All model outputs as JSON?

[jdebacker]

OG-Core has been using pkl as the format for saving model outputs such as SS_vars.pkl, TPI_vars.pkl, and model_params.pkl.

The binary file type makes is impossible to see diffs, is not human readable, and poses more security risks.

In addition, the custom Specifications class object saved to model_params.pkl is saved with CloudPickle and cannot be saved and opened in different version of Python and so is not a good file type for long-term storage or for sharing across users/platforms.

What about using JSON for all model output?

• Steady-state results saved to SS_vars.pkl are in a single dictionary with values that are scalars or Numpy arrays. This dictionary could easily be saved to JSON.
• Transition path results saved to SS_vars.pkl are in a single dictionary with values that are scalars or Numpy arrays. This dictionary could easily be saved to JSON.
• Model parameters saved to model_params.pkl are almost all data types that can be put into a JSON. The exception is if tax functions use the mono or mono2D type.

Should we move over to JSON? The issue with certain tax function types (that are rarely used at this point) could be solved by saving just those function to a pickle file in parameterizations of the model that use those functions.

cc @rickecon

[jdebacker]

Two notes about JSON for model output:

1. The TPI_vars.pkl is about 33MB. A JSON with the same data is about 5x as large: 154 MB.
2. To save SS and TPI output to JSON, all arrays need to be converted to lists (and the reverse would be needed when reading them back in).