estimagic

Introduction

estimagic is a Python package that provides high-quality and user-friendly tools to fit large scale empirical models to data and make inferences about the estimated model parameters. It is especially suited to solve difficult constrained optimization problems.

estimagic provides several advantages over similar packages, including a unified interface that supports a large number of local and global optimization algorithms and the possibility of monitoring the optimization procedure via a beautiful interactive dashboard.

estimagic provides tools for nonlinear optimization, numerical differentiation and statistical inference.

Optimization

• estimagic wraps all algorithms from scipy.optimize and many more become available when installing optional dependencies.
• estimagic can automatically implement many types of constraints via reparametrization, with any optmizer that supports simple box constraints.
• estimagic encourages name-based parameters handling. Parameters are specified as pandas DataFrames that can have any kind of single or MultiIndex. This is especially useful when specifying constraints.
• The complete history of parameters and function evaluations are saved in a database for maximum reproducibility and displayed in real time via an interactive dashboard.

Numerical differentiation

• estimagic can calculate precise numerical derivatives using Richardson extrapolations.
• Function evaluations needed for numerical derivatives can be done in parallel with pre-implemented or user provided batch evaluators.

Statistical Inference

• estimagic provides asymptotic standard errors for maximum likelihood and method of simulated moments.
• estimagic also provides bootstrap confidence intervals and standard errors. Of course the bootstrap procedures are parallelized.

Installation

The package can be installed via conda. To do so, type the following commands in a terminal:

$ conda config --add channels conda-forge
$ conda install -c opensourceeconomics estimagic

The first line adds conda-forge to your conda channels. This is necessary for conda to find all dependencies of estimagic. The second line installs estimagic and its dependencies.

Documentation

The documentation is hosted (on rtd)

Citation

If you use Estimagic for your research, please do not forget to cite it. Many people worked on this software and you should recognize their effort.

@Unpublished{Gabler2020,
  Title  = {A Python Tool for the Estimation of (Structural) Econometric Models.},
  Author = {Janos Gabler},
  Year   = {2020},
  Url    = {https://github.com/OpenSourceEconomics/estimagic}
}

Warning

Estimagic is still in beta status and the API might still change. With version 0.2.0 we start to mark experimental functionality with warnings. Everything that does not throw such warnings will not change without notice. With version 0.3.0 we will try to keep the API for full estimation workflows (optimization, inference and sensitivity analysis) more stable and use deprecation cycles for breaking changes.

Roadmap

Version 0.2.0

• Raise warnings for all things that are still considered experimental. Conversely, everything that does not throw an experimental warning is from now on stable and breaking changes require deprecations.
• estimagic.optimization.utilities becomes estimagic.utilities (breaking change)
• Return of first_derivative becomes a dict (breaking change)
• Logging disabled by default (potentially breaking change)
• Support bootstrap (experimental)
• Support basic sensitivity analysis for moments based estimation (experimental)
• Support scaling of optimization problems (experimental)

Version 0.2.1

• Support multi start optimizations as in TikTak (experimental)

Version 0.2.2

• Improve packaging and upload on conda-forge

• Split up into several packages to keep runtimes for the test suites manageable

  • estimagic-optimization
  • estimagic-differentiation
  • estimagic-inference

• Internal refactoring of constraints code

Version 0.2.x

Improve interfaces of inference functions but still keep them experimental

Version 0.3.0 (getting rid of experimental warnings)

• Find a good example model that can be estimated with ML, MSM, GMM and II for test cases and documentation
• Have stable interfaces for all inference and sensitivity analysis functions
• Make tutorials that show full workflows for each estimation principle
• Promote estimagic via blogposts

Additionally planned Features

• Wrap nlopt and ipopt
• Implement a flexible toolkit for optimization of noisy functions
• Wrap pygmo
• Add code for numerical second derivatives
• Make dashboard options configurable in GUI
• Improve the appearance of the current Dashboard
• Add a dashboard tab for numerical derivatives
• Add a dashboard tab for bootstrapping