时间序列预测、平稳性分析

Question

I have a yearly data of carbon emission from 1990 to 2017(28 datapoints) and I want to perform Time series forecasting on it. The rolling mean of the data with a window of 2 keeps increasing with time that indicates my data is not stationary but dicky-fuller test gives a p-value of 0.04 which is less then 0.05, hence this indicates that the data is stationary. What can be done in this situation?

def rolling_stats(time_data):
     rolling_mean=time_data.rolling(2).mean()
     rolling_std=time_data.rolling(2).std()
     raw_data=plt.plot(time_data,color="blue",label="Original Data")
     mean=plt.plot(rolling_mean,color="green",label="mean")
     std=plt.plot(rolling_std,color="red",label="std")
     plt.legend(loc="best")
     plt.title("Mean and standard deviation")
     plt.xlabel("Years")
     plt.ylabel("Carbon emitted in megatonnes")
rolling_stats(CarbonEmitted)

Answer 1

Without seeing the code/data it's hard to see what you did wrong, for example, you could have performed adfuller on the standard deviation or something. Your data should be of the format

Date	Emission
1992-01-01	21
1993-01-01	20.5

Date is usually set as the index and your dates should be of the type datetime

Answer 2

The ADF test is used to determine if differencing the data will create staionairty, thus inferring that the current data is not. I don't have your data so I cannot tell if it does, but what you want to test is whether differencing the data results in random noise centered at 0. If you do not get stationary data when performing ADF then that may be why ADF is failing on your original data.

There are other tests you can look into, sometimes differencing isn't required to make data stationary and a log transform is more appropriate, experiment with your data and see.