Hands-On Machine Learning for Algorithmic Trading

Defining and training the model

We use a similar architecture with two stacked LSTM layers with 12 and 6 units, respectively, followed by a fully-connected layer with 10 units. The output layer has two units, one for each time series. We compile them using mean absolute loss and the recommended RMSProp optimizer, as follows:

n_features = output_size = 2lstm1_units = 12lstm2_units = 6rnn = Sequential([    LSTM(units=lstm1_units,         dropout=.2,         recurrent_dropout=.2,         input_shape=(window_size, n_features), name='LSTM1',         return_sequences=True),    LSTM(units=lstm2_units,         dropout=.2,         recurrent_dropout=.2,         name='LSTM2'),    Dense(10, name='FC1'),    Dense(output_size, name='Output')])rnn.compile(loss='mae', optimizer='RMSProp')

The model has 1,268 parameters, ...