I'm practicing with some data with a LSTM neural nets to come up with predicted data, comparing with actual data. I generated an image to show what I came up with.
The blue line is actual data, and red line is predicted data. You can see near the end (right side), after the blue line ends, that's where the prediction starts for future data (the red line before that is for predictions on trained data, so it makes sense to have the error low).
What I am noticing is that the predicted future data just seems very consistent (straight line), it doesn't seem to have any curves to it. How I trained it is by taking the past 20 data points, getting average and then the output is the average. So the output is 1 value for the future. Then what I do is take the previous 19 plus the new predicted, and use that as the new input to get the 2nd future value. I repeat that until I get 20 future values (each one has a accumulated error of the previous one).
Is this how people usually handle trying to get multiple future values?