curvefit.uncertainty.draws.Draws

A class for generating draws: predictions plus random residuals according to provided ResidualModel

This class simulates possible trajectories by adjusting out-of-sample part of the prediction with expected residuals from ResidualModel. The user-controlled parameters are passed via init; the parameters which need to be consistent to other parts of the pipeline (like model, solver, evaluation_space, etc) are passed via create_draws(...).

Arguments

• num_draws (int): the number of draws to take
• prediction_times (np.array): which times to produce final predictions (draws) at

Methods

create_draws

Fills in the `draws` dictionary by generating `num_groups` sample trajectories per location

- `data (curvefit.core.data.Data)`: data
- `model_prototype (curvefit.models.base.Model)`: a model object
- `solver_prototype (curvefit.solvers.solver.Solver)`: a solver used to fit the model
- `residual_model (curvefit.uncertainty.residual_model._ResidualModel)' a residual model object, should be
     fitted before being passed here.
- `evaluation_space (callable)`: which space to generate draws in. It should be the same as used
    in `PredictiveValidity` instance which generated the residual matrix passed to the `residual_model`
- `theta (float)`: between 0 and 1, how much scaling of the residuals to do relative to the prediction mean

get_draws

Returns generated draws

- `group (str)`: Group for which the draws are requested. Defaults to None, in which case returns
    a dictionary of all draws indexed by groups.

get_draws_summary

Returns mean and (lower, higher) quantiles of draws

- `group (str)`: Group for which draws summary is requested. Defaults to None, in which case returns
    a dictionary of statistics for all draws indexed by groups.
- `quantiles (float)`: which quantiles to generate. Should be between 0 and 0.5. For instance,
    if quantiles = 0.05 is passed then it returns 5'th and 95'th percentiles.

Usage

In ModelRunner.run():

    draws = self.draws.create_draws(
        data=self.data,
        model_prototype=self.model,
        solver_prototype=self.solver,
        residual_model=self.residual_model,
        evaluation_space = self.predictive_validity.evaluation_space,
        theta=self.predictive_validity.theta
    ).get_draws()