Well you could try to do that. In this example I have used the data from 2015 and 2016 as test and the rest as training. With the model and the data for 2018 you could also predict the target. You just have to be careful what your target means. If it is the sales from 2018 the question would be when would you know the Target and when the other data. So in oder to make an actual prediction you would need the explaining variables before you would need the prediction (Target).
I the example workflow the best model is H2O GLM. Although the prediction for 2015 is not very good.
Row0 is 2015 and Row1 is 2016 (H2O omitted the RowIDs). 2016 is OK, 2015 not so much.
And as has been mentioned in the other discussions I referenced. With such few data points and a complicated market it is not so easy to make an accurate prediction. Some additional data might help. Or you could eliminate non-standard years (if there is such a thing).