I understand that the integration is used to 'convert' a non-stationary process into a stationary process using the method of differencing. I'm just unsure exactly what differencing by $d > 1$ means.

Say if we have a series $X_{t}$ and forecast with $d>1$ in the $\text{ARIMA}(p, d, q)$ model is this equivalent to forecasting $Y_{t} = X_{t} - X_{t-1} - \dots - X_{t-d}$ with $\text{ARMA}(p, q)$ or have I misunderstood?


For a fixed $d$, you just repeat the process of differentiation $d$ times, recursively. For example, if you want to get to $d = 3$, start with the first difference:

$$\Delta X_t = X_t - X_{t-1} =: Y_t$$

Then, $Y_t$ is just some series, which you can difference just like you did $X_t$. That would be the 2nd difference:

$$\Delta^2 X_t = \Delta (\Delta X_t) = \Delta Y_t = Y_t - Y_{t-1} = (X_t - X_{t-1}) - (X_{t-1} - X_{t-2}) = X_t - 2X_{t-1} + X_{t-2}$$

And then once more to get to the third difference:

$$ \Delta^3 X_t = \Delta (\Delta^2 X_t) = (X_t - 2X_{t-1} + X_{t-2}) - (X_{t-1} - 2X_{t-2} + X_{t-3}) = X_t - 3X_{t-1} + 3 X_{t-2} - X_{t-3}$$

So, $\Delta^dX_t$ is indeed a linear combination of $X_t, X_{t-1}, ..., X_{t-d}$, but not the one you suggested.

In an ARIMA(p,d,q) model, $\Delta^dX_t$ follows an ARMA(p,q) process, that's correct.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.