auto.arima 在 R 版本 3.2.0 和 R 版本 3.2.2 中 d=1 的不同结果

Different results for auto.arima with d=1 in R version 3.2.0 and R version 3.2.2

我在一台计算机上安装了 R 3.2.0 版,在另一台计算机上安装了 3.2.2 版,两台计算机上安装了相同版本的软件包 "forecast" (6.1)。当我将带有 d=1auto.arima 函数应用于 3.2.0 版计算机中的向量 fec1(下面的数据)时,它会给我一个 ARIMA(1,1,1)。在 3.2.2 版的计算机中,它给了我一个带漂移的 ARIMA(1,1,0)。为什么会出现这样的差异?当我省略选项 d=1 时,我在两台计算机上得到相同的结果。

fec1 <- c(9.53408021, 9.48993631, 9.44261613, 9.38585482, 9.31611858, 
          9.24743801, 9.19483663, 9.11690414, 9.05586032, 8.96781034, 
          8.85437896, 8.73956021, 8.62587246, 8.49758284, 8.41877774, 
          8.29737132, 8.02294327, 7.62604834, 7.32936532, 7.10006394, 
          6.87039464, 6.56332757, 6.27195705, 5.97089310, 5.69537808, 
          5.46074967, 5.19061116, 4.90749261, 4.61822473, 4.39635507, 
          3.96883450, 3.71230296, 3.43604354, 3.15150833, 2.82864081, 
          2.51857560, 2.27264470, 2.06785600, 1.86116599, 1.61452534, 
          1.41214847, 1.26673153, 1.15033592, 1.00646045, 0.83493055, 
          0.65399085, 0.47450499, 0.30408840, 0.18216540, 0.09006937, 
          0.00000000) 

参见 PR#16278, "nested arima model has higher log-likelihood", which as noted in the change log for 3.2.1,影响 d >= 1 的 arima 计算:

arima(*, xreg = .) (for d >= 1) computes estimated variances based on a the number of effective observations as in R version 3.0.1 and earlier. (PR#16278)