,     1077,
  7,   2017,      994,
  8,   2017,      968,
  9,   2017,      838,
  10,   2017,      805,
  11,   2017,      804,
  12,   2017,      749
)

# aggregated data
terror_byMonth_Train = ts(data = terror$monthly,
                          start = c(2015,1), 
                          end = c(2016,12),
                          frequency=12) 

terror_byMonth_Test = ts(data = terror$monthly,
                         start = c(2017,1), 
                         end = c(2017,12),
                         frequency=12)

# arima instead of exp smooth
m_arima <- auto.arima(terror_byMonth_Train)
#> Warning in value[[3L]](cond): The chosen test encountered an error, so no
#> seasonal differencing is selected. Check the time series data.

# fit exp smooth model
m_ets = ets(terror_byMonth_Train) 

# Get length of terror_byMonth_Test set
size <- length(terror_byMonth_Test)

# forecast for 2017 using multiple forecast (Davis Style)
f_arima_multi <- m_arima %>%
  forecast(h = size)

f_arima_multi %>%
  autoplot()


# forecast ARIMA 2017 (Orininal Style)
f_arima<-forecast(m_arima,h=12)
f_arima %>%
  autoplot()


# forecast ETS 2017
f_ets = forecast(m_ets, h=12) 
f_ets %>%
  autoplot()


# check accuracy ETS
acc_ets <- accuracy(m_ets) 

#check accuracy ARIMA, between train and test sets
acc_arima_TrainVSTest <- accuracy(f_arima_multi, x = terror_byMonth_Test)

# check accuarcy ARIMA
acc_arima <- accuracy(f_arima)

# MAPE(ETS)= 20.03 < MAPE(ARIMA) = 22.05
# ETS model chosen 

# Compare to 2017 data
accuracy(f_ets, terror_byMonth_Test)
#>                     ME      RMSE       MAE       MPE      MAPE      MASE
#> Training set -14.30982  90.08823  70.06438 -1.606862  5.900178 0.5790445
#> Test set     303.53575 316.03133 303.53575 23.986363 23.986363 2.5085599
#>                       ACF1 Theil's U
#> Training set  0.0008690031        NA
#> Test set     -0.2148651254  2.356116

由reprex 包（v0.2.1）于 2019 年 2 月 13 日创建

4

1 回答 1

1

问题在于您如何定义terror_byMonth_Test. 它应该是，例如，

terror_byMonth_Test <- ts(data = tail(terror$monthly, 12),
                          start = c(2017, 1), 
                          end = c(2017, 12),
                          frequency = 12)

也就是说，仅仅提供开始和结束日期并不足以ts知道 24 个观测值中的哪 12 个观测值terror$monthly。这将 MAPE 降低到 10.4%。

于 2019-02-17T02:42:13.907 回答