Heterogeneous Agents Resources and toolKit (HARK)¶
Click the Badge for Citation Info. DOI
Table of Contents:
- 1. Introduction
- 2. Quick start guide
- 3. List of files in repository
- 4. Warnings and disclaimers
- 5. License Information
This document will tell you how to get HARK up and running on your machine, how to get started using it, and give you an overview of the main elements of the toolkit.
Other useful resources are:
II. QUICK START GUIDE¶
HARK is an open source project that is compatible with both python 2 and 3. But we recommend using python 3; eventually support for python 2 will end.
Installing HARK with pip¶
The simplest way to install HARK is to use pip.
To install HARK with pip, at a command line type
pip install econ-ark.
(If you want to install a release that is not the default stable release, for instance if you want to install a development release, you’ll need to use a “pinned” release number:
pip install econ-ark==0.10.1.dev1, substituting “0.10.1.dev1” for your desired release number.)
If you are installing via pip, we recommend using a virtual environment such as virtualenv. Creation of a virtual environment isolates the installation of
econ-ark from the installations of any other python tools and packages.
virtualenv, then to create an environment named
econ-ark, and finally to activate that environment:
cd [directory where you want to store the econ-ark virtual environment] pip install virtualenv virtualenv econ-ark source econ-ark/bin/activate
Using HARK with Anaconda¶
If you intend ever to use the toolkit for anything other than running the precooked material we have provided, you should probably install Anaconda, which will install python along with many packages that are frequently used in scientific computing.
- Download Anaconda for your operating system and follow the installation instructions at Anaconda.com.
- Anaconda includes its own virtual environment system called
condawhich stores environments in a preset location (so you don’t have to choose). So in order to create and activate an econ-ark virtual environment:
conda create -n econ-ark anaconda conda activate econ-ark conda install -c conda-forge econ-ark
Open Spyder, an interactive development environment (IDE) for Python (specifically, iPython). You may be able to do this through Anaconda’s graphical interface, or you can do so from the command line/prompt. To do so, simply open a command line/prompt and type
To verify that spyder has access to HARK try typing
pip install econ-arkinto the iPython shell within Spyder. If you have successfully installed HARK as above, you should see a lot of messages saying ‘Requirement satisfied’.
- If that doesn’t work, you will need to manually add HARK to your Spyder environment. To do this, you’ll need to get the code from Github and import it into Spyder. To get the code from Github, you can either clone it or download a zipped file.
- If you have
gitinstalled on the command line, type
git clone email@example.com:econ-ark/HARK.gitin your chosen directory (more details here).
- If you do not have
gitavailable on your computer, you can download the GitHub Desktop app and use it to make a local clone
- If you do not have
- If you don’t want to clone HARK, but just to download it, go to the HARK repository on GitHub. In the upper righthand corner is a button that says “clone or download”. Click the “Download Zip” option and then unzip the contents into your chosen directory.
Once you’ve got a copy of HARK in a directory, return to Spyder and navigate to that directory where you put HARK. This can be done within Spyder by doing
import osand then using
os.chdir()to change directories.
chdirworks just like cd at a command prompt on most operating systems, except that it takes a string as input:
os.chdir('Music')moves to the Music subdirectory of the current working directory.
- Most of the modules in HARK are just collections of tools. There are a few demonstration applications that use the tools that you automatically get when you install HARK – they are listed below in Application Modules. A much larger set of uses of HARK can be found at two repositories:
You will want to obtain your own local copy of these repos using:
git clone https://github.com/econ-ark/DemARK.git
and similarly for the REMARK repo. Once you have downloaded them, you will find that each repo contains a
notebooks directory that contains a number of jupyter notebooks. If you have the jupyter notebook tool installed (it is installed as part of Anaconda), you should be able to launch the
jupyter notebook app from the command line with the command:
and from there you can open the notebooks and execute them.
We have a set of 30-second Elevator Spiels describing the project, tailored to people with several different kinds of background.
- A full replication of the Iskhakov, Jørgensen, Rust, and Schjerning paper for solving the discrete-continuous retirement saving problem
- Structural-Estimates-From-Empirical-MPCs is an example of the use of the toolkit in a discussion of a well known paper. (Yes, it is easy enough to use that you can estimate a structural model on somebody else’s data in the limited time available for writing a discussion)
- Start with A Gentle Introduction to HARK to get your feet wet
- A simple indirect inference/simulated method of moments structural estimation along the lines of Gourinchas and Parker’s 2002 Econometrica paper or Cagetti’s 2003 paper is performed by the SolvingMicroDSOPs REMARK; this code implements the solution methods described in the corresponding section of these lecture notes
Making changes to HARK¶
If you want to make changes or contributions (yay!) to HARK, you’ll need to have access to the source files. Installing HARK via pip (either at the command line, or inside Spyder) makes it hard to access those files (and it’s a bad idea to mess with the original code anyway because you’ll likely forget what changes you made). If you are adept at GitHub, you can fork the repo. If you are less experienced, you should download a personal copy of HARK again using
git clone (see above) or the GitHub Desktop app.
- Navigate to wherever you want to put the repository and type
git clone firstname.lastname@example.org:econ-ark/HARK.git(more details here). If you get a permission denied error, you may need to setup SSH for GitHub, or you can clone using HTTPS: ‘git clone https://github.com/econ-ark/HARK.git’.
- Then, create and activate a virtual environment.
For Mac or Linux:
Install virtualenv if you need to and then type:
virtualenv econ-ark source econ-ark/bin/activate
virtualenv econ-ark econ-ark\\Scripts\\activate.bat
- Once the virtualenv is activated, you may see
(econ-ark)in your command prompt (depending on how your machine is configured)
- Make sure to change to HARK directory, and install HARK’s requirements into the virtual environment with
pip install -r requirements.txt.
- To check that everything has been set up correctly, run HARK’s tests with
python -m unittest.
Trouble with installation?¶
We’ve done our best to give correct, thorough instructions on how to install HARK but we know this information may be inaccurate or incomplete. Please let us know if you run into trouble so we can update this guide! Here’s a list of platforms and versions this guide has been verified for:
| Installation Type | Platform | Python Version | Date Tested | Tested By | | ————- |:————-:| —–:| —–:|—–:| | basic pip install | Linux (16.04) | 3 | 2019-04-24 | @shaunagm | | anaconda | Linux (16.04) | 3 | 2019-04-24 | @shaunagm | | basic pip install | MacOS 10.13.2 “High Sierra” | 2.7| 2019-04-26 | @llorracc |
III. LIST OF FILES IN REPOSITORY¶
This section contains descriptions of the main files in the repo.
- README.md: The file you are currently reading.
- Documentation/HARKdoc.pdf: A mini-user guide produced for a December 2015 workshop on HARK, unofficially representing the alpha version. (Substantially out of date).
- Documentation/HARKmanual.pdf: A user guide for HARK, written for the beta release at CEF 2016 in Bordeaux. Should contain 90% fewer lies relative to HARKdoc.pdf.
- Documentation/HARKmanual.tex: LaTeX source for the user guide. Open source code probably requires an open source manual as well.
- Documentation/ConsumptionSavingModels.pdf: Mathematical descriptions of the various consumption-saving models in HARK and how they map into the code.
- Documentation/ConsumptionSavingModels.tex: LaTeX source for the “models” writeup.
- Documentation/NARK.pdf: Variable naming conventions for HARK, plus concordance with LaTeX variable definitions. Still in development.
- HARK/core.py: Frameworks for “microeconomic” and “macroeconomic” models in HARK. We somewhat abuse those terms as shorthand; see the user guide for a description of what we mean. Every model in HARK extends the classes AgentType and Market in this module. Does nothing when run.
- HARK/utilities.py: General purpose tools and utilities. Contains literal utility functions (in the economic sense), functions for making discrete approximations to continuous distributions, basic plotting functions for convenience, and a few unclassifiable things. Does nothing when run.
- HARK/estimation.py: Functions for estimating models. As is, it only has a few wrapper functions for scipy.optimize optimization routines. Will be expanded in the future with more interesting things. Does nothing when run.
- HARK/simulation.py: Functions for generating simulated data. Functions in this module have names like drawUniform, generating (lists of) arrays of draws from various distributions. Does nothing when run.
- HARK/interpolation.py: Classes for representing interpolated function approximations. Has 1D-4D interpolation methods, mostly based on linear or cubic spline interpolation. Will have ND methods in the future. Does nothing when run.
- HARK/parallel.py: Early version of parallel processing in HARK. Works with instances of the AgentType class (or subclasses of it), distributing commands (as methods) to be run on a list of AgentTypes. Only works with local CPU. The module also contains a parallel implentation of the Nelder-Mead simplex algorithm, poached from Wiswall and Lee (2011). Does nothing when run.
- A “tractable” model of consumption and saving in which agents face one simple risk with constant probability: that they will become permanently unemployed and receive no further income. Unlike other models in HARK, this one is not solved by iterating on a sequence of one period problems. Instead, it uses a “backshooting” routine that has been shoehorned into the AgentType.solve framework. Solves an example of the model when run, then solves the same model again using MarkovConsumerType.
- Consumption-saving models with idiosyncratic shocks to income. Shocks are fully transitory or fully permanent. Solves perfect foresight model, a model with idiosyncratic income shocks, and a model with idiosyncratic income shocks and a different interest rate on borrowing vs saving. When run, solves several examples of these models, including a standard infinite horizon problem, a ten period lifecycle model, a four period “cyclical” model, and versions with perfect foresight and “kinked R”.
- Consumption-saving models with idiosyncratic shocks to income and multi- plicative shocks to utility. Currently has two models: one that extends the idiosyncratic shocks model, and another that extends the “kinked R” model. The second model has very little new code, and is created merely by merging the two “parent models” via multiple inheritance. When run, solves examples of the preference shock models.
- Consumption-saving models with a discrete state that evolves according to a Markov rule. Discrete states can vary by their income distribution, interest factor, and/or expected permanent income growth rate. When run, solves four example models: (1) A serially correlated unemployment model with boom and bust cycles (4 states). (2) An “unemployment immunity” model in which the consumer occasionally learns that he is immune to unemployment shocks for the next N periods. (3) A model with a time-varying permanent income growth rate that is serially correlated. (4) A model with a time- varying interest factor that is serially correlated.
- Consumption-saving models with idiosyncratic and aggregate income shocks. Currently has a micro model with a basic solver (linear spline consumption function only, no value function), and a Cobb-Douglas economy for the agents to “live” in (as a “macroeconomy”). When run, solves an example of the micro model in partial equilibrium, then solves the general equilibrium problem to find an evolution rule for the capital-to-labor ratio that is justified by consumers’ collective actions.
- A very serious model about choosing to dress as a jock or a punk. Used to demonstrate micro and macro framework concepts from HARKcore. It might be the simplest model possible for this purpose, or close to it. When run, the module solves the microeconomic problem of a “fashion victim” for an example parameter set, then solves the general equilibrium model for an entire “fashion market” constituting many types of agents, finding a rule for the evolution of the style distribution in the population that is justi- fied by fashion victims’ collective actions.
- Conducts a very simple structural estimation using the idiosyncratic shocks model in ConsIndShocksModel. Estimates an adjustment factor to an age-varying sequence of discount factors (taken from Cagetti (2003)) and a coefficient of relative risk aversion that makes simulated agents’ wealth profiles best match data from the 2004 Survey of Consumer Finance. Also demonstrates the calculation of standard errors by bootstrap and can construct a contour map of the objective function. Based on section 9 of Chris Carroll’s lecture notes “Solving Microeconomic Dynamic Stochastic Optimization Problems”.
- Conducts the estimations for the paper “The Distribution of Wealth and the Marginal Propensity to Consume” by Carroll, Slacalek, Tokuoka, and White (2016). Runtime options are set in SetupParamsCSTW.py, specifying choices such as: perpetual youth vs lifecycle, beta-dist vs beta-point, liquid assets vs net worth, aggregate vs idiosyncratic shocks, etc. Uses ConsIndShockModel and ConsAggShockModel; can demonststrate HARK’s “macro” framework on a real model.
- Makes various figures for the text of the cstwMPC paper. Requires many output files produced by cstwMPC.py, from various specifications, which are not distributed with HARK. Has not been tested in quite some time.
- Makes various figures for the slides for the cstwMPC paper. Requires many output files produced by cstwMPC.py, from various specifications, which are not distributed with HARK. Has not been tested in quite some time.
Parameter and data modules:
- Defines dictionaries with the minimal set of parameters needed to solve the models in ConsIndShockModel, ConsAggShockModel, ConsPrefShockModel, and ConsMarkovModel. These dictionaries are used to make examples when those modules are run. Does nothing when run itself.
- Imports 2004 SCF data for use by SolvingMicroDSOPs/StructEstimation.py.
- Loads calibrated model parameters for cstwMPC.py, chooses specification.
- Example parameters for FashionVictimModel.py, loaded when that module is run.
- Early version of unit testing for HARK, still in development. Compares the perfect foresight model solution to the idiosyncratic shocks model solution with shocks turned off; also compares the tractable buffer stock model solution to the same model solved using a “Markov” description.
- Early version of unit testing for HARK, still in development. Defines a few wrapper classes to run unit tests on subclasses of AgentType.
- Early version of unit testing for HARK, still in development. Runs a test on TractableBufferStockModel.
- Demonstrates the multithreading functionality in HARKparallel.py. When run, it solves oneexample consumption-saving model with idiosyncratic shocks to income, then solves many such models serially, varying the coefficient of relative risk aversion between rho=1 and rho=8, displaying the results graphically and presenting the timing. It then solves the same set of many models using multithreading on the local CPU, displays the results graphically along with the timing.
- SCF 2004 data for use in SolvingMicroDSOPs/StructEstimation.py, loaded by SolvingMicroDSOPs/EstimationParameters.py.
- SCF 2004 data with just net worth and data weights, for use by cstwMPC.py
- U.S. mortality data from the Social Security Administration, for use by cstwMPC.py when running a lifecycle specification.
- Mortality adjusters by education and age (columns by sex and race), for use by cstwMPC.py when running a lifecycle specification. Taken from Brown et al. (2002).
Other files that you don’t need to worry about:
- A file used by Sphinx when generating html documentation for HARK. Users don’t need to worry about it. Several copies are found throughout HARK.
- A file that tells git which files (or types of files) might be found in the repository directory tree, but should be ignored (not tracked) for the repo. Currently ignores compiled Python code, LaTex auxiliary files, etc.
- License text for HARK, Apache 2.0. Read it if you’re a lawyer!
- Contour plot of the objective function for SolvingMicroDSOPs/StructEstimation.py. Generated when that module is run, along with a PDF version.
- A placeholder file because git doesn’t like empty folders, but cstwMPC.py needs the /Figures directory to exist when it runs.
- A configuration file for producing html documentation with Sphinx, generated by sphinx-quickstart.
- A very small file used by Sphinx to produce documentation.
- A list of modules to be included in HARK’s Sphinx documentation. This should be edited if a new tool or model module is added to HARK.
- A markdown file with instructions for how to set up and run Sphinx. You don’t need to read it.
- Another markdown file with instructions for how to set up and run Sphinx.
- A batch file for producing Sphinx documentation, generated by sphinx-quickstart.
- Another Sphinx auxiliary file generated by sphinx-quickstart.
- LaTeX style file with notation definitions.
- LaTeX class file with document layout for the user manual.
- LaTeX style file with notation definitions.
- LaTeX style file with notation definitions.
- Image for the front cover of the user guide, showing the consumption function for the KinkyPref model.
IV. WARNINGS AND DISCLAIMERS¶
This is a beta version of HARK. The code has not been extensively tested as it should be. We hope it is useful, but there are absolutely no guarantees (expressed or implied) that it works or will do what you want. Use at your own risk. And please, let us know if you find bugs by posting an issue to the GitHub page!
All of HARK is licensed under the Apache License, Version 2.0 (ALv2). Please see the LICENSE file for the text of the license. More information can be found at: http://www.apache.org/dev/apply-license.html