This is what a default config looks like.

It has 4 top level keys :

We may add a 5th key for whatifs scenarios, but that depends on how formally intergrated that becomes to the end to end pipeline.

Fitting

data :
- tables : list of tables to retrieve from Athena DB
- schema : name of the schema in Athena DB
- staging_dir : name of the S3 staging directory
- label : label used to name all associated artifacts
model :
variable_param_ranges :
default_params :
optimiser :
optimiser_params : str. Can be GridSearch or BO_Hyperopt
- num_evals: int
- algo: str. Can be tpe (TPE), atpe (adaptive TPE), or rand (random search)
- seed: int
split :
- start_date: date. Start of training period. Default null
- end_date: date. End of training period. Default null
- train_period: int. Length of training period.
- val_period: int. Length of validation period.
- test_period: int. Length of test period. Giving train, val and test period is a must. If both start and end date are null, it fits on latest data. Both can't be given, only 1 of them can be given. Otherwise an error will be thrown.
loss :
- loss_method : str. Can be rmse, mape, or rmsle
- loss_compartments : [str]
- loss_weights : [float]

Forecast

forecast_days : int.
num_trials_to_plot : int. Num trials to plot in the plot_topk_trials functions
plot_topk_trials_for_columns : [str]. Which columns to plot topk trials for.
plot_ptiles_for_columns :[str]. Which columns to plot percentiles for.

Parameters for each uncertainty class come inside uncertainty_params.

For ABMAUncertainty :

num_evals : int
optimiser: str. Can be GridSearch or BO_Hyperopt For GridSearch -
optimiser_params:
- parallelise: bool Specify the range for beta -
variable_param_ranges:
- beta: [[0, 10], 101] (Eg)
construct_percentiles_day_wise: bool. If true, percentiles are constructed day wise.
date_of_sorting_trials : date
sort_trials_by_column : str
loss :
- loss_method : str. Can be rmse, mape, or rmsle
- loss_compartments : [str]
- loss_weights : [float]
percentiles: [float]. Which percentiles to compute.