Starting to work with ml_walkforward, I noticed there is not an input for step_size/num_steps/timesteps/etc used in timeseries models (LSTM for example). Are timeseries models not supported with ml_walkforward?
Data window sizes are controlled by the
rolling_train arguments to
ml_walkforward. The model will successively be fit with the resulting windows of data. For incremental training, the walkforward analyzer will call
model.partial_fit(...) for scikit-learn models and
model.train_on_batch(...) for keras models. For non-incremental training,
model.fit(...) will be called.
ml_walkfoward may not be appropriate for every model or use case. Some users train their models without using
ml_walkforward then incorporate the model predictions into standard Moonshot or Zipline.