在使用R语言的 Arima 进行预测时，其时间序列很平稳，但是预测结果与理想结果相差很远，请教高人帮忙看看

  1. 使用的时间序列参数如下 data<-c(20.94,-22.76,4.33,-4.6,21.75,-24.86,4.87,-5.4,21.48,-23.46,4.22,-4.57,21.92,-22.97,4.26,-4.53,21.79,-26.0,4.54,-5.02) plot(data,type="l")2. acf图 与pacf 图
  2. 开始预测 data.arima<-auto.arima(data) 输出:
Series: data 
ARIMA(1,0,1) with zero mean 

Coefficients:
          ar1      ma1
      -0.4605  -0.8257
s.e.   0.2021   0.1301

sigma^2 estimated as 96.22:  log likelihood=-74.01
AIC=154.01   AICc=155.51   BIC=157

预测命令

data.forecast<-forecast(data.arima,5)

输出:

   Point Forecast     Lo 80     Hi 80     Lo 95    Hi 95
21     -2.9580835 -15.53001  9.613843 -22.18519 16.26902
22      1.3622049 -19.11904 21.843454 -29.96116 32.68557
23     -0.6272988 -22.41997 21.165376 -33.95632 32.70172
24      0.2888727 -21.77189 22.349635 -33.45015 34.02790
25     -0.1330266 -22.25022 21.984170 -33.95836 33.69231

预测结果图：

  1. 残差正态性

    qqnorm(data.arima$residuals)
qqline(data.arima$residuals)

    图片说明

  2. 无关性检验

Box.test(data.arima$residuals, type = "Ljung-Box")

        Box-Ljung test

data:  data.arima$residuals
X-squared = 0.28894, df = 1, p-value = 0.5909
  1. 是否平稳 adf.test ,平稳性检验0.01，为平稳性数据
library(tseries)
adf.test(data)

        Augmented Dickey-Fuller Test

data:  data
Dickey-Fuller = -59.373, Lag order = 2, p-value = 0.01
alternative hypothesis: stationary

Warning message:
In adf.test(data) : p-value smaller than printed p-value

用的模型不好，Arima不能有效学习出周期规律。试试看用rnn/lstm网络去预测。

模型参数没有通过显著性检验，不可以用模型预测