最近也正在搞這個問題，

模型選擇準則的話，

1.先選擇限制少的模型，也就是趨勢+截距模型，看是否unit root，有unit root，則再檢定趨勢項是否不為0，若不為0, 以一般t值來檢定adf值。

2.若前一步驟的趨勢為0，則選擇截距模型做adf檢定，當然要先看是否有unit root。檢定完unit root後，再檢定截距是否為0。步驟同上。

3.若截距為0，則不顯著，所以這次選擇none做adf檢定。

不過有疑問的是，在檢定這三個模型時，lag數需不需要設成一樣來做檢定？

ADF test is used to test whether a time series is stationary (vs. Alternative Hypothesis: Nonstationary). It is based on an artifical difference equation which can consists of several lags.

Step 1: Model selection (include intercept, deterministic trend, etc). It is an art, you can decide your model by observing the graph first.

Step 2: use Akaki or SBC to decide the number of lags.

Step 3: If there is a unit root (nonstationary), reserve your judgment because the test has low power.

Step 4: use KPSS test to examine an opposite hypothesis.

Step 5: If KPSS result also indicates nonstationarity, good!. Otherwise, you have to rely on your sujective decision.