I'm quite newbie to time series analysis and I have to understand what's the difference between differencing time series (i.e considering $Y_t= X_t-X_{t-1}$) and detrending (using linear regression for example) the series to make a time series stationary. I've read in my book that these are two diffent approaches but I don't understand which is better in which context.
2 Answers
Hi: It depends on what the DGP of the original process is. Is the process trend stationary or difference stationary ? If it's trend stationary then de-trending is the way to go. If it's difference stationary, then differencing is the way to go.
The two models are quite different:
Trend Stationary: $y_t = \beta_{0} + \beta_1 \times t + \epsilon_t$
Difference Stationary:
$y_t = u_{t} + \epsilon_t $
$u_t = u_{t-1} + \omega_t$
In the early 1980's, Nelson and Plosser (link to paper below) found that a lot of econometric series that were though to be trend-stationary were actually difference stationary and this caused an explosion of research on the question of difference versus trend stationary.
-
$\begingroup$ Thank you, maybe i get it. But from a statistic point of view, how I can distinguish the real DGP? I'm trying to model netflix (NFLX) Adjusted Close Price and related log-return but I really can't find it. $\endgroup$– perseoDec 24, 2019 at 18:26
-
$\begingroup$ More details: I was trying to model the time series with an ARIMA(p,1,q)+GARCH(m,n) in R but i got no stable model for the ARIMA part; in fact i got many NAs in fitted value, probably because of Hessian troubles. Now i just de-trended the ts with a simple linear model and I got an AR(1). I guess ADF test can relate but i can't understand how to distinguish the two models. $\endgroup$– perseoDec 24, 2019 at 18:42
-
1$\begingroup$ There are tests that can be used but your project sounds large. If your trying to model Netflix, I would use returns rather than prices. That might get rid of a lot of the non-stationarity, be it trend or difference. See how that goes because using prices for prediction is not a good idea. Prices are EXTREMELY non-stationary. Returns might be but not as much. I would also do things in steps. Don't just fit arima-garch. Try to see what kind of arima model looks reasonable and then worry about adding garch in after that. $\endgroup$ Dec 24, 2019 at 20:22
Let me try to write formulae to explain the differences:
- When $X_t=a+b\,t + c\,\xi_t$, where $\xi_t$ is an iid centered and reduced noise (ie $\mathbb{E}\xi=0$ and $\mathbb{E}\xi^2=1$.
With $X_(t+1)-X_t=b + c\Delta\xi$, you read that you increased the amplitude of the noise $\xi$ by a factor $\sqrt{2}$, you removed $a$ and you have no more time dependent.
- When $X_t=a+b\,t + c\,W_t$ where $W$ is a Wiener process, ie $dX=b\, dt + c\, dW$ and $dW\sim {\cal N}(0,1)$.
Here it is more natural to immediately look at $dX$: you removed the constant, and this time you reduced the amplitude of the noise. Moreover, since $\mathbb{E}W_t^2=t$ you removed the heteroskedasticity of the process (see Bollerslev's papers).
For stock returns, just take $dX=\frac{dP}{P}$ and you are closer to the second case. If you consider daily prices, we know the model should be more sophisticated:
- you should have a jump component because $dW$ is too regular (see Cont and Tankov's book)
- you should write $c_t$ because the volatility is time dependent, and especially it is clustered (have a look at Rob Engle's Nobel lecture).