Release Notes#

Introduction#

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.

Changes#

0.17.3 (dev)#

Release Date: TBD

Release Notes#

(None yet)

Major Changes#

  • Raises the minimum supported Python to 3.12 and adds 3.14, following SPEC 0, which drops a Python version three years after release. Python 3.11 left that window in October 2025, and numpy and scipy both already require 3.12 or newer, so installing HARK on 3.10 or 3.11 resolved a dependency stack from two years ago rather than the one HARK is developed against. The CI matrix now sweeps 3.12, 3.13 and 3.14 on Linux and covers 3.13 on macOS and Windows.

  • Breaking: ConsAggIndMarkovModel is rewritten: AggIndMrkvConsumerType(MarkovConsumerType) replaces the former AggIndMarkovConsumerType(AgentType) (hierarchical macro+micro Markov states via shocks["Mrkv"], overridable get_macro_markov_states/get_micro_markov_states, pure-MarkovConsumerType fallback when the hierarchical counts are unset). The old name is removed, not aliased - the old class’s contract (MrkvCombined, AgentType base) differed enough that a silent alias would mislead; imports fail loudly instead. KrusellSmithType reparents onto the new class (with a no-op sim_death preserving its no-mortality RNG stream); its default simulation path is unchanged, evidenced by the untouched seeded KS test suite. #1798

  • Breaking (Krusell-Smith only): the aggregate Markov state key on KrusellSmithType and KrusellSmithEconomy is renamed "Mrkv" → "MrkvAgg" (shock_vars_, shocks, sow_vars, track_vars, sow_init, and economy.history). This frees the "Mrkv" name for the idiosyncratic Markov state in the hierarchical-Markov refactor and removes the collision between agent-level and aggregate-level state keys. AggShockMarkovConsumerType and the Cobb-Douglas Markov economies are unchanged. Migration: replace history["Mrkv"]/shocks["Mrkv"] with "MrkvAgg" in KS-based code. #1797

  • future item

  • future item

  • future item

Minor Changes#

  • Fixes HARK.dual_measure indexing the Q income process one period ahead of P whenever cycles != 1: _draw_Q_shocks_indshock chose IncShkDstn_Q[t] where get_shocks uses IncShkDstn[t - 1], so an infinite-horizon agent with T_cycle > 1 drew the Q sample from the wrong period’s distribution and scaled it by the wrong PermGroFac. Invisible until now because every fixture used T_cycle == 1, where indices 0 and -1 name the same element.

  • Fixes MarkovProcess.draw(shuffle=True) returning uninitialized memory for an agent whose source state has no row in the transition matrix. The output buffer is now sentinel-filled and verified, so those agents raise IndexError (as the unshuffled path already did) instead of silently inheriting the previous period’s Mrkv values.

  • Fixes a division by zero at LivPrb == 1 in compute_mean_pLvl, and a wrong limit in the corresponding guard in compute_pLvl_factor. Both compute the newborn share of a stationary population; it is now one shared helper returning 1 / T_age at the no-mortality limit rather than nan or 0.

  • MarkovConsumerType.get_shocks now records newborn base draws under ("newborn", j), the key HARK.dual_measure already looked up and nothing wrote, so Markov newborns share their uniforms with the Q measure like every other cell. The P stream is unchanged.

  • setup_Q_measure refuses, rather than silently mispricing, a composition whose get_Rport reads P-side state (KinkedRconsumerType, KinkyPrefConsumerType, and the ConsRiskyAssetModel branch, which would give the Q agent the P agent’s realized portfolio share).

  • setup_Q_measure no longer emits one degenerate-distribution warning per period; a stock init_lifecycle produced 25, which buried the aggregate warning that matters. The periods are recorded in Q_degenerate_periods instead.

  • MarkovConsumerType.get_markov_states warns and falls back when balanced_transitions is set without a pLvl to sort on, matching AggIndMrkvConsumerType instead of raising a bare KeyError; both now share one helper.

  • DualMeasureMixin.sim_one_period calls AgentType._sim_period_prologue/_sim_period_epilogue instead of hand-copying them, so a future change to the prologue cannot silently skip dual mode. Verified bit-identical.

  • Declares init_shuffle in PerfForesightConsumerType_simulation_defaults and init_indshk_markov, where its readers live; it was declared only on IndShockConsumerType and worked through a getattr fallback.

  • Restores ruff’s default file discovery (extend-include rather than include), so ruff check <dir> no longer reports “All checks passed” after inspecting zero Python files.

  • AgentType.get_states now raises when transition() returns more values than there are states. States are assigned by position, so the loop silently dropped the tail. A return shorter than the state list stays legal, because it is deliberate: GenIncProcessConsumerType declares five states and returns three, writing the rest by name in get_poststates. The comment there now also records that reordering state_vars silently reassigns every value, which is how three models came to declare an aNrm that nothing wrote.

  • MarkovProcess.draw(shuffle=True) now warns when a source state has too few agents for deterministic counts and falls back to iid. The exact transition counts the shuffled path advertises were silently withdrawn for those agents while the rest of the population kept them; both normalization mixins already warn on their analogous skips. Aggregated into one warning naming the affected source states, rather than one per state per period.

  • AgentType._sim_period_prologue now blanks each period’s ndarray states with nan instead of np.empty. A state that no later step writes previously held whatever was in the freed buffer, which in practice is usually the previous period’s values, so the gap read as plausible data rather than as a defect; it now surfaces as nan rather than as plausible numbers. The same change in AgentSimulator’s newborn path (HARK.simulator) replaces an np.empty under a comment promising to “clear” the variable; both now route through one type-dispatching blank helper, which was already nan-filling elsewhere in that file. #1809

  • Fixes three models reporting an uninitialized aNrm: GenIncProcessConsumerType, MedShockConsumerType, and MedExtMargConsumerType. The variable is declared in state_vars but these models work in levels, so nothing in transition writes it and only sim_birth ever touched it; every continuing agent carried whatever the per-period blanking left behind. On a 100-agent, 10-period GenIncProcess run, 998 of 1000 tracked cells disagreed with aLvl / pLvl, taking values like 3.96e-319 – subnormals, i.e. freed memory. MedShockConsumerType was worse: all 1600 cells of a 200-agent, 8-period run. The definition now lives in one place, GenIncProcessConsumerType.set_aNrm_from_levels, which the two get_poststates overrides call, so a further override cannot silently reopen it. Found because the nan blanking above made the second and third instances visible.

  • Removes a duplicate "mLvl" from MedShockConsumerType.state_vars, which appended a name the inherited list already contained.

  • setup_Q_measure now also refuses an agent with normalize_pLvl=True, and PermanentIncomeNormalizationMixin refuses an agent already in dual mode. Same class of defect as normalize_shocks below by a different mechanism: the per-cohort pLvl adjustment runs in post_state_hook and the Q pipeline does not mirror it at all. Measured at 1000 agents over 10 periods, enabling it moves the P history by 1.45e-2 and leaves the Q history bit-identical; across 10 seeds at 2000 agents the standard deviation of the final-period mean pLvl falls from 9.71e-3 to 5.91e-4 for P while Q stays at 1.03e-2. Both refusal messages now say the other flag is not a workaround, since the previous one-sided guard told users to turn off normalize_shocks and the module’s own docstring example enables both.

  • setup_Q_measure now refuses an agent with normalize_shocks=True, and ShockNormalizationMixin refuses an agent already in dual mode. Composing the two variance-reduction features reversed the result dual mode exists to demonstrate: normalization rescales shocks["PermShk"] in place after the base uniforms were recorded and before the Q pipeline inverts them, so P’s cross-sectional shock mean was pinned exactly while Q kept all of its sampling noise. Measured over 12 seeds at 2000 agents, the cross-seed standard deviation of the period-mean deviation is 2.03e-3 for both measures with normalization off, and 6.1e-17 for P against 2.03e-3 for Q with it on – so a user enabling both, which nothing warned against, would have concluded the neutral measure increases variance. Making them genuinely composable requires normalizing the Q draws to the Q measure’s own mean, PermGroFac * E[psi^2] / E[psi]^2, which is a design decision rather than a fix, so this refuses instead of guessing.

  • ShockNormalizationMixin’s zero-mean guard is now relative to the shock scale rather than an absolute 1e-16. The guard protects a division by the empirical mean, so what matters is the mean’s size next to the values it came from: the absolute form skipped a group whose shocks were legitimately all near 1e-18, and accepted a mean of 1e-10 among values of order 1e6, where the rescale factor is order 1e16. No shipped calibration reaches either end; the simulation fingerprint is unchanged.

  • Removes the unreachable first branch of AggIndMrkvConsumerType.get_macro_markov_states. Its docstring advertised self.EconomyMrkvNow as the primary lookup, but nothing in HARK assigns that attribute, so the hasattr guard always fell through to shocks["MrkvAgg"].

  • Renames the examples workflow from “Test examples as a cron job” to “Test examples”, which is what it does: besides the nightly run it also fires on every push to main and every PR against main. Those pre-merge runs are the ones that catch a change breaking an example notebook, so the name was corrected rather than the triggers.

  • Declares the public API of the recently added modules: __all__ for HARK.simulation.normalization and HARK.ConsumptionSaving.ConsAggIndMarkovModel, and an API-reference section for HARK.simulation.normalization on the Simulation tools page. #1811

  • Excludes scipy 1.18.0, whose PPoly-family objects (e.g. CubicHermiteSpline) cannot be deepcopy-ed (TypeError: cannot pickle 'module' object), breaking ValueFuncCRRA construction and the existing test suite wherever that scipy version is resolved. #1788

  • Adds opt-in markov_shuffle and balanced_transitions parameters to MarkovConsumerType.get_markov_states: quota-exact Markov transitions via MarkovProcess.draw(shuffle=True), optionally with systematic sampling by pLvl. Default False; the default call is unchanged. #1793

  • Adds opt-in low-variance draw modes to the distributions layer: MarkovProcess.draw(shuffle=, sort_key=, draws=) (quota-exact state transitions with optional stratified rank assignment), DiscreteDistribution.draw_events(shuffle=), and DiscreteDistribution.draw(replicates=) (exact full-coverage samples). The leftover-slot allocation that all of these share now lives in one function, HARK.distributions.base.allocate_remainder_slots, rather than being written out once per call site: the two earlier copies diverged, with the MarkovProcess one still allocating leftovers proportional to the transition row instead of to the fractional remainders, which overweighted the modal target by up to 6.5% and starved the rarest by 15% whenever N_j * P[j,k] was not an integer. replicates now rejects non-positive values instead of silently returning an empty sample, accepts zero-probability atoms, and warns only when the minimal sample is larger than the rarest atom alone requires. sort_key=/draws= warn when passed with shuffle=False rather than being silently ignored. Every default path is textually identical to the previous code and pinned by RNG-stream golden tests. #1786

  • Adds HARK.simulation.normalization: opt-in mixins that pin simulated cross-sectional moments to their analytical values, removing sampling noise from aggregates without requiring special population sizes. ShockNormalizationMixin rescales each period’s drawn shocks so their cross-sectional means are exact; because HARK stores psi * PermGroFac in shocks["PermShk"] (growth is folded into the permanent “shock”), the target for PermShk is PermGroFac, not 1.0 - normalizing it to 1.0 would delete permanent income growth rather than sampling noise. PermanentIncomeNormalizationMixin pins per-cohort log-pLvl moments; a cohort reading t_age == k inside post_state_hook has already taken k + 1 permanent shocks (newborns are not exempt: get_shocks redraws a random PermShk for them and pins only TranShk), and the targets accumulate period by period over each cohort’s realized income-process history, so life-cycle calibrations with age-varying PermGroFac or PermShkStd get the right age profile instead of period 0’s parameters extrapolated. Both are Markov-capable, with per-state targets and an automatic mean-only mode under state-dependent growth, and both warn rather than degrade silently when they cannot deliver exactness (small groups or cohorts, read_shocks replay, staggered entry, a model whose sim_one_period never reaches post_state_hook). Wiring is through AgentType.post_state_hook rather than an overridden sim_one_period, so models with their own simulation pipeline are not shadowed. Purely additive: defaults change no behavior and no existing file is modified. #1784

  • make_hierarchical_mrkv_array auto-detects a general nested [i][j] conditional-matrix format (source-and-destination conditioning, Krusell-Smith style) alongside the existing flat destination-conditioned format (unchanged for existing callers); adds its inverse extract_cond_mrkv_arrays; KrusellSmithEconomy.make_MrkvArray now also stores MacroMrkvArray/CondMrkvArrays and the KS agent’s market_vars distributes them (additive plumbing for the hierarchical refactor). extract_cond_mrkv_arrays validates its input: it raises ValueError if MrkvIndArray is not (M*N) x (M*N), or if any block is not the macro probability times a row-stochastic matrix, which is the necessary and sufficient condition for the extracted arrays to be transition matrices. #1796

  • Adds an opt-in init_shuffle parameter (PerfForesightConsumerType.sim_birth, mirrored for MarkovConsumerType’s initial Markov states): exact-marginal initial-state draws via DiscreteDistribution.draw(shuffle=True), removing sampling noise in the initial cross-section. Default False; the kwarg is passed only when enabled, so duck-typed continuous init distributions keep working; pinned by a behavior-golden test. #1791

  • Makes CubicHermiteInterp deepcopy-able and picklable independent of scipy internals: the wrapped scipy spline is excluded from serialized state and deterministically rebuilt on restore, so attribute caching like scipy 1.18.0’s unpicklable module objects (scipy issue #25489) can no longer break serialization of HARK solutions. #1802

  • Removes the !=1.18.0 scipy exclusion added in #1788, which was a stopgap for the deepcopy failure the entry above fixes at its root. Note that pickles written by this version cannot be loaded by earlier HARK, since _chs is no longer stored in serialized state. #1802

  • Restores calc_expectation as a DeprecationWarning-bearing alias of expected_with_loop (renamed in 0.17.2), preserving import compatibility for downstream code pinned to earlier versions - including frozen reproduction archives that cannot be edited. Slated for removal in a future release. #1800

  • Exports KrusellSmithType, KrusellSmithEconomy, init_KS_agents, init_KS_economy in HARK.ConsumptionSaving.ConsAggShockModel.__all__ (they were defined but unlisted); fixes a stale sentence in the KrusellSmithType example notebook. #1795

  • Adds AgentType.post_state_hook(): a no-op extension point invoked by sim_one_period between get_states() and get_controls(), for mixins that adjust states before controls are computed (e.g. variance-reduction normalization). Default behavior is bit-identical (pinned by a behavior-golden test). #1787

  • Adds an opt-in death_shuffle parameter (PerfForesightConsumerType.sim_death, mirrored in MarkovConsumerType): for each distinct death probability, the number of deaths is set by floor-plus-remainder and the agents who die are drawn uniformly at random from that group. Each agent’s marginal death probability is still DiePrb and the expected number of deaths is unchanged; deaths within a group become negatively correlated (-1/(N-1)), which is what removes the binomial noise. The reduction scales with N_group * DiePrb, so it is exact only for a single large group with T_age=None (death count constant at every period) and partial otherwise: on a 65-age lifecycle calibration the variance falls 34% at AgentCount=1000 and 96% at AgentCount=20000, and with T_age set, old-age deaths are added afterward and are not de-noised. Default False; the default RNG path is preserved verbatim and pinned by a behavior-golden test. #1790

  • Fixes DiscreteDistribution.draw(shuffle=True), which distributed the leftover slots after floor(N*pmv) in proportion to pmv itself rather than to the fractional remainders, so E[count_j] was not N*pmv[j]. The remainder is now allocated by systematic sampling, whose inclusion probability is exactly the remainder. The error was O(M/N): invisible at population-sized draws, and large for the small cohorts sim_birth redraws, where the rarest atom came out about half as often as it should. draw(0, shuffle=True) also raised rather than returning an empty array, which sim_birth hits in any period with no deaths. Because ConsAggShockModel draws shuffled income shocks, the simulated Krusell-Smith economy changes slightly and AFunc[0].slope in testAggShockMarkovConsumerType moves from 1.05654 to 1.06030.

  • Adds HARK.dual_measure: opt-in Harmenberg neutral-measure (Q) parallel tracking via DualMeasureMixin, plus standalone aggregation helpers (compute_mean_pLvl, compute_pLvl_factor). Purely additive: no existing class or default behavior changes, and the default-off path is bit-identical to the plain agent including RNG stream position. simulate() delegates each period’s P-measure recording to AgentType.simulate instead of reimplementing its loop, so enabling dual mode cannot change what the P pipeline records; an earlier draft dropped the base loop’s getattr fall-through and left history["MPCnow"] all-NaN whenever dual_measure was on. _transition_Q stores kNrm and bNrm rather than computing bNrm and discarding it, and _lag_Q_states blanks with np.nan rather than np.empty, so a Q state nobody writes reads back as NaN instead of as recycled buffer contents that are finite and in range for the variable they are standing in for. Markov newborns redraw their Q permanent shock from IncShkDstn_Q[0][j] and gate TranShk on NewbornTransShk, matching MarkovConsumerType.get_shocks rather than pinning psi to 1. make_Q_measure_dstn and setup_Q_measure warn when the permanent shock is degenerate, instead of quietly handing back a Q measure equal to P. #1783

  • AgentType.make_shock_history gains an opt-in shuffle= keyword (default False): the pre-drawn shock history can be generated with the low-variance draw modes temporarily enabled. The original body survives verbatim as _make_shock_history; the default path delegates to it unchanged and is pinned by a stream-golden test. #1794

  • AggIndMrkvConsumerType.get_micro_markov_states gains the opt-in markov_shuffle branch: quota-exact micro-state transitions per (macro, source-micro) cell via MarkovProcess.draw(shuffle=True), supporting both conditional-matrix formats, with balanced_transitions (systematic sampling by pLvl) available. Default remains iid RNG.choice, unchanged. A macro transition that carries agents but has zero probability under MacroMrkvArray now raises ValueError naming the (macro_prev, macro_next) cell, instead of leaving those agents with uninitialised micro states; balanced_transitions=True without a pLvl in state_now now warns rather than silently falling back to unbalanced shuffling. #1801

  • Extends income_shuffle to MarkovConsumerType.get_shocks (per-state and newborn draws). Default False; original RNG paths preserved verbatim. #1792

  • Adds an opt-in income_shuffle parameter to IndShockConsumerType.get_shocks: exact floor-plus-leftover shock frequencies per period (via DiscreteDistribution.draw(shuffle=True)) instead of iid sampling. Default False; the default RNG path is preserved verbatim and pinned by a stream-golden test. #1789

  • Wires HARK.dual_measure’s base-draw cache through IndShockConsumerType.get_shocks and MarkovConsumerType.get_shocks: under the _cache_base_shock_draws flag, income-shock uniforms are recorded for Q-CDF inversion using the same draws and the same inversion as the default path; the P-stream is bit-identical with the flag on or off (tested). setup_Q_measure now turns the flag on and registers IncShkDstn_Q with self.distributions, so the shared base draws the module documents actually happen and survive repeated initialize_sim() calls; pass _cache_base_shock_draws = False afterwards for independent Q draws. Setting income_shuffle and the cache together records no uniforms, and now warns instead of silently decoupling P from Q (this covers the income_shuffle collision specifically, not every way P and Q can decouple; see the normalize_shocks entry below for the other one). initialize_sim clears _base_shock_draws so a later run cannot consume a previous run’s uniforms. #1799

  • future item

  • future item

  • future item

0.17.2#

Release Date: May 1, 2026

Release Notes#

This is a moderately sized release with several exciting new features, as well as many small improvements and fixes. Most of the breaking changes (see below) are very small adjustments to parameter names or formats; two functions also had their name change. The only significant breaking change is a reworking of the interaction between AgentType instances and their associated Market with respect to aggregate-level parameters.

The new features are headlined by the addition of two models with consumption habits in the new ConsHabitModel module. Additionally, HARK’s automatic HA-SSJ construction method has been extended to life-cycle models, rather than only infinite horizon models.

There are some breaking changes:

  • AgentType subclasses that had a get_economy_data method now use the general AgentType.get_market_params method, which exactly replicates their prior operation. See #1719

  • As a consequence of the above, random seeds on the distributions of some AgentType subclasses will change because the order in which they are created during instantiation has changed.

  • Parameter PortfolioBool has been deprecated. To allow portfolio choice for RiskyAssetConsumerType, just set RiskyShareFixed=None. #1740

  • The parameter BeqCRRA has been deprecated; agents with a warm glow bequest motive must use the same CRRA as their ordinary utility function. #1758

  • “Terminal bequest parameters” have been deprecated; agents have the same bequest motive in period T as they do in all other periods. #1758

  • calc_expectation has been renamed to expected_with_loop; use expected and pass vectorized=False for this functionality. #1763

  • The argument dist in expected has been renamed to dstn. #1763

  • The function make_exponential_grid has been renamed to make_polynomial_grid to reduce confusion with make_grid_exp_mult. #1762

Major Changes#

  • The new way to set up AgentType instances with an associated Market is to create them (with the agents in the Market’s agents attribute), then invoke the Market’s new give_agent_params() method. #1719

  • The above method calls each agent’s get_market_params() method, which references the market_vars class attribute for the names of objects to take from the associated Market.

  • All interpolator classes now have default derivative methods using finite differences. These are fallback methods, and are already overridden by most subclasses. #1723

  • New consumption-saving model with habit formation has been added; extends IndShockConsumerType model. #1739

  • Added habit-formation model with portfolio allocation, along with example notebooks. #1748

  • Simulator class has new method simulate_shock_by_grids to perturb the steady state distribution and then simulate by matrix transition methods. #1754

  • Simplify parameters in ConsBequestModel.py to eliminate “terminal” bequest parameters and different CRRA for bequests than consumption. #1758

  • The make_basic_SSJ method can now handle life-cycle models (cycles=1) as well as standard infinite horizon models. #1718

Minor Changes#

  • The special constructor get_it_from can now interpret the referenced attribute being a single value (any numeric or string) and will simply copy it to the new name. #1719

  • A Market’s calc_dynamics function/method can now use arguments other than those named in track_vars; HARK will look for those names as attributes of the Market. #1719

  • The _derY method for LowerEnvelope2D and LowerEnvelope3D were previously bugged and returned nonsense, now fixed. #1723

  • Updated syntax in a few places that tried to convert singleton array to a float, to ensure compatibility with NumPy 2.4+ #1725

  • Add new income shock constructor that incorporates Velasquez-Giraldo’s representation of medical expenses as negative transitory income shocks. #1724

  • Add parameter dictionary with Fulford and Low’s estimates for all expenses (not just medical) for use by MedShockConsumerType. #1724

  • Refactoring of representative agent model solver and the “labeled” submodule. #1727

  • Example notebooks for all models with portfolio choice have been significantly expanded and improved. #1740

  • Example notebooks for models in ConsAggShockModel.py have been improved and expanded from their prior form. #1738

  • The labels argument now works as intended with distribution.expected. #1742

  • Example notebook Transition_Matrix_Example.ipynb has been cleaned up and expanded. #1744

  • AggIndMarkovConsumerType added for models with both aggregate (shared) and idiosyncratic discrete states; KrusellSmithType refactored to extend it. #1747

  • Light safety fixes to the new HabitConsumerType. #1753

  • Example notebooks for models in ConsBequestModel.py have been improved and expanded from their prior form. #1754

  • Example notebooks for KinkedRconsumerType, MarkovConsumerType, LaborIntMargConsumerType, and TractableBufferStockConsumerType have been improved and expanded. #1743

  • Handling of income shocks for model “newborns” has been made consistent across models, with transitory shocks optional. #1760

  • Tests added to handle a variety of unusual corner cases. #1761

  • Computing expectations now always uses expected; if the function cannot accept vector arguments, pass vectorized=False. #1763

  • Matrix transition methods (including HA-SSJ) now support multi-exponential grids, as well as fully custom grids. #1762

  • HARK.interpolation refactored to reduce repetition and code clutter. #1765

  • Small documentation notebook for life-cycle HA-SSJ construction has been added. #1718

  • HARK.simulator and experimental Monte Carlo submodule refactored to reduce repetition. #1766

  • HARK.distributions refactored to reduce repetition and improve structures. #1767

  • Additional refactoring in Labeled, SSJutils, utilities, and metric to reduce code repetition. #1768

0.17.1#

Release Date: February 2, 2026

Release Notes#

This is a relatively small release that includes various adjustments and improvements (see Minor Changes), as well as several new features and an algebraic revision to some models (Major Changes).

There are some breaking changes:

  • The exact_match option for DiscreteDistribution.draw has been renamed to shuffle, and its behavior has changed slightly. See #1691.

  • Both AgentType subclasses in ConsPrefShockModel have had their utility function adjusted, moving the preference shock inside the CRRA term. See #1708.

  • If calc_expectation is used with a DiscreteDistributionLabeled, the function must reference indices of the distribution by name, not position number. See #1713.

  • Method NewKeynesianConsumerType.compute_steady_state has been renamed to compute_pe_steady_state. See #1711.

Major Changes#

  • Added find_target method to AgentType, automating search for target value of state variables. #1698

  • Utility function for PrefShockConsumerType and KinkyPrefConsumerType was algebraically rearranged. There is no functional difference, but the scale of preference shocks that yields a given level of consumption variation will be different. #1708

  • The format of the utility function for MedShockConsumerType has been revised; prior distributions of MedShk will need to be adjusted. See #1706.

  • The policy function representation for MedShockConsumerType has been revised, and old classes have been moved to LegacyOOsolvers.

  • The utility function for MedShockConsumerType has been algebraically rearranged, moving MedShk inside of the second CRRA term and adding a new parameter MedShift (default near zero). #1706

  • Function plot_func_slices has been added to HARK.utilities for convenient in-line plotting of multivariate functions #1695

  • New method AgentType.export_to_df added to flexibly export simulated history to a pandas.DataFrame. #1712

Minor Changes#

  • Revised exact_match option for DiscreteDistribution.draw to shuffle to be more robust to population draw size. #1691

  • multi_thread_commands[_fake] no longer requires empty parentheses to be included with each method name (now optional). #1692

  • Added repr method for DiscreteDistribution (and subclasses) to display basic information about itself.

  • All AgentTypes now have sensible defaults for track_vars if none is provided. #1693

  • AgentType.unpack and the new simulation structure appropriately handle solutions represented as dictionaries. #1709

  • calc_expectation now works with DiscreteDistributionLabeled instances when func references RVs by name, but not by position numbers. #1713

  • Repository now includes AI prompts to aid users when updating their project code from one version of HARK to another. #1696

  • 2D, 3D, and 4D interpolator classes no longer require that their arguments have the same size/shape; now they must only be jointly broadcastable. #1701

  • Method name change for NewKeynesianConsumerType: compute_steady_state is now compute_pe_steady_state. #1711

  • Life-cycle parameter calibrations from Carroll 1997 (QJE) have been added to ConsIndShockModel. #1715

0.17.0#

Release Date: January 4, 2026

Release Notes#

This release has many small improvements and fixes to existing HARK capabilities, listed below under Minor Changes. It also includes expanded and improved documentation/learning materials in examples/Gentle-Intro. To copy those example notebooks into a local working directory for easy use, simply execute these two commands in a Python environment and then follow the prompts:

from HARK import install_examples install_examples()

Four new consumption-saving models have been added, listed below under Major Changes.

There are some breaking changes:

  • TimeVaryingDiscreteDistribution has been removed; use IndexDistribution instead, and see #1592.

  • FixedPortfolioShareRiskyAssetConsumerType is removed, but now incorporated as RiskyAssetConsumerType with PortfolioBool=False. Default behavior of latter class is unchanged; see #1607.

  • The content of HARK.parallel has been moved to HARK.core, and the former is deprecated. Import from HARK.core and see #1614.

  • parse_ssa_life_table now returns one fewer survival probability by default, to match output length of parse_income_spec; pass terminal=True to restore old behavior. Argument min_age has been renamed to age_min for consistency. See #1629.

  • The parameter tau in RiskyContribModel has been renamed to WithdrawTax to match HARK notation style; see #1639.

  • Simulation method get_Rfree() has been renamed to get_Rport(), but no functional changes; see #1646.

  • The parameter DeprFac has been renamed to DeprRte to reflect its actual usage.

  • All distributions now default to using a random seed if none is provided. If your code relied on HARK defaulting to a specific seed, it will not reproduce exactly. See #1641.

  • The function apply_flat_income_tax has been removed, but it has not been used at all since 2016.

  • Content from ConsLabeledModel has been split up into files in the Labeled submodule. See #1684.

Major Changes#

  • Basic health investment model added in new module ConsHealthModel. #1567

  • Extensive margin medical care choice model added to ConsMedModel. #1595

  • TRP-style wealth-in-utility model without portfolio choice added in new module ConsWealthUtilityModel. #1634

  • “Capitalist spirit” style wealth-in-utility model added in new module ConsWealthUtilityModel. #1634

Minor Changes#

  • Fixed terminal solution initialization in IndShockConsumerTypeFast for proper numba compatibility, added CRRA=1 validation with clear error message, and expanded test coverage. #1649

  • Turns off use_infimum feature in ConsIndShock solver because it did not work properly when vFunc=True #1589

  • Consolidates TimeVaryingDiscreteDistribution into IndexDistribution. For time-varying discrete behavior, use IndexDistribution(distributions=[...]). #1592

  • Krusell-Smith model guide added to documentation. #1594

  • Added additional options and simplified syntax for non-default constructors when instantiating agents. #1591

  • Added options for custom indexer and pre-computation of coefficients to LinearInterp. #1593

  • Fixed bug that prevented combine_indep_dstn from working with Bernoulli distributions. #1581

  • Introductory / instructional notebooks significantly expanded. #1597

  • Lognormal discrete approximation math has been simplified. #1598

  • Directory structure for consumption-saving examples regularized. #1596

  • Fixed share model has been combined with RiskyAssetConsumerType’s PortfolioBool=False option. #1607

  • Deprecate HARK.parallel, moving the three functions there to HARK.core. #1614

  • Test coverage expanded to cover almost all content #1606 #1610 #1617 #1619 #1623 #1624 #1625 #1626 #1628 #1684

  • Consumption-saving models now aliased at HARK.models and HARK.ConsumptionSaving; some calibration tools also aliased at HARK.Calibration #1629

  • AgentType.solve() can be passed postsolve=False to skip post-processing call to post_solve(). #1631

  • The /examples directory can be copied to a directory of user’s choice with HARK.install_examples() #1630

  • Improved and expanded features for Parameters class in HARK.core #1627

  • Fixed the representation of the terminal period solution in ConsPrefShock #1638

  • Renamed tau to WithdrawTax in RiskyContribModel #1639

  • Valid bounds checking on make_grid_exp_mult #1640

  • Ensure utility functions return NaN for negative consumption #1640

  • Fixed a bug with resetting the RNG of IndexDistributions, restoring replicability of simulations #1643

  • Legacy simulation methods now use get_Rport() instead of get_Rfree() #1646

  • Fixed a bug that occured when changing an AgentType’s AgentCount attribute after simulating #1647

  • Add describe_distance() method to MetricObject, generating text description of how “distance” is calculated for an object #1648

  • Default behavior of seeds for distribution classes has been revised. #1641

  • Terminal solution representation for the “fast” solvers (using numba) has been cleaned up. #1649

  • Refactored ConsLabeledModel to use new HARK.Labeled subpackage with modular architecture (config, factories, transitions, solvers, solution, agents). Added comprehensive input validation, runtime warnings for numerical issues, and expanded test coverage. #1650

0.16.1#

Release Date: July 24, 2025

This release includes various small changes and improvements, as well as one significant new feature: (almost) all AgentType subclasses can now construct HA-SSJs (for use with the sequence_jacobian toolkit) in standard infinite horizon problems for arbitrary shock variables and arbitrary model outputs. This capability is powered by a new simulation structure that uses YAML-based model files to define dynamics, which in turn can be used to automatically transform HARK’s model solution representations (policy functions over continuous spaces) into the grid-based representation needed for efficient computation of the fake news algorithm. Don’t worry, that all happens under the hood.

See documentation notebooks in /examples/SequenceSpaceJacobians/ . The capabilities of our SSJ calculator will be expanded in the near future to include lifecycle models.

Major Changes#

  • Adds a new simulator structure based on YAML model files, replicating legacy simulation results #1545

  • Can convert solved HARK models into transition matrix-based discretized grid representations #1545

  • Can produce sequence-space Jacobians for infinite horizon problems for all HARK AgentType subclasses for which this is appropriate #1545

Minor Changes#

  • Allows lifecycle models to be solved backward starting from non-terminal period (with custom solution) #1545

  • Adds new interpolator class IndexedInterp with alternative notation for functions with mixed discrete-continuous domain #1545

  • New notebook with tutorial for (old and new) simulation methods #1545

  • Constructor make_grid_exp_mult allows linearly spaced grid with timestonest=-1 #1545

  • Adds documentation for new simulator structure and basic SSJ calculator #1545

  • Fixed a rare bug that could occur with unusual constructor dependencies resulting in incomplete updates. #1575

  • Added a reference to a trivial constructor that was missing from the WealthPortfolio model. #1583

  • Documentation files have been moved from /Documentation/ to /docs/ #1579

  • All tests have been consolidated into a single directory, rather than being scattered about. #1578

  • Add a special README so that the robots know we’re on their side when the singularity arrives. #1577

0.16.0#

Release Date: June 9, 2025

The items listed as “Developmental Features” are an independent system that is not connected to HARK’s existing model structure.

The most likely code-breaking change in this release is the reorganization of HARK.distribution. If your project code tells you that it can’t find the module HARK.distribution, just change the import name to HARK.distributions (note the plural s).

Additionally, several parameters have been lightly renamed:

aNrmInitMean –> kLogInitMean aNrmInitStd –> kLogInitStd pLvlInitMean –> pLogInitMean pLvlInitStd –> pLogInitStd

Finally, the legacy option for Rfree to be time-invariant has been removed from most models to allow the code to be simplified. If you used time-invariant Rfree, you will need to change your parameterization from Rfree = Rfree_value to T_cycle*[Rfree_value].

Major Changes#

  • Reorganizes the HARK.distribution file into HARK.distributions submodule with various files for readability and extensibility #1496

  • Regularizes AgentType initialization methods and moves all constructed model objects to constructors #1529 and #1530

Developmental Features#

  • Adds a discretize method to DBlocks and RBlocks #1460

  • Allows structural equations in model files to be provided in string form #1427

  • Introduces `HARK.parser’ module for parsing configuration files into models #1427

  • Allows construction of shocks with arguments based on mathematical expressions #1464

  • YAML configuration file for the normalized consumption and portolio choice #1465

Minor Changes#

  • Fixes bug in AgentPopulation that caused discretization of distributions to not work. #1275

  • Adds support for distributions, booleans, and callables as parameters in the Parameters class. #1387

  • Removes a specific way of accounting for ``employment’’ in the idiosyncratic-shocks income process. #1473

  • Adds income process constructor for the discrete Markov state consumption-saving model. #1484

  • Changes the behavior of make_lognormal_RiskyDstn so that the standard deviation represents the standard deviation of log(returns)

  • Adds detailed parameter and LaTeX documentation to most models.

  • Add PermGroFac constructor that explicitly combines idiosyncratic and aggregate sources of growth. 1489

  • Suppress warning from calc_stable_points when it would be raised by inapplicable AgentType subclasses. 1493

  • Fixes notation errors in IndShockConsumerType.make_euler_error_func from prior changes. 1495

  • Fixes typos in IdentityFunction interpolator class. 1492

  • Expands functionality of Cobb-Douglas aggregator for CRRA utility. 1363

  • Improved tracking of the bounds of support for distributions, and (some) solvers now respect those bounds when computing the “worst outcome”. 1524

  • Adds a new function for using Tauchen’s method to approximate an AR1 process. #1521

  • Adds additional functionality to the CubicHermiteInterp class, imported from scipy.interpolate. #1020

  • Allows users to pass a generic solution object to agent solvers to be used as the initial condition of backward induction. #1543

  • Adds support for Python 3.13 and related package updates. #1549

  • Move sim_birth methods to constructed distributions, lightly rename parameters. #1553

  • Cleans up warnings for the distance metric and prevents simulation history from being returned as output. #1563

  • Assorted small code cleanup tasks proposed by codex. [#1555] [#1556] [#1557] [#1558] [#1559] [#1560] [#1561] [#1562]

  • Updated and expanded documentation of sequence space Jacobian examples. [#1564] [#1568] [#1501] [#1490] [#1481] [#1475]

  • Improved documentation of cycles and timing of microeconomic models. #1571

0.15.1#

Release Date: June 15, 2024

This minor release was produced prior to CEF 2024 to enable public usage of HARK with the SSJ toolkit.

Major Changes#

none

Minor Changes#

  • Adds example of integration of HARK with SSJ toolkit. #1447

  • Maintains compatibility between EconForge interpolation and numba #1457

  • Renanmes ‘SSJ_example’ to ‘HANKFiscal_example’ so that it is more informative. #1509

0.15.0#

Release Date: June 4, 2024

Note: Due to major changes on this release, you may need to adjust how AgentTypes are instantiated in your projects using HARK. If you are manually constructing “complicated” objects like MrkvArray, they should be assigned to your instances after initialization, not passed as part of the parameter dictionary. See also the new constructor methodology for how to pass parameters for such constructed inputs.

This release drops support for Python 3.8 and 3.9, consistent with SPEC 0, and adds support for Python 3.11 and 3.12. We expect that all HARK features still work with the older versions, but they are no longer part of our testing regimen.

Major Changes#

  • Drop official support for Python 3.8 and 3.9, add support for 3.11 and 3.12. #1415

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

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

  • AgentTypeMonteCarloSimulator now requires model shock, parameter, and dynamics information to be organized into ‘blocks’. The DBlock object is introduced. #1411

  • RBlock object allows for recursive composition of DBlocks in models, as demonstrated by the AgentTypeMonteCarloSimulator #1417

  • Transtion, reward, state-rule value function, decision value function, and arrival value function added to DBlock #1417

  • All methods that construct inputs for solvers are now functions that are specified in the dictionary attribute constructors. #1410

  • Such constructed inputs can use alternate parameterizations / formats by changing the constructor function and providing its arguments in parameters.

  • Move HARK.datasets to HARK.Calibration for better organization of data and calibration tools. #1430

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

  • Removed ‘parameters’ from new block definitions; these are now ‘calibrations’ provided separately.

  • Create functions for well-known and repeated calculations in single-function solvers. 1395

  • Re-work WealthPortfolioSolver to use approximate EGM method #1404

  • Default parameter dictionaries for AgentType subclasses have been “flattened”: all parameters appear in one place for each model, rather than inheriting from parent models’ dictionaries. The only exception is submodels within a file when only 1 or 2 parameters are added or changed. #1425

  • Fix minor bug in HARK.distributions.Bernoulli to allow conversion into DiscreteDistributionLabeled. #1432

0.14.1#

Release date: February 28, 2024

Major Changes#

none

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

0.14.0#

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

0.13.0#

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.distributions.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 interpolation.py for fast multilinear interpolation. #1151

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

  • Adds class DecayInterp to econforgeinterp.py. 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.distributions. 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 ConsIndShockModel.py 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.distributions.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 ConsIndShockModel.py 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

0.12.0#

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 ConsGenIncProcessModel.py 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

0.11.0#

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](econ-ark/HARK#] (#897)[https://github.com/econ-ark/HARK/pull/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 Model.call; 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 MANIFEST.in. 957

0.10.8#

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

0.10.7#

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

0.10.6#

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#

0.10.5#

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#

0.10.4#

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

0.10.3#

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 simultation.py. (#408)

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

0.10.2#

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)

0.10.1.dev5#

Release Date: 09-25-2019

Minor Changes#

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

0.10.1.dev4#

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)

0.10.1.dev3#

Release Date: 07-23-2019

Minor Changes#

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

0.10.1.dev2#

Release Date: 07-22-2019

Minor Changes#

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

0.10.1.dev1#

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.

0.10.1#

Release Date: 05-30-2019

No changes from 0.10.0.dev3.

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)

0.10.0.dev2#

Release Date: 04-18-2019

Major Changes#

None

Minor Changes#

  • Fix verbosity check in ConsIndShockModel. (250)

Other Changes#

None

0.10.0.dev1#

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 econ-ark.org. (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.