I am registering this as an issue, but acknowledge that the following functionality may have always been out of scope for the package, considering the focus on auto-regression methods. Nonetheless, I think it worth mentioning.
Currently the toolbox does not appear to support discrete-time models (whether one-step or multi-step prediction) that do not included past values of the system output as an input.
To test whether this functionality exists, I tried passing "auto_order=0" to a NARX model. The result is a concatenation error in utils.generate_lag_features at the line:
lag_features = np.concatenate(lag_feature_list, axis=1)
This appears to be because a 1-D array associated with the output feature y is assigned to lag_feature_list. This empty 1-D array has different dimensionality to all the other lag_features (which are 2-D), and hence the concatenation error:
I raise the issue because it seems like it could be relatively straightforward to fix by adding a unique case for auto_order = 0 that doesn't generate a lag feature, which will avoid the failure of concatenation. I haven't looked at the potential implications for the predict and forecast function, but for now I will try and investigate this in a separate fork.
One of the things i appreciate about FireTS is its simplicity of execution and would like to keep using it, so having this additional utility for discrete-time models without autoregression on the output would make it even more flexible and useful for me!
I am registering this as an issue, but acknowledge that the following functionality may have always been out of scope for the package, considering the focus on auto-regression methods. Nonetheless, I think it worth mentioning.
Currently the toolbox does not appear to support discrete-time models (whether one-step or multi-step prediction) that do not included past values of the system output as an input.
To test whether this functionality exists, I tried passing "auto_order=0" to a NARX model. The result is a concatenation error in utils.generate_lag_features at the line:
lag_features = np.concatenate(lag_feature_list, axis=1)This appears to be because a 1-D array associated with the output feature y is assigned to lag_feature_list. This empty 1-D array has different dimensionality to all the other lag_features (which are 2-D), and hence the concatenation error:
I raise the issue because it seems like it could be relatively straightforward to fix by adding a unique case for auto_order = 0 that doesn't generate a lag feature, which will avoid the failure of concatenation. I haven't looked at the potential implications for the predict and forecast function, but for now I will try and investigate this in a separate fork.
One of the things i appreciate about FireTS is its simplicity of execution and would like to keep using it, so having this additional utility for discrete-time models without autoregression on the output would make it even more flexible and useful for me!