Suppose I have a non-stationary limited data. Do I have to make it stationary before making forecasts? Can I use exponential smoothing, moving averages or even Holt Winters methods without making my data stationary? I've read that time series should be stationary to use such methods? Is that correct?
$\begingroup$
$\endgroup$
Add a comment
|
3 Answers
$\begingroup$
$\endgroup$
1
I don't know what "non-stationary limited data" means. So I will assume you mean "non-stationary data".
Exponential smoothing methods including Holt-Winters methods are appropriate for (some kinds of) non-stationary data. In fact, they are only really appropriate if the data are non-stationary. Using an exponential smoothing method on stationary data is not wrong but is sub-optimal.
If by "moving averages", you mean forecasting using a moving average of recent observations, then that is also ok for some kinds of non-stationary data. But it obviously won't work well with trends or seasonality.
If by "moving averages", you mean a moving average model (i.e. a model consisting of a linear combination of past error terms), then you do need a stationary time series.
answered Nov 20, 2013 at 3:58
-
$\begingroup$ I'm not entirely following your reply: many text books mention that moving average models and various exponential smoothing models are equivalent. For example $ARIMA(0,1,1) \simeq \text {simple exponential smoothing} $ - but ARIMA(0,1,1) and MA(q) models require stationarity while you are saying that ES methods are suboptimal on stationary series? $\endgroup$ Commented Jan 18, 2018 at 19:03
$\begingroup$
$\endgroup$
There is no problem with forecasting nonstationary data directly. Take the random walk
$$Y_t=Y_{t-1}+\epsilon_t$$
The best predictor of $Y_{t+1}$ given the information set $\{Y_t,Y_{t-1},...\}$ simply is
$$E(Y_{t+1}|Y_t)=E(Y_t|Y_{t})+E(\epsilon_{t+1}|Y_t)=Y_t$$
answered Feb 18, 2015 at 15:54
$\begingroup$
$\endgroup$
1
Stationarity refers to uniformity in the properties of the data. If you know that the data is non-stationary, it means that the useful properties of the data cannot be assumed to be the same for the entire series. Under such an assumption, why do you want to apply the same filter or model to the entire series ?
My suggestion is to look for properties that stay the same for a stretch of data and then changes but again stays the same for another stretch. then look for a criteria to transition between the two different stretches of data.
Alternatively, search for locally stationary series.
Also if smoothing is what you want, then I would suggest some non-parametric smoothing methods like kernel smoothing.
Edit after first comment : if you know the precise form of non-stationarity, or can approximate a functional form to the series, then use the properties of the form for your prediction.
-
$\begingroup$ This answer is extremely misleading. There are very predictable non-stationary series, because the cause of non-stationarity may come from the deterministic part. What matters is the power of the deterministic component to the power of the stochastic component in the whole. For example $y_t = e^{y_{t-1}}+gaussian error$ can be filtered very easily although the series explodes and non-stationary by any definition. $\endgroup$ Commented Nov 20, 2013 at 14:00