Transition Matrix Methods #1155

wdu9 · 2022-07-21T07:09:53Z

Please ensure your pull request adheres to the following guidelines:

[x ] Tests for new functionality/models or Tests to reproduce the bug-fix in code.
[x ] Updated documentation of features that add new functionality.
[x ] Update CHANGELOG.md with major/minor changes.

codecov-commenter · 2022-07-21T07:19:38Z

Codecov Report

Base: 73.62% // Head: 72.83% // Decreases project coverage by -0.78% ⚠️

Coverage data is based on head (59c16d3) compared to base (e6c936b).
Patch coverage: 32.74% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1155      +/-   ##
==========================================
- Coverage   73.62%   72.83%   -0.79%     
==========================================
  Files          72       72              
  Lines       11509    11735     +226     
==========================================
+ Hits         8473     8547      +74     
- Misses       3036     3188     +152

Impacted Files	Coverage Δ
HARK/utilities.py	`29.63% <10.71%> (-5.23%)`	⬇️
HARK/ConsumptionSaving/ConsIndShockModel.py	`79.27% <36.36%> (-5.99%)`	⬇️
...nsumptionSaving/tests/test_IndShockConsumerType.py	`76.73% <100.00%> (+1.57%)`	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

llorracc · 2022-07-21T15:48:08Z

Will, If you think this is ready to merge, could you add a tag to that effect to it (we have a predefined tag for that).

…

On Thu, Jul 21, 2022 at 3:10 AM wdu9 ***@***.***> wrote: Please ensure your pull request adheres to the following guidelines: - [x ] Tests for new functionality/models or Tests to reproduce the bug-fix in code. - [x ] Updated documentation of features that add new functionality. - [x ] Update CHANGELOG.md with major/minor changes. ------------------------------ You can view, comment on, or merge this pull request online at: #1155 Commit Summary - eb2d369 <eb2d369> Transition Matrix Methods - 6d55939 <6d55939> Merge branch 'econ-ark:master' into Transition_Matrix - 01f1dad <01f1dad> transition Matrix methods - 6dd504c <6dd504c> Merge branch 'Transition_Matrix' of https://github.com/wdu9/HARK into Transition_Matrix File Changes (5 files <https://github.com/econ-ark/HARK/pull/1155/files>) - *M* Documentation/CHANGELOG.md <https://github.com/econ-ark/HARK/pull/1155/files#diff-642febef01c09084c712d527a5b0a1c04eb1fb108202c2a1445947f31db48f07> (2) - *M* HARK/ConsumptionSaving/ConsIndShockModel.py <https://github.com/econ-ark/HARK/pull/1155/files#diff-7f09d28a3d4136ae35d8835cd1352f060da6e0114fe617c34374c8b57d41a57e> (300) - *M* HARK/ConsumptionSaving/tests/test_IndShockConsumerType.py <https://github.com/econ-ark/HARK/pull/1155/files#diff-46adcdf1fe37ad7df97b9cb94fdba2e287e910387c68b9b1158c146eb5ccba85> (47) - *M* HARK/utilities.py <https://github.com/econ-ark/HARK/pull/1155/files#diff-30bbf6349341e5b97b6a40790de996e53763b5e37e896cb12ab6254ee024e439> (150) - *A* examples/ConsIndShockModel/IndShockConsumerType_Transition_Matrix_Example.ipynb <https://github.com/econ-ark/HARK/pull/1155/files#diff-ee32ef5bc9227ad4f730a96c49d67410b11809c46fa69f69cb5493e29c7a092a> (1075) Patch Links: - https://github.com/econ-ark/HARK/pull/1155.patch - https://github.com/econ-ark/HARK/pull/1155.diff — Reply to this email directly, view it on GitHub <#1155>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAKCK7Y7S67URBPGYXE7ENTVVDZUZANCNFSM54GQDPOQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- - Chris Carroll

wdu9 · 2022-07-22T00:02:40Z

Will, If you think this is ready to merge, could you add a tag to that effect to it (we have a predefined tag for that).

This is ready to be merged.

I think we discussed this before but I don't think I am able to add tags.

AMonninger · 2022-07-22T17:04:37Z

Happy to review this

alanlujan91 · 2022-07-29T13:20:17Z

HARK/ConsumptionSaving/ConsIndShockModel.py

+                #Dist_pGrid is taken to cover most of the ergodic distribution
+                p_variance = self.PermShkStd[0]**2 #set variance of permanent income shocks
+                max_p = max_p_fac*(p_variance/(1-self.LivPrb[0]))**0.5 #Maximum Permanent income value
+                one_sided_grid = make_grid_exp_mult(1.05+1e-3, np.exp(max_p), num_points, 2)


min_p should maybe be a parameter to this model instead of hardcoded

Agreed, good catch.

AMonninger · 2022-08-01T15:47:37Z

For calc_ergodic_dist: (based on numerical summer school lecture)
As we are only interested in one eigenvalue, we can transform the problem:

b = vector(Nx1) with element (1,1) = 1, rest zeros
transition matrix (T): replace first row with zeros and element (1,1)=1
solve: s = T^(-1)*b
Normalize: s = s/sum(s)

Should be faster then searching for the eigenvalues with sp.linalg.eigs

alanlujan91 · 2022-08-01T15:53:43Z

HARK/ConsumptionSaving/ConsIndShockModel.py

+    jump_to_grid,
+    jump_to_grid_fast,


More descriptive names would be useful for extensions/alternative uses. In the past, we've used _fast to refer to numba-fied methods that do the same as their reference methods, but here, both are using numba.jit.

alanlujan91 · 2022-08-01T16:12:55Z

HARK/ConsumptionSaving/ConsIndShockModel.py

+                for i in range(m_density):
+                    axtra_shifted = np.delete(aXtra_Grid,-1) 
+                    axtra_shifted = np.insert(axtra_shifted, 0,1.00000000e-04)
+                    dist_betw_pts = aXtra_Grid - axtra_shifted
+                    dist_betw_pts_half = dist_betw_pts/2
+                    new_A_grid = axtra_shifted + dist_betw_pts_half
+                    aXtra_Grid = np.concatenate((aXtra_Grid,new_A_grid))
+                    aXtra_Grid = np.sort(aXtra_Grid)
+                self.dist_mGrid =  aXtra_Grid


is this "thickening" the grid of the distribution of cash-on-hand? what is particular about this construction that is different from make_grid_exp_mult? If this is doing something that can't be accomplished by make_grid_exp_mult maybe it should be its own utility method.

It does thicken if that means to increase the density. Chris had asked me to do this a while back but I'm not sure whethe rmake_grid_exp_mult can thicken, I feel as though it just concentrates points near the bottom end.

alanlujan91 · 2022-08-01T16:16:57Z

HARK/ConsumptionSaving/ConsIndShockModel.py

+        elif self.cycles > 1:
+            raise Exception('define_distribution_grid requires cycles = 0 or cycles = 1')
+
+        elif self.T_cycle != 0:


So this condition is (cycles == 1 and T_cycle !=0)

is T_cycle < 1 ever?

yes this condition is meant as (cycles == 1 and T_cycle !=0).

T_cycle is never less than 1 I believe

alanlujan91 · 2022-08-01T16:21:32Z

HARK/ConsumptionSaving/ConsIndShockModel.py

+        else:
+            m_points = num_pointsM
+
+        if self.cycles == 0:


This could be
if (self.cycles == 0) or (self.cycles == 1 and self.T_cycle !=0):
to avoid code repetition I think

The T_cycle !=0 problem is exclusive of the cycles=0 problem so it is not unfortunately.

alanlujan91 · 2022-08-01T16:23:27Z

HARK/ConsumptionSaving/ConsIndShockModel.py

+            LivPrb = self.LivPrb[0] # Update probability of staying alive
+
+            #New borns have this distribution (assumes start with no assets and permanent income=1)
+            NewBornDist = jump_to_grid(tran_shks,np.ones_like(tran_shks),shk_prbs,dist_mGrid,dist_pGrid)


Is this a probability distribution? Would it be useful to use our DiscreteDistribution constructor here?

This is a function that takes in a grid, a set of values that may lie between the gridpoints, and finally the associated probabilities of each the values and returns the probability mass of each gridpoint. I don't think itd be useful to use DiscreteDistribution here

alanlujan91 · 2022-08-01T16:24:53Z

HARK/ConsumptionSaving/ConsIndShockModel.py

+        elif self.cycles > 1:
+            raise Exception('calc_transition_matrix requires cycles = 0 or cycles = 1')
+
+        elif self.T_cycle!= 0 :


Is there a model that has cycles==1 and T_cycle==0? I'm thinking this should never happen, and if it does, it's probably a bug.

I'm not sure, that is why I was so specific on the conditional cases.

alanlujan91

Left a few in-line comments. Overall looks great!

Comments about new tools or using DiscreteDistribution can probably be addressed later.

llorracc · 2022-09-19T01:32:24Z

@wdu9 ,

Glad it is now passing tests. Will ask @alanlujan91 to make a final review before merging. (Thanks to @AMonninger for earlier comments; maybe he could take a final look with an eye to how it might work for his durables problem; but @alanlujan91 should be the decider about merging).

@wdu, one thing I'd ask you to add to the chain is some insight about why it was failing before and is now working. What did you change? If it's another instance of our tests just defaulting to a degree of precision that doesn't make sense, that would be a useful insight. If it's something else, that would also be useful to know.

wdu9 · 2022-09-19T01:37:39Z

For calc_ergodic_dist: (based on numerical summer school lecture) As we are only interested in one eigenvalue, we can transform the problem:
1. b = vector(Nx1) with element (1,1) = 1, rest zeros

2. transition matrix (T): replace first row with zeros and element (1,1)=1

3. solve: s = T^(-1)*b

4. Normalize: s = s/sum(s)
Should be faster then searching for the eigenvalues with sp.linalg.eigs

I had this in the code however sometimes it would fail for unknown reasons. It was indeed twice as fast at very large grid dimensions.

wdu9 · 2022-09-19T01:49:35Z

I reduced the tolerance of the self.assertalmostequal function to resolve the original problem where the self assertion was off by a difference of 10e-6.
This minuscule difference was due to the randomness in the sp.linalg.eigs function under the calc_ergodic_dist method.
Although, I eliminated the randomness in my own code and the test was passing in my directory, it was failing in the tests here.

AMonninger · 2022-09-29T18:04:47Z

Hi Will, some additional comments:

Can you make the linear algebra described above as an option?
For transition matrices in a lifecycle model, it saves a lot of RAM when you make it sparse as well. Eg.

TranMatrix_sparse = sp.csr_matrix(TranMatrix)
self.tran_matrix.append(TranMatrix_sparse)
Helps for many grid points or large number of periods. If you use the transitionmatrices afterwards, you can use: MODEL.tran_matrix[i].A

llorracc assigned AMonninger Jul 22, 2022

alanlujan91 self-requested a review July 22, 2022 18:52

alanlujan91 removed the Ready-To-Merge Has been reviewed and when branch is updated and checks pass it should be merged label Jul 22, 2022

alanlujan91 reviewed Jul 29, 2022

View reviewed changes

alanlujan91 reviewed Aug 1, 2022

View reviewed changes

wdu9 closed this Sep 9, 2022

wdu9 force-pushed the Transition_Matrix branch from e64c32d to ec3a7bf Compare September 9, 2022 07:22

wdu9 reopened this Sep 9, 2022

wdu9 closed this Sep 30, 2022

wdu9 force-pushed the Transition_Matrix branch from 3c2d1b7 to e6c936b Compare September 30, 2022 08:04

wdu9 added 2 commits September 30, 2022 04:12

Update pull requests with recently merged pull requests

178ac84

add Example Notebook on TranMat Methods

59c16d3

wdu9 reopened this Sep 30, 2022

alanlujan91 merged commit dc6cbc6 into econ-ark:master Oct 4, 2022

sbenthall added this to the 0.13.0 milestone Jan 4, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Transition Matrix Methods #1155

Transition Matrix Methods #1155

wdu9 commented Jul 21, 2022

codecov-commenter commented Jul 21, 2022 •

edited

Loading

llorracc commented Jul 21, 2022 via email

wdu9 commented Jul 22, 2022 •

edited by llorracc

Loading

AMonninger commented Jul 22, 2022

alanlujan91 Jul 29, 2022

wdu9 Aug 1, 2022

AMonninger commented Aug 1, 2022

alanlujan91 Aug 1, 2022

wdu9 Aug 1, 2022

alanlujan91 Aug 1, 2022

wdu9 Aug 1, 2022

alanlujan91 Aug 1, 2022

alanlujan91 Aug 1, 2022

wdu9 Aug 1, 2022

alanlujan91 Aug 1, 2022

wdu9 Aug 5, 2022

alanlujan91 Aug 1, 2022

wdu9 Aug 1, 2022

alanlujan91 Aug 1, 2022

wdu9 Aug 1, 2022

alanlujan91 left a comment

llorracc commented Sep 19, 2022

wdu9 commented Sep 19, 2022

wdu9 commented Sep 19, 2022

AMonninger commented Sep 29, 2022

Transition Matrix Methods #1155

Transition Matrix Methods #1155

Conversation

wdu9 commented Jul 21, 2022

codecov-commenter commented Jul 21, 2022 • edited Loading

Codecov Report

llorracc commented Jul 21, 2022 via email

wdu9 commented Jul 22, 2022 • edited by llorracc Loading

AMonninger commented Jul 22, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AMonninger commented Aug 1, 2022

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

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

alanlujan91 left a comment

Choose a reason for hiding this comment

llorracc commented Sep 19, 2022

wdu9 commented Sep 19, 2022

wdu9 commented Sep 19, 2022

AMonninger commented Sep 29, 2022

codecov-commenter commented Jul 21, 2022 •

edited

Loading

wdu9 commented Jul 22, 2022 •

edited by llorracc

Loading