Add proper Krusell-Smith model #762

mnwhite · 2020-07-20T21:25:37Z

CDC has long wanted HARK to be able to solve the Krusell-Smith (1998) model verbatim; this PR adds that capability. Previously, we had "fake KS", as our Markov ConsAggShock model didn't have serially correlated unemployment.

This PR includes new AgentType and Market subclasses in ConsAggShockModel.py, as the Krusell-Smith model is so specialized that trying to shoehorn trivial versions of existing HARK model objects would just overcomplicate things. The one compromise with HARK is that the aggregate saving rule (regressing log A_t on log M_t conditional on the state of the economy) is still the equilibrium object, rather than a straight log K_t on log K_{t-1} like in the Krusell-Smith paper. HARK's existing models use the "A on M" approach because we have aggregate risk other than the binary state of the macroeconomy, so K_{t-1} is not a sufficient statistic for the aggregate state in period t.

Example has been added to the ConsAggShock example file, replacing the old "fake KS" example. Tests have also been added. Default parameters are set to those in the KS paper, with the exception of LbrInd (\tilde{\ell}), whose value is not provided in the paper as far as I can tell. Their results are nearly identically replicated.

This code does not handle the KS version of discount factor heterogeneity, as CDC did not reply to my query about whether he wanted that included.

CDC really wants to be able to solve KS, so I wrote a solver just for that model. Completely untested, and needs an AgentType and Market to go with it. First 2/3 of the solution code can be done in pre-computational step.

Walked Mridul through KS model math and solver and began writing AgentType subclass for it. Still need to do the obnoxious arithmetic of the good-bad bad bad-good idiosyncratic transitions.

Made tiny example file, solveKrusellSmith seems to work as intended. Need to make a special Market subclass for it.

Had to force idiosyncratic shocks to exactly match specified probabilities, but it works. Need to clean up model file and do a bit of work to speed up the process.

Default KS model parameters now more closely match version in their paper, but it also takes a shockingly long time to solve. Moved dictionaries into model file, will move example code next.

Also deleted figure-producing progress code lines.

These should pass and don't add much to test time.

alanlujan91 · 2020-07-20T21:48:45Z

I have ell tilde in my class notes as 0.3271, if that helps

mnwhite · 2020-07-20T23:17:53Z

This does help, as I had hand-calibrated it to 0.325 to fit the KS results. I was worried that they used 1/3 or 0.33 and there was still an error in the code. Given that all the other calibrated numbers are very rough and the entire model has no scale, do you know where/why they used this specific number to 4 digits? Is it in the paper and I missed it?

…

On Mon, Jul 20, 2020, 5:49 PM alanlujan91 ***@***.***> wrote: I have ell tilde in my class notes as 0.3271, if that helps — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#762 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADKRAFIWUV5MAZ2XHIBWL4TR4S3UZANCNFSM4PCZ675Q> .

alanlujan91 · 2020-07-20T23:42:24Z

I don't think the number is on the paper. As far as I remember, the source of this number is my prof. emailing one of the authors.

mnwhite · 2020-07-21T00:04:25Z

Thank you very much for confirming this.

…

On Mon, Jul 20, 2020, 7:42 PM alanlujan91 ***@***.***> wrote: I don't think the number is on the paper. As far as I remember, the source of this number is my prof. emailing one of the authors. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#762 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADKRAFLIAS5SGD2Y4T7CKOTR4TI63ANCNFSM4PCZ675Q> .

sbenthall · 2020-07-21T01:11:33Z

HARK/ConsumptionSaving/ConsAggShockModel.py

+        # This can be written more concisely, but is intentionally explicit for
+        # the sake of making it very easy to read
+        if self.MrkvNow == 0: # Bad macroeconomic conditions
+            these = np.logical_not(self.EmpNow) # Unemployed agents


maybe name this variable unemployed_agents instead of these.
Save you having to explain it in comment :)

I suppose, but this is used throughout HARK for binary indexing arrays.

fair enough. just a suggestion.

sbenthall · 2020-07-21T01:15:15Z

HARK/ConsumptionSaving/ConsAggShockModel.py

+        # Add consumer-type specific objects
+        self.time_vary = []
+        self.time_inv = ['DiscFac', 'CRRA',]
+        self.poststate_vars = ['aNow','EmpNow']


Not a critique of this code but a general question...

I see that the poststate_vars are listed explicitly here.
Would it be silly to list (a) the regular states, (b) the control variables, and (c) the shocks in lists like these?
I mean, in a future implementation.

That would make it easier to do things like #760 and #761

That can be done. They're listed here and in all HARK models because initializeSim() initializes them.

Cool. I think I can demo how this would look in a PR for the next weekly meeting.

sbenthall · 2020-07-21T01:16:25Z

HARK/ConsumptionSaving/ConsAggShockModel.py

+        Mcount = self.Mgrid.size
+
+        # Make tiled array of end-of-period idiosyncratic assets (order: a, M, s, s')
+        aNow_tiled = np.tile(np.reshape(self.aGrid, [aCount, 1, 1, 1]), [1, Mcount, 4, 4]) # shape (aCount, Mcount, 4, 4)


In general, I think the code formatting of this PR is really nice.
I was going to ask if you used black for it?

I'm surprised by this long line (in terms of number of characters) though.

No, I didn't use black. All of my comments with the shape of arrays can be removed (will do that now). I put those in while explaining the code the Mridul.

mnwhite · 2020-07-21T01:18:24Z

Ack, just caught that I forgot to put pre-computed arrays into the solveKrusellSmith docstring.

sbenthall · 2020-07-21T01:27:00Z

I've made a few minor comments.
Otherwise, this looks good to me.
I'm probably not the best person to judge the economics of it, but I'd be happy to merge if there's no objections, after Matt gets in his "ack!' fix.

Fixed and expanded some comments, added comparison of our results to KS' published numbers (in the example). Adjusted LbrInd parameter to use value that Alan Lujan says his prof got from KS. Future readers: please cite this chain for the parameter value missing from Krusell-Smith (1998).

sbenthall · 2020-07-21T19:26:21Z

HARK/ConsumptionSaving/ConsAggShockModel.py

+        self.makeGrid()
+        self.updateSolutionTerminal()
+
+    def getEconomyData(self, Economy):


Should be lower case -- see how the GitHub syntax highlighting misformats this as a class name because it is capitalized.

I see this is common to other uses of getEconomyData though.

sbenthall · 2020-07-21T19:29:02Z

HARK/ConsumptionSaving/ConsAggShockModel.py

+        self.MrkvIndArray = Economy.MrkvIndArray     # Transition probabilities among discrete states
+        self.MrkvAggArray = Economy.MrkvArray        # Transition probabilities among aggregate discrete states
+        self.addToTimeInv('Mgrid', 'AFunc', 'DeprFac', 'CapShare', 'UrateB', 'LbrInd',\
+                          'UrateG', 'ProdB', 'ProdG', 'MrkvIndArray', 'MrkvAggArray')


See #764 . If those aspects of the economy that needed to be brought into the Agent were available in a list (as most of them are here), then it would be possible to write this method generically as a loop over those values, rather than writing out each one individually.

This is common across all uses of getEconomyData

mnwhite · 2020-07-21T19:54:14Z

But not all of them are just "take this attribute of that and make it an attribute of self". Some of them change names or do a very small bit of coding. Yes, that can be done in the lines below. Change it as you see fit.

…

On Tue, Jul 21, 2020 at 3:29 PM Sebastian Benthall ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In HARK/ConsumptionSaving/ConsAggShockModel.py <#762 (comment)>: > + self.T_sim = Economy.act_T # Need to be able to track as many periods as economy runs + self.kInit = Economy.KSS # Initialize simulation assets to steady state + self.MrkvInit = Economy.MrkvNow_init # Starting Markov state for the macroeconomy + self.Mgrid = Economy.MSS*self.MgridBase # Aggregate market resources grid adjusted around SS capital ratio + self.AFunc = Economy.AFunc # Next period's aggregate savings function + self.DeprFac = Economy.DeprFac # Rate of capital depreciation + self.CapShare = Economy.CapShare # Capital's share of production + self.LbrInd = Economy.LbrInd # Idiosyncratic labor supply (when employed) + self.UrateB = Economy.UrateB # Unemployment rate in bad state + self.UrateG = Economy.UrateG # Unemployment rate in good state + self.ProdB = Economy.ProdB # Total factor productivity in bad state + self.ProdG = Economy.ProdG # Total factor productivity in good state + self.MrkvIndArray = Economy.MrkvIndArray # Transition probabilities among discrete states + self.MrkvAggArray = Economy.MrkvArray # Transition probabilities among aggregate discrete states + self.addToTimeInv('Mgrid', 'AFunc', 'DeprFac', 'CapShare', 'UrateB', 'LbrInd',\ + 'UrateG', 'ProdB', 'ProdG', 'MrkvIndArray', 'MrkvAggArray') See #764 <#764> . If those aspects of the economy that needed to be brought into the Agent were available in a list (as most of them are here), then it would be possible to write this method generically as a loop over those values, rather than writing out each one individually. This is common across all uses of getEconomyData — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#762 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADKRAFIL47CWPTZGFVUU4DLR4XUA7ANCNFSM4PCZ675Q> .

sbenthall · 2020-07-21T19:59:30Z

Made #766 for that issue.
I'll work on it after this PR is in so as not to block it.

sbenthall · 2020-07-21T19:59:51Z

Is this PR ready to merge, from your perspective, @mnwhite ?

mnwhite · 2020-07-21T20:11:56Z

Yes, I think so.

…

On Tue, Jul 21, 2020 at 4:00 PM Sebastian Benthall ***@***.***> wrote: Is this PR ready to merge, from your perspective, @mnwhite <https://github.com/mnwhite> ? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#762 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADKRAFK3V6FQ4FNXUMD6NM3R4XXUTANCNFSM4PCZ675Q> .

mnwhite · 2020-07-21T20:28:42Z

Ah, the one thing left is to make the ipynb example match the py. There's an automated thing for that?

…

On Tue, Jul 21, 2020 at 4:25 PM Sebastian Benthall ***@***.***> wrote: Merged #762 <#762> into master. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#762 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADKRAFPPVIJNO52STJGYY7TR4X2SFANCNFSM4PCZ675Q> .

sbenthall · 2020-07-21T20:36:25Z

! Oh, perhaps I merged too soon !

Yes, I think what you are looking for is jupytext

https://jupytext.readthedocs.io/en/latest/using-cli.html

mnwhite · 2020-07-21T20:47:02Z

We can just do it in another branch.

…

On Tue, Jul 21, 2020 at 4:36 PM Sebastian Benthall ***@***.***> wrote: ! Oh, perhaps I merged too soon ! Yes, I think what you are looking for is jupytext https://jupytext.readthedocs.io/en/latest/using-cli.html — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#762 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADKRAFJ5J3ZV5RKWZMS26QLR4X35RANCNFSM4PCZ675Q> .

mnwhite added 7 commits July 16, 2020 21:30

First draft of Krusell-Smith solver

345fef9

CDC really wants to be able to solve KS, so I wrote a solver just for that model. Completely untested, and needs an AgentType and Market to go with it. First 2/3 of the solution code can be done in pre-computational step.

Live coding work with Mridul

0f9f364

Walked Mridul through KS model math and solver and began writing AgentType subclass for it. Still need to do the obnoxious arithmetic of the good-bad bad bad-good idiosyncratic transitions.

KS idiosyncratic solver works

48c9b14

Made tiny example file, solveKrusellSmith seems to work as intended. Need to make a special Market subclass for it.

KS model works

0df69fb

Had to force idiosyncratic shocks to exactly match specified probabilities, but it works. Need to clean up model file and do a bit of work to speed up the process.

Clean up KS model, adjust default parameters

5d233b5

Default KS model parameters now more closely match version in their paper, but it also takes a shockingly long time to solve. Moved dictionaries into model file, will move example code next.

Move KS example code

ff3d19c

Also deleted figure-producing progress code lines.

Add Krusell-Smith tests

9f1cd68

These should pass and don't add much to test time.

sbenthall reviewed Jul 21, 2020

View reviewed changes

sbenthall mentioned this pull request Jul 21, 2020

Aiyagari chimera econ-ark/DARKolo#12

Open

sbenthall reviewed Jul 21, 2020

View reviewed changes

sbenthall mentioned this pull request Jul 21, 2020

compress getEconomyData #766

Open

Merge branch 'master' into AggShockMicroMarkov

5ee69a8

sbenthall merged commit 112a01a into master Jul 21, 2020

MridulS deleted the AggShockMicroMarkov branch August 24, 2020 18:06

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add proper Krusell-Smith model #762

Add proper Krusell-Smith model #762

mnwhite commented Jul 20, 2020

alanlujan91 commented Jul 20, 2020

mnwhite commented Jul 20, 2020 via email

alanlujan91 commented Jul 20, 2020

mnwhite commented Jul 21, 2020 via email

sbenthall Jul 21, 2020

mnwhite Jul 21, 2020

sbenthall Jul 21, 2020

sbenthall Jul 21, 2020

mnwhite Jul 21, 2020

sbenthall Jul 21, 2020

sbenthall Jul 21, 2020

mnwhite Jul 21, 2020

mnwhite commented Jul 21, 2020

sbenthall commented Jul 21, 2020

sbenthall Jul 21, 2020

sbenthall Jul 21, 2020

mnwhite commented Jul 21, 2020 via email

sbenthall commented Jul 21, 2020

sbenthall commented Jul 21, 2020

mnwhite commented Jul 21, 2020 via email

mnwhite commented Jul 21, 2020 via email

sbenthall commented Jul 21, 2020

mnwhite commented Jul 21, 2020 via email

Add proper Krusell-Smith model #762

Add proper Krusell-Smith model #762

Conversation

mnwhite commented Jul 20, 2020

alanlujan91 commented Jul 20, 2020

mnwhite commented Jul 20, 2020 via email

alanlujan91 commented Jul 20, 2020

mnwhite commented Jul 21, 2020 via email

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mnwhite commented Jul 21, 2020

sbenthall commented Jul 21, 2020

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mnwhite commented Jul 21, 2020 via email

sbenthall commented Jul 21, 2020

sbenthall commented Jul 21, 2020

mnwhite commented Jul 21, 2020 via email

mnwhite commented Jul 21, 2020 via email

sbenthall commented Jul 21, 2020

mnwhite commented Jul 21, 2020 via email