AgentPopulation solution is sometimes returning a negative RiskyShare

[sbenthall]

Under some conditions, the AgentPopulation solution's ShareFunc is returning a negative risky share from its ShareFunc.

Since we use this value to create the agents' target value of share ownership, this wreaks havoc on the simulation.

This is one way to reproduce the error with some standard scripts:

https://gist.github.com/sbenthall/d0b4b581c683a0e3bd0e10b59dd4011d

[sbenthall]

Please include an automated unit test along with the fix for this issue. A failing test that reproduces this error under simple, isolated conditions is a good contribution in its own right, as right now it's hard to know under what conditions the solution function has this error.

[alanlujan91]

(sharkfin-dev) C:\Users\alujan\GitHub\alanlujan91\SHARKFin>python simulate/run_any_simulation.py --simulation Attention --seed 790 --quarters 3 --attention 0.99 --dphm 1500000 --p1 0.1 --p2 .1 --dividend_growth_rate 1.002 output/20221006_ras_d
Traceback (most recent call last):
  File "C:\Users\alujan\GitHub\alanlujan91\SHARKFin\simulate\run_any_simulation.py", line 10, in <module>
    from simulate.parameters import (
ModuleNotFoundError: No module named 'simulate'

How do I set this up?

[sbenthall]

Did you go through the installation steps in the README?

(If you did, and are still getting this error, that's a documentation problem I'll need to fix)

[alanlujan91]

Yes, and can run tests locally and they all pass. but can't seem to run a simulation on command line

[sbenthall]

Dang. Try git pull master to get the most recent commit and then running the script again.

[alanlujan91]

The trouble seems to be my environment not recognizing the subdirectory simulate as a module. This subdirectory is outside of the sharkfin subdirectory.

[sbenthall]

I agree with your analysis. I've made another change. Try:

• pulling from master
• running pip install -e . again
• running the run_any_simulation script again

[alanlujan91]

great, that worked, I'll look into the bug now

[sbenthall]

This error happens sometimes when agents extrapolate RiskyShare solutions beyond the grid for which these functions have been evaluated.

A temporary fix, which is increasing the size of the solution grid to capture a broader span of values, has been implemented in #161. Documenting this issue is #162 .

There are several more robust solutions, which are now in other tickets:

• The higher-order solution interpolaters could be extended with decay and intercept features to control the extrapolation behavior, see #162
• The HARK Uniform distribution discretizer could be modified to optionally include the endpoints: https://github.com/econ-ark/HARK/issues/1178 . This way on the low end, the grid can always have two points with 0 risky share value, preventing negative extrapolations
• Some sort of system should be built for taking the output of a simulation run and using it to calibrate the model parameters recursively.