Can I ask a very simple question about LSTM time-series data prediction?
I have 1258 data points which are Amazon's historical stock price by day unit. And I build the below data structure in order to train the model.
trainX.shape : (840, 28, 1)
trainY.shape : (840, 1)
testX.shape : (360, 28, 1)
testY.shape : (360, 1)
For this training, I took a single feature that is 'close value' and just ignore other features (eg. open, high, low, volume so on..)
and below is the model structure that I use.
nNeuron = 512
batch_size = 40
seq_length = 28
input_data_column_cnt = 1
dropout = 0.7
epoch_num = 100
model = Sequential()
model.add(LSTM(nNeuron, batch_input_shape = (batch_size, seq_length, input_data_column_cnt), return_sequences = True, stateful = True, dropout = dropout))
model.add(LSTM(nNeuron, input_shape=(batch_size, input_data_column_cnt), return_sequences = True, dropout = dropout))
model.add(LSTM(nNeuron, return_sequences = True, stateful = True, dropout = dropout))
model.add(LSTM(nNeuron, return_sequences = True, stateful = True, dropout = dropout))
model.add(Dense(1))
model.add(Activation('softsign'))
model.summary()
model.compile(loss = 'mean_squared_error', optimizer = 'adam')
early_stopping = [ EarlyStopping(monitor = 'loss', patience = 5, mode = 'auto'),
ModelCheckpoint(filepath = 'test_mode.h5', monitor = 'loss', save_best_only = True)]
hist = model.fit(trainX, trainY, epochs = epoch_num, batch_size = batch_size, verbose = 1, callbacks = early_stopping)
Above model is working now.
My Question is, If I want to predict multi-step data(for example, I want to predict 7 days of future data), the data structure above should be like this?
trainX.shape : (840, 28, 1)
trainY.shape : (840, 7)
testX.shape : (360, 28, 1)
testY.shape : (360, 7)
Due to my stupid English skill, I cannot fully understand many blog tutorials. Sorry for the low-level question.
Thank you.
