time-series-forecasting-rnn-tensorflow
time-series-forecasting-rnn-tensorflow copied to clipboard
jiegzhan
Why you shuffle the sequence of training data?
In the pre- processing before entering network model, why you shuffle the sequence of training data? Since the sequence is important in time series!
Train data is a normalized view of the raw data that has been split into windows. Please see this logic: # Split data into windows raw = [] for index in range(len(data) - adjusted_window): raw.append(data[index: index + adjusted_window])
Meaning that train data is also cut into windows, E.g. for window size four you get:
[[ 0. -0.00289975 0.00517813 -0.01367026 -0.04971002] [ 0. 0.00810137 -0.01080183 -0.04694641 -0.02928957] [ 0. -0.01875129 -0.0546054 -0.03709046 -0.00535751] ... [ 0. -0.01493441 -0.01937437 -0.04702321 -0.02583249] [ 0. -0.00450727 -0.03257529 -0.01106331 -0.03605818] [ 0. -0.0281951 -0.00658572 -0.03169376 -0.033546 ]]
Where first four elements of each sub-array are inputs (X) and the last one is the actual value (Y).
It is perfectly valid to shuffle those sub-arrays as all that matters is that the sequence of elements is preserved within each individual sub-array.
So shuffling is not an issue.
