I have a code which predicts values using three different models: lm, auto.arima and randomForest. I do forecast for the next day, using a rolling window. For instance, I take a time window of 500 days and forecast the next day's value. Then move the widow to data from 2 to 501, and forecast for the day 502 and so on.
I wonder if for this kind of forecast I can use sMAPE measure, since I am only forecasting one day ahead.
Whether you can use the sMAPE does not depend on how many days ahead you are forecasting.
If you evaluate only this one time point for each series, then each series is not averaged, so you only get a sAPE for each separate series. But you will likely later average these sAPEs over many series, ending up with an sMAPE.
Note that the sMAPE shares many shortcomings of the plain MAPE. One issue is particular to the sMAPE: if the actual is $y=0$, then the sAPE is 200%, regardless of the forecast $\hat{y}$. This is usually not what you want, since for an actual of $y=0$, a forecast of $\hat{y}=1$ is better than one of $\hat{y}=10$ - but the sAPE cannot distinguish between the two. Thus, the sMAPE is not very useful if you have many zeros.
