Heterogeneous Agents Resources and toolKit (HARK) beta release - June 25, 2016

Table of Contents:

• I. Introduction
• II. Quick start guide
• III. List of files in repository
• IV. Warnings and disclaimers
• V. License Information

I. INTRODUCTION

Welcome to HARK! We are tremendously excited you're here. HARK is very much a work in progress, but we hope you find it valuable. We really hope you find it so valuable that you decide to contribute to it yourself. This document will tell you how to get HARK up and running on your machine, and what you will find in HARK once you do.

If you have any comments on the code or documentation, we'd love to hear from you! Our email addresses are:

• Chris Carroll: ccarroll@llorracc.org
• Matthew White: mnwhite@gmail.com
• Nathan Palmer: Nathan.Palmer@ofr.treasury.gov
• David Low: David.Low@cfpb.gov
• Alexander Kaufman: akaufman10@gmail.com

GitHub repository: https://github.com/econ-ark/HARK

Online documentation: https://econ-ark.github.io/HARK

User guide: /Documentation/HARKmanual.pdf (in the repository)

II. QUICK START GUIDE

This is going to be easy, friend. HARK is written in Python, specifically the Anaconda distribution of Python. Follow these easy steps to get HARK going:

1. Go to https://www.continuum.io/downloads and download Anaconda for your operating system; be sure to get the version for Python 2.7

2. Install Anaconda, using the instructions provided on that page. Now you have installed everything you need to run most of HARK. But you still need to get HARK on your machine.

3. To get HARK on your machine, you should know that HARK is managed with version control software called "Git". HARK is hosted on a website called "GitHub" devoted to hosting projects managed with Git.

If you don't want to know more than that, you don't have to. Go to HARK's page on GitHub (https://github.com/econ-ark/HARK), click the "Clone or download" button in the upper right hand corner of the page, then click "Download ZIP". Unzip it into an empty directory. Maybe call that directory /HARK ? The choice is yours.

You can also clone HARK off GitHub using Git. This is slightly more difficult, because it involves installing Git on your machine and learning a little about how to use Git. We believe this is an investment worth making, but it is up to you. To learn more about Git, read the documentation at https://git-scm.com/documentation or visit many other great Git resources on the internet.

4. Open Spyder, an interactive development environment (IDE) for Python (specifically, iPython). On Windows, open a command prompt and type "spyder". On Linux, open the command line and type "spyder". On Mac, open the command line and type "spyder".

5. Navigate to the directory where you put the HARK files. This can be done within Spyder by doing "import os" and then using os.chdir() to change directories.
   chdir works 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.

6. Run one of HARK's modules. You can either type "run MODULENAME" after navigating to the correct directory (see step 5), or click the green arrow "run" button in Spyder's toolbar after opening the module in the editor. Every module should do something when run, but that something might not be very interesting in some cases. For starters, check out /ConsumptionSavingModel/ConsIndShockModel.py See section III below for a full list of modules that produce non-trivial output.

7. The Python environment can be cleared or reset with ctrl+. Note that this will also change the current working directory back to its default. To change the default directory (the "global working directory"), see Tools-->Preferences-->Global working directory; you might need to restart Spyder for the change to take effect.

8. Read the more complete documentation in HARKmanual.pdf.

9. OPTIONAL: If you want to use HARK's multithreading capabilities, you will need two Python packages that do not come automatically with Anaconda: joblib and dill. Assuming you have the necessary permissions on your machine, the easiest way to do this is through Anaconda. Go to the command line, and type "conda install joblib" and then "conda install dill" (accept defaults if prompted). If this doesn't work, but you have Git, you can just clone the packages directly off GitHub. Go to the command line and navigate to the directory you want to put these packages in. Then type "git clone https://github.com/joblib/joblib.git" and then "git clone https://github.com/uqfoundation/dill". Joblib should work after this, but there is one more step to get dill working. Navigate to dill's directory in the command line, and then type "python setup.py build". Then you should have joblib and dill working on your machine.

Note: If you did not put joblib and dill in one of the paths in sys.path, you will need to add the joblib and dill directories to sys.path. The easiest way to do this is to open up Anaconda, and type:

import sys sys.path.append('path_to_joblib_directory') sys.path.append('path_to_dill_directory')

III. LIST OF FILES IN REPOSITORY

This section contains descriptions of every file included in the HARK repository at the time of the beta release, categorized for convenience.

Documentation files:

• 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. Somewhat 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.

Tool modules:

• HARKcore.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.
• HARKutilities.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.
• HARKestimation.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.
• HARKsimulation.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.
• HARKinterpolation.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.
• HARKparallel.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.

Model modules:

• ConsumptionSavingModel/TractableBufferStockModel.py: 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.
• ConsumptionSavingModel/ConsIndShockModel.py: 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".
• ConsumptionSavingModel/ConsPrefShockModel.py: 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.
• ConsumptionSavingModel/ConsMarkovModel.py: 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.
• ConsumptionSavingModel/ConsAggShockModel.py: 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.
• FashionVictim/FashionVictimModel.py: 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.

Application modules:

• SolvingMicroDSOPs/StructEstimation.py: 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".
• cstwMPC/cstwMPC.py: 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.
• cstwMPC/MakeCSTWfigs.py: 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.
• cstwMPC/MakeCSTWfigsForSlides.py: 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:

• ConsumptionSaving/ConsumerParameters.py: 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.
• SolvingMicroDSOPs/SetupSCFdata.py: Imports 2004 SCF data for use by SolvingMicroDSOPs/StructEstimation.py.
• cstwMPC/SetupParamsCSTW.py: Loads calibrated model parameters for cstwMPC.py, chooses specification.
• FashionVictim/FashionVictimParams.py: Example parameters for FashionVictimModel.py, loaded when that module is run.

Test modules:

• Testing/ComparisonTests.py: 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.
• Testing/ModelTesting.py: Early version of unit testing for HARK, still in development. Defines a few wrapper classes to run unit tests on subclasses of AgentType.
• Testing/ModelTestingExample.py An example of ModelTesting.py in action, using TractableBufferStockModel.
• Testing/TBSunitTests.py: Early version of unit testing for HARK, still in development. Runs a test on TractableBufferStockModel.
• Testing/MultithreadDemo.py: 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.

Data files:

• SolvingMicroDSOPs/SCFdata.csv: SCF 2004 data for use in SolvingMicroDSOPs/StructEstimation.py, loaded by SolvingMicroDSOPs/EstimationParameters.py.
• cstwMPC/SCFwealthDataReduced.txt: SCF 2004 data with just net worth and data weights, for use by cstwMPC.py
• cstwMPC/USactuarial.txt: U.S. mortality data from the Social Security Administration, for use by cstwMPC.py when running a lifecycle specification.
• cstwMPC/EducMortAdj.txt: Mortality adjusters by education and age (columns by sex and race), for use by cstwMPC.py when running a lifecycle specification. Taken from an appendix of PAPER.

Other files that you don't need to worry about:

• */index.py: 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.
• .gitignore: 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: License text for HARK, Apache 2.0. Read it if you're a lawyer!
• SolvingMicroDSOPs/SMMcontour.png: Contour plot of the objective function for SolvingMicroDSOPs/StructEstimation.py. Generated when that module is run, along with a PDF version.
• cstwMPC/Figures/placeholder.txt: A placeholder file because git doesn't like empty folders, but cstwMPC.py needs the /Figures directory to exist when it runs.
• cstwMPC/Results/placeholder.txt: A placeholder file because git doesn't like empty folders, but cstwMPC.py needs the /Results directory to exist when it runs.
• Documentation/conf.py: A configuration file for producing html documentation with Sphinx, generated by sphinx-quickstart.
• Documentation/includeme.rst: A very small file used by Sphinx to produce documentation.
• Documentation/index.rst: 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.
• Documentation/instructions.md: A markdown file with instructions for how to set up and run Sphinx. You don't need to read it.
• Documentation/simple-steps-getting-sphinx-working.md: Another markdown file with instructions for how to set up and run Sphinx.
• Documentation/make.bat: A batch file for producing Sphinx documentation, generated by sphinx-quickstart.
• Documentation/Makefile: Another Sphinx auxiliary file generated by sphinx-quickstart.
• Documentation/econtex.sty: LaTeX style file with notation definitions.
• Documentation/econtex.cls: LaTeX class file with document layout for the user manual.
• Documentation/econtexSetup.sty: LaTeX style file with notation definitions.
• Documentation/econtexShortcuts.sty: LaTeX style file with notation definitions.
• Documentation/UserGuidePic.pdf: Image for the front cover of the user guide, showing the consumption function for the KinkyPref model.

IV. WARNINGS AND DISCLAIMERS

This is an early 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!

V. License

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