Release Notes#


This document contains the release notes of HARK. HARK aims to produce an open source repository of highly modular, easily interoperable code for solving, simulating, and estimating dynamic economic models with heterogeneous agents.

For more information on HARK, see our Github organization.


0.15.0 (in development)#

Release Date: TBA

Major Changes#

  • Replace object-oriented solvers with single function versions. 1394

  • Object-oriented solver code has been moved to /HARK/ConsumptionSaving/, for legacy support of downstream projects.

Minor Changes#

  • Add option to pass pre-built grid to LinearFast. 1388

  • Moves calculation of stable points out of ConsIndShock solver, into method called by post_solve #1349

  • Adds cubic spline interpolation and value function construction to “warm glow bequest” models.

  • Fixes cubic spline interpolation for ConsMedShockModel.

  • Moves computation of “stable points” from inside of ConsIndShock solver to a post-solution method. 1349

  • Corrects calculation of “human wealth” under risky returns, providing correct limiting linear consumption function. 1403


Release date: February 28, 2024

Major Changes#


Minor Changes#

  • Fixes a bug in make_figs arising from the metadata argument being incompatible with jpg. #1386

  • Reverts behavior of the repr method of the Model class, so that long strings aren’t generated. Full description is available with describe(). #1390


Release Date: February 12, 2024

Major Changes#

  • Adds HARK.core.AgentPopulation class to represent a population of agents with ex-ante heterogeneous parametrizations as distributions. #1237

  • Adds HARK.core.Parameters class to represent a collection of time varying and time invariant parameters in a model. #1240

  • Adds HARK.simulation.monte_carlo module for generic Monte Carlo simulation functions using Python model configurations. 1296

Minor Changes#

  • Adds option sim_common_Rrisky to control whether risky-asset models draw common or idiosyncratic returns in simulation. #1250,#1253

  • Addresses #1255. Makes age-varying stochastic returns possible and draws from their discretized version. #1262

  • Fixes bug in the metric that compares dictionaries with the same keys. #1260

  • Fixes bug in the calc_jacobian method. #1342

  • Fixes bug that prevented risky-asset consumer types from working with time-varying interest rates Rfree. 1343

  • Overhauls and expands condition checking for the ConsIndShock model #1294. Condition values and a description of their interpretation is stored in the bilt dictionary of IndShockConsumerType.

  • Creates a models/ directory with Python model configurations for perfect foresight and Fisher 2-period models. 1347

  • Fixes bug in AgentType simulations where ‘who_dies’ for period t was being recorded in period t-1 in the history Carlo simulation functions using Python model configurations.1296

  • Removes unused .1296

  • Fixes bug that default seed was being used in the initializing of income shock distributions. 1380


Release Date: February 16, 2023

Major Changes#

  • Updates the DCEGM tools to address the flaws identified in issue #1062. PR: 1100.

  • Updates IndexDstn, introducing the option to use an existing RNG instead of creating a new one, and creating and storing all the conditional distributions at initialization. 1104

  • make_shock_history and read_shocks == True now store and use the random draws that determine newborn’s initial states #1101.

  • FrameModel and FrameSet classes introduced for more modular construction of framed models. FrameAgentType dedicated to simulation. #1117

  • General control transitions based on decision rules in FrameAgentType. #1117

  • Adds distr_of_function tool to calculate the distribution of a function of a discrete random variable. #1144

  • Changes the DiscreteDistribution class to allow for arbitrary array-valued random variables. #1146

  • Adds IndShockRiskyAssetConsumerType as agent which can invest savings all in safe asset, all in risky asset, a fixed share in risky asset, or optimize its portfolio. #1107

  • Updates all HARK models to allow for age-varying interest rates. #1150

  • Adds DiscreteDistribution.expected method which expects vectorized functions and is faster than HARK.distribution.calc_expectation. #1156

  • Adds DiscreteDistributionXRA class which extends DiscreteDistribution to allow for underlying data to be stored in a xarray.DataArray object. #1156

  • Adds keyword argument labels to expected() when using DiscreteDistributionXRA to allow for expressive functions that use labeled xarrays. #1156

  • Adds a wrapper for for fast multilinear interpolation. #1151

  • Adds support for the calculation of dreivatives in the wrappers. #1157

  • Adds class DecayInterp to It implements interpolators that “decay” to some limiting function when extrapolating. #1165

  • Add methods to non stochastically simulate an economy by computing transition matrices. Functions to compute transition matrices and ergodic distribution have been added #1155.

  • Fixes a bug that causes t_age and t_cycle to get out of sync when reading pre-computed mortality. #1181

  • Adds Methods to calculate Heterogenous Agent Jacobian matrices. #1185

  • Enhances combine_indep_dstns to work with labeled distributions (DiscreteDistributionLabeled). #1191

  • Updates the numpy random generator from RandomState to Generator. #1193

  • Turns the income and income+return distributions into DiscreteDistributionLabeled objects. #1189

  • Creates UtilityFuncCRRA which is an object oriented utility function with a coefficient of constant relative risk aversion and includes derivatives and inverses. Also creates UtilityFuncCobbDouglas, UtilityFuncCobbDouglasCRRA, and UtilityFuncConstElastSubs. #1168

  • Reorganizes HARK.distribution. All distributions now inherit all features from scipy.stats. New ContinuousFrozenDistribution and DiscreteFrozenDistribution to use scipy.stats distributions not yet implemented in HARK. New Distribution.discretize(N, method = "***") replaces Distribution.approx(N). New DiscreteDistribution.limit attribute describes continuous origin and discretization method. #1197.

  • Creates new class of labeled models under ConsLabeledModel that use xarray for more expressive modeling of underlying mathematical and economics variables. #1177

Minor Changes#

  • Updates the lognormal-income-process constructor from to use IndexDistribution. #1024, #1115

  • Allows for age-varying unemployment probabilities and replacement incomes with the lognormal income process constructor. #1112

  • Option to have newborn IndShockConsumerType agents with a transitory income shock in the first period. Default is false, meaning they only have a permanent income shock in period 1 and permanent AND transitory in the following ones. #1126

  • Adds benchmark utility to profile the performance of HARK solvers. #1131

  • Fixes scaling bug in Normal equiprobable approximation method. 1139

  • Removes the extra-dimension that was returned by calc_expectations in some instances. #1149

  • Adds HARK.distribution.expected alias for DiscreteDistribution.expected. #1156

  • Renames attributes in DiscreteDistribution: X to atoms and pmf to pmv. #1164, #1051, #1159.

  • Remove or replace automated tests that depend on brittle simulation results. #1148

  • Updates asset grid constructor from to allow for linearly-spaced grids when aXtraNestFac == -1. #1172

  • Renames DiscreteDistributionXRA to DiscreteDistributionLabeled and updates methods #1170

  • Renames HARK.numba to HARK.numba_tools #1183

  • Adds the RNG seed as a property of DiscreteDistributionLabeled #1184

  • Updates the approx method of HARK.distributions.Uniform to include the endpoints of the distribution with infinitesimally small (zero) probability mass. #1180

  • Refactors tests to incorporate custom precision HARK_PRECISION = 4. #1193

  • Cast DiscreteDistribution.pmv attribute as a np.ndarray. #1199

  • Update structure of dynamic interest rate. #1221


Release Date: December 14, 2021

Major Changes#

  • FrameAgentType for modular definitions of agents #865 #1064

  • Frame relationships with backward and forward references, with plotting example #1071

  • PortfolioConsumerFrameType, a port of PortfolioConsumerType to use Frames #865

  • Input parameters for cyclical models now indexed by t #1039

  • A IndexDistribution class for representing time-indexed probability distributions #1018.

  • Adds new consumption-savings-portfolio model RiskyContrib, which represents an agent who can save in risky and risk-free assets but faces frictions to moving funds between them. To circumvent these frictions, he has access to an income-deduction scheme to accumulate risky assets. PR: #832. See this forthcoming REMARK for the model’s details.

  • ‘cycles’ agent property moved from constructor argument to parameter #1031

  • Uses iterated expectations to speed-up the solution of RiskyContrib when income and returns are independent #1058.

  • ConsPortfolioSolver class for solving portfolio choice model replaces solveConsPortfolio method #1047

  • ConsPortfolioDiscreteSolver class for solving portfolio choice model when allowed share is on a discrete grid #1047

  • ConsPortfolioJointDistSolver class for solving portfolio chioce model when the income and risky return shocks are not independent #1047

Minor Changes#

  • Using Lognormal.from_mean_std in the forward simulation of the RiskyAsset model #1019

  • Fix bug in DCEGM’s primary kink finder due to numpy no longer accepting NaN in integer arrays #990.

  • Add a general class for consumers who can save using a risky asset #1012.

  • Add Boolean attribute ‘PerfMITShk’ to consumption models. When true, allows perfect foresight MIT shocks to be simulated. #1013.

  • Track and update start-of-period (pre-income) risky and risk-free assets as states in the RiskyContrib model 1046.

  • distribute_params now uses assign_params to create consistent output #1044

  • The function that computes end-of-period derivatives of the value function was moved to the inside of ConsRiskyContrib’s solver #1057

  • Use np.fill(np.nan) to clear or initialize the arrays that store simulations. #1068

  • Add Boolean attribute ‘neutral_measure’ to consumption models. When true, simulations are more precise by allowing permanent shocks to be drawn from a neutral measure (see Harmenberg 2021). #1069

  • Fix mathematical limits of model example in example_ConsPortfolioModel.ipynb #1047

  • Update to use calc_expectation method #1072

  • Fix bug in calc_normal_style_pars_from_lognormal_pars due to math error. #1076

  • Fix bug in distribute_params so that AgentCount parameter is updated. #1089

  • Fix bug in ‘vFuncBool’ option for ‘MarkovConsumerType’ so that the value function may now be calculated. #1095


Release Date: March 4, 2021

Major Changes#

  • Converts non-mathematical code to PEP8 compliant form #953

  • Adds a constructor for LogNormal distributions from mean and standard deviation #891

  • Uses new LogNormal constructor in ConsPortfolioModel #891

  • calcExpectations method for taking the expectation of a distribution over a function [#884](] (#897)[

  • Implements the multivariate normal as a supported distribution, with a discretization method. See #948.

  • Centralizes the definition of value, marginal value, and marginal marginal value functions that use inverse-space interpolation for problems with CRRA utility. See #888.

  • MarkovProcess class used in ConsMarkovModel, ConsRepAgentModel, ConsAggShockModel #902 #929

  • replace HARKobject base class with MetricObject and Model classes #903

  • Add repr and eq methods to Model class #903

  • Adds SSA life tables and methods to extract survival probabilities from them #986.

  • Adds the U.S. CPI research series and tools to extract inflation adjustments from it #930.

  • Adds a module for extracting initial distributions of permanent income (pLvl) and normalized assets (aNrm) from the SCF #932.

  • Fix the return fields of dcegm/calcCrossPoints#909.

  • Corrects location of constructor documentation to class string for Sphinx rendering #908

  • Adds a module with tools for parsing and using various income calibrations from the literature. It includes the option of using life-cycle profiles of income shock variances from Sabelhaus and Song (2010). See #921, #941, #980.

  • remove “Now” from model variable names #936

  • remove; use Model init in Market and AgentType init to standardize on parameters dictionary #947

  • Moves state MrkvNow to shocks[‘Mrkv’] in AggShockMarkov and KrusellSmith models #935

  • Replaces ConsIndShock’s init_lifecycle with an actual life-cycle calibration #951.

Minor Changes#

  • Move AgentType constructor parameters docs to class docstring so it is rendered by Sphinx.

  • Remove uses of deprecated time.clock #887

  • Change internal representation of parameters to Distributions to ndarray type

  • Rename IncomeDstn to IncShkDstn

  • AgentType simulate() method now returns history. #916

  • Rename DiscreteDistribution.drawDiscrete() to draw()

  • Update documentation and warnings around IncShkDstn #955

  • Adds csv files to 957


Release Date: Nov. 05 2020

Major Changes#

  • Namespace variables for the Market class #765

  • We now have a Numba based implementation of PerfForesightConsumerType model available as PerfForesightConsumerTypeFast #774

  • Namespace for exogenous shocks #803

  • Namespace for controls #855

  • State and poststate attributes replaced with state_now and state_prev namespaces #836

Minor Changes#

  • Use shock_history namespace for pre-evaluated shock history #812

  • Fixes seed of PrefShkDstn on initialization and add tests for simulation output

  • Reformat code style using black


Release Date: 08-08-2020

Major Changes#

  • Add a custom KrusellSmith Model #762

  • Simulations now uses a dictionary history to store state history instead of _hist attributes #674

  • Removed time flipping and time flow state, “forward/backward time” through data access #570

  • Simulation draw methods are now individual distributions like Uniform, Lognormal, Weibull #624

Minor Changes#

  • unpackcFunc is deprecated, use unpack(parameter) to unpack a parameter after solving the model #784

  • Remove deprecated Solution Class, use HARKObject across the codebase #772

  • Add option to find crossing points in the envelope step of DCEGM algorithm #758

  • Fix reset bug in the behaviour of AgentType.resetRNG(), implemented individual resetRNG methods for AgentTypes #757

  • Seeds are set at initialisation of a distribution object rather than draw method #691 #750, #729

  • Deal with portfolio share of ‘bad’ assets #749

  • Fix bug in make_figs utilities function #755

  • Fix typo bug in Perfect Foresight Model solver #743

  • Add initial support for logging in ConsIndShockModel #714

  • Speed up simulation in AggShockMarkovConsumerType #702

  • Fix logic bug in DiscreteDistribution draw method #715

  • Implemented distributeParams to distributes heterogeneous values of one parameter to a set of agents #692

  • NelderMead is now part of estimation #693

  • Fix typo bug in parallel #682

  • Fix DiscreteDstn to make it work with multivariate distributions #646

  • BayerLuetticke removed from HARK, is now a REMARK #603

  • cstwMPC removed from HARK, is now a REMARK #666

  • SolvingMicroDSOPs removed from HARK, is now a REMARK #651

  • constructLogNormalIncomeProcess is now a method of IndShockConsumerType #661

  • Discretize continuous distributions #657

  • Data used in cstwMPC is now in HARK.datasets #622

  • Refactor checkConditions by adding a checkCondition method instead of writing custom checks for each condition #568

  • Examples update #768, #759, #756, #727, #698, #697, #561, #654, #633, #775


Release Date: 17-04-2020

Major Changes#

  • Add Bellman equations for cyclical model example #600

  • read_shocks now reads mortality as well #613

  • Discrete probability distributions are now classes #610

Minor Changes#


Release Date: 24-03-2020

Major Changes#

  • Default parameters dictionaries for ConsumptionSaving models have been moved from ConsumerParameters to nearby the classes that use them. #527

  • Improvements and cleanup of ConsPortfolioModel, and adding the ability to specify an age-varying list of RiskyAvg and RiskyStd. #577

  • Rewrite and simplification of ConsPortfolioModel solver. #594

Minor Changes#


Release Date: 05-03-2020

Major Changes#

  • Last release to support Python 2.7, future releases of econ-ark will support Python 3.6+ #478

  • Move non-reusable model code to examples directory, BayerLuetticke, FashionVictim now in examples instead of in HARK code #442

  • Load default parameters for ConsumptionSaving models #466

  • Improved implementaion of parallelNelderMead #300

Minor Changes#

  • Notebook utility functions for determining platform, GUI, latex (installation) are available in HARK.utilities #512

  • Few DemARKs moved to examples #472

  • MaxKinks available in ConsumerParameters again #486


Release Date: 12-12-2019

Major Changes#

  • Added constrained perfect foresight model solution. (#299

Minor Changes#

  • Fixed slicing error in minimizeNelderMead. (#460)

  • Fixed matplotlib GUI error. (#444)

  • Pinned sphinx dependency. (#436)

  • Fixed bug in ConsPortfolioModel in which the same risky rate of return would be drawn over and over. (#433)

  • Fixed sphinx dependency errors. (#411)

  • Refactored (#408)

  • AgentType.simulate() now throws informative errors if attributes required for simulation do not exist, or initializeSim() has never been called. (#320)


Release Date: 10-03-2019

Minor Changes#

  • Add some bugfixes and unit tests to HARK.core. (#401)

  • Fix error in discrete portfolio choice’s AdjustPrb. (#391)


Release Date: 09-25-2019

Minor Changes#

  • Added portfolio choice between risky and safe assets (ConsPortfolioModel). (#241)


Release Date: 09-19-2019

Minor Changes#

  • Fixes cubic interpolation in KinkedRSolver. (#386)

  • Documentes the procedure for constructing value function inverses and fixes bug in which survival rate was not included in absolute patience factor. (#383)

  • Fixes problems that sometimes prevented multiprocessing from working. (#377)


Release Date: 07-23-2019

Minor Changes#

  • Missed pre-solve fix (see #363 for more context). (#367)


Release Date: 07-22-2019

Minor Changes#

  • Revert pre-solve commit due to bug. (#363)


Release Date: 07-20-2019

Breaking Changes#

  • See #302 under minor changes.

Major Changes#

  • Adds BayerLuetticke notebooks and functionality. (#328)

Minor Changes#

  • Fixes one-asset HANK models for endowment economy (had MP wired in as the shock). (#355)

  • Removes jupytext *.py files. (#354)

  • Reorganizes documentation and configures it to work with Read the Docs. (#353)

  • Adds notebook illustrating dimensionality reduction in Bayer and Luetticke. (#345)

  • Adds notebook illustrating how the Bayer & Luetticke invoke the discrete cosine transformation(DCT) and fixed copula to reduce dimensions of the problem.(#344)

  • Makes BayerLuetticke HANK tools importable as a module. (#342)

  • Restores functionality of SGU_solver. (#341)

  • Fixes datafile packaging issue. (#332)

  • Deletes .py file from Bayer-Luetticke folder. (#329)

  • Add an empty method for preSolve called checkRestrictions that can be overwritten in classes inheriting from AgentType to check for illegal parameter values. (#324)

  • Adds a call to updateIncomeProcess() in preSolve() to avoid solutions being based on wrong income process specifications if some parameters change between two solve() calls. (#323)

  • Makes checkConditions() less verbose when the checks are not actually performed by converting a print statement to an inline comment. (#321)

  • Raises more readable exception when simultate() is called without solving first. (#315)

  • Removes testing folder (part of ongoing test restructuring). (#304)

  • Fixes unintended behavior in default simDeath(). Previously, all agents would die off in the first period, but they were meant to always survive. (#302) Warning: Potentially breaking change.


Release Date: 05-30-2019

No changes from 0.10.0.dev3.


Release Date: 05-18-2019

Major Changes#

  • Fixes multithreading problems by using Parallels(backend=’multiprocessing’). (287)

  • Fixes bug caused by misapplication of check_conditions. (284)

  • Adds functions to calculate quadrature nodes and weights for numerically evaluating expectations in the presence of (log-)normally distributed random variables. (258)

Minor Changes#

  • Adds method decorator which validates that arguments passed in are not empty. (282

  • Lints a variety of files. These PRs include some additional/related minor changes, like replacing an exec function, removing some lambdas, adding some files to .gitignore, etc. (274, 276, 277, 278, 281)

  • Adds vim swp files to gitignore. (269)

  • Adds version dunder in init. (265)

  • Adds flake8 to requirements.txt and config. (261)

  • Adds some unit tests for IndShockConsumerType. (256)


Release Date: 04-18-2019

Major Changes#


Minor Changes#

  • Fix verbosity check in ConsIndShockModel. (250)

Other Changes#



Release Date: 04-12-2019

Major Changes#

  • Adds tools to solve problems that arise from the interaction of discrete and continuous variables, using the DCEGM method of Iskhakov et al., who apply the their discrete-continuous solution algorithm to the problem of optimal endogenous retirement; their results are replicated using our new tool here. (226)

  • Parameters of ConsAggShockModel.CobbDouglasEconomy.updateAFunc and ConsAggShockModel.CobbDouglasMarkovEconomy.updateAFunc that govern damping and the number of discarded ‘burn-in’ periods were previously hardcoded, now proper instance-level parameters. (244)

  • Improve accuracy and performance of functions for evaluating the integrated value function and conditional choice probabilities for models with extreme value type I taste shocks. (242)

  • Add calcLogSum, calcChoiceProbs, calcLogSumChoiceProbs to HARK.interpolation. (209, 217)

  • Create tool to produce an example “template” of a REMARK based on SolvingMicroDSOPs. (176)

Minor Changes#

  • Moved old utilities tests. (245)

  • Deleted old files related to “cstwMPCold”. (239)

  • Set numpy floating point error level to ignore. (238)

  • Fixed miscellaneous imports. (212, 224, 225)

  • Improve the tests of buffer stock model impatience conditions in IndShockConsumerType. (219)

  • Add basic support for Travis continuous integration testing. (208)

  • Add SciPy to requirements.txt. (207)

  • Fix indexing bug in bilinear interpolation. (194)

  • Update the build process to handle Python 2 and 3 compatibility. (172)

  • Add MPCnow attribute to ConsGenIncProcessModel. (170)

  • All standalone demo files have been removed. The content that was in these files can now be found in similarly named Jupyter notebooks in the DEMARK repository. Some of these notebooks are also linked from (229, 243)

Other Notes#

  • Not all changes from 0.9.1 may be listed in these release notes. If you are having trouble addressing a breaking change, please reach out to us.