generalize configuration of income process

[sbenthall]

Most models now inherit the lognormal income process from ConsIndShock:

HARK/HARK/ConsumptionSaving/ConsIndShockModel.py

Line 2120 in 7dd38fc

IncomeDstn, PermShkDstn, TranShkDstn = self.constructLognormalIncomeProcessUnemployment()

HARK/HARK/ConsumptionSaving/ConsIndShockModel.py

Line 2581 in 7dd38fc

def constructLognormalIncomeProcessUnemployment(self):

But some models have subtle changes to the process, requiring some overwriting of the model function:

https://github.com/econ-ark/HARK/blob/0.10.6/HARK/cstwMPC/cstwMPC.py#L51-L74

It may be possible to design this so that the income process distributions are defined when a model is parameterized. This would lead to a more consistent API. See #620

[llorracc]

Incidentally, I have long felt that our default should be using a lognormal truncated at some (large) number of standard deviations (like, 5). This would ameliorate (though not fix) the fact that as we increase the number of gridpoints, the minimum possible realization moves substantially. This matters because a good bit of the logic of the model depends on the value of the minimum possible realization.

An even better approach would be to institute an infinitesimal probability of a "minimum income event" corresponding exactly to the truncation bound of the distribution. That would effectively eliminate the problem because the minimum possible income would remain the same as long as the truncation threshold remained the same, regardless of how many (or few) points were used in the approximation.

[mnwhite]

I strongly, strongly endorse this. I think we've had this discussion? But with Sebastian's in progress refactoring of our distributions as objects (as they should be been all along), our UnivariateDistribution class can simply have attributes for its minimum and maximum value. Right now, HARK would correctly interpret the addition of a point mass with machine epsilon probability at zero, but it looks a little funny to do it that way. Much more clear to be able to ask for TranShkDstn.Xmin (I know we're moving away from X).

…

On Tue, May 5, 2020 at 2:20 PM Christopher Llorracc Carroll < ***@***.***> wrote: Incidentally, I have long felt that our default should be using a lognormal truncated at some (large) number of standard deviations (like, 5). This would ameliorate (though not fix) the fact that as we increase the number of gridpoints, the minimum possible realization moves substantially. This matters because a good bit of the logic of the model depends on the value of the minimum possible realization. An even better approach would be to institute an infinitesimal probability of a "minimum income event" corresponding exactly to the truncation bound of the distribution. That would effectively eliminate the problem because the minimum possible income would remain the same as long as the truncation threshold remained the same, regardless of how many (or few) points were used in the approximation. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#673 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADKRAFLRSIGBKJY77M7DNILRQBKFVANCNFSM4MZTP4JA> .

[sbenthall]

This issue is now quite complicated because it's not clear whether it refers to the original scope or the one CDC introduced in the comment.

It looks like #804 is a more fleshed out version of the issues raised in the comment here.

So this ticket refers to the original item now,

[sbenthall]

Now redundant with #620