Not matching the LSTM forecast with training model

Question

I use R and create a cosine vector just for testing the LSTM model and define a window (time step) lag=3 and want to predict from the 4th position (I mean shift target value