Hands-On Machine Learning for Algorithmic Trading

Feedforward autoencoder with sparsity constraints

The addition of regularization is fairly straightforward. We can apply it to the dense encoder layer using Keras' activity_regularizer, as follows:

encoding_l1 = Dense(units=encoding_size,                    activation='relu',                    activity_regularizer=regularizers.l1(10e-5),                    name='Encoder_L1')(input_)

The input and decoding layers remain unchanged. In this example, with a compression of factor 24.5, regularization negatively affects performance with a test RMSE of 0.2946.