**Why do we need precarity plots?**

- Precarity showcases the latent behavior that arises from the history of the decisions made by the individual. Removing precarity plots does not really make sense since if we remove it and try to use colors as the feature that show cases the precarity, then it could be seen as us picking two individuals and showing that they behave differently and does not really justify the introduction of precarity. It would just look like two different individuals forced to behave in two different ways.
- One other point we had was the reviewers were not questioning the validity of using the precarity plots.
- Also, even though we do set initial precarity explicitly, we do not force the system to follow any specific precarity path. We only let the system evolve and use that data as a measure of precarity. We do agree that the initial precarity we use might not be the best option, we still do not intervene with other precarity values and the initial precarity is just one small thing.
- Adding plots such as utility instead of precarity does not make sense. same plots of different variable plus the utility plots are always very hard to distinguish.
- if we remove precarity, the paper would be like any other asset simulation tool out there
- the sequecne or history argument through the whole paper would go away
- if we used it as a concept only and do merely monetary plots, then one would say who says this phenomenon is precarity? it could be called phenomenon "xyz" since socioeconomics is very vast and without proper quantification how can we link it to a specific socioeconomic concept.
- overall, we think the precarity metric and plots should remain (fig 2 and 3). They would look the same style & type-wise but some details in the shocks, classifier, etc. might change.

*Meeting discussion points*
- The context is not clear. We do not have a good sense other than bigger is bad. How do we give the readers a sense as to the difference between difference gaps
- Calculate precarity values for the whole population and observe the distribution shifts
- Claim on latent precarity: If you do not observe latent precarity you cannot do well
- We can talk about latent factors such as having to help family members and such that (refer to Google docs).

**How do we evolve the system without stacking the decisions against individuals? How do we capture the instability? What is the difference between a permanenet worker and a gig worker?**
- use real word examples and statistics as discussed before. use real-world statistical plots.
- We would need to use the shock profiles to capture the instability in the system. One other knob we can use is the income distribution and how it changes over time for different individuals.
- For instance, a permanent worker might have a income distribution that is stable while the income distribution of the gig worker changes over time. (They start off with the same income distribution and the gig worker's have more instability over time)
- To evolve the system, we think we should give two individuals same shocks and then have one premanent large shock that would change things between the two individuals. The hypothesis is even if we simulate in favor of more precarious people (since in reality they get more shocks cite, cite cite), one large shock could have lasting effects that might not be immediately visible.

*Meeting discussion points*
- Try to get to: The latent property not observed by the individuals changes peoples' ability to react to shock.
- Both get the same negative shocks but some other properties are different.
- Try to see how the utility maximization strategy changes.
- When the minimum subsistence changes how the people with the same setting changes over time.
- We might need a new debt term or a way for their assets to go faster.
- Figuring out real world narratives that makes sense and rationalize the ways the shocks change on different narratives.

**How do we do the interventions?**

- We do not think we can do optimal interventions. Instead we would like to try to do interventions that has some temporal relationship to the shocks. Ex: interventions that are close to the shock event and maybe spread over time until there is stability.
- we could do different types of interventions as well. https://www.investopedia.com/terms/s/stimulus-package.asp

*Meeting discussion points*
- Just show this as a example thing we could do.

**What do we do for our classifier?**

- We do think we would need a more sophisticated classifier than the fixed average-based decision boundaries we have right now.
- idea: the evolution based dataset (@Kanchana)
- Data driven approach: We think we should find a dataset that contains information such as income, assets and link that to a measure such as loan default likelihood, credit risk, etc. and use them to come up with a classification model using python's built-in methods. There are couple of issues with this though. There are no public data out there that have both income and assets and some general finanacial label like credit (already checked this). Also, the clf for a specific golden label would mean that the decision would be just for that one specific purpose. Like loan approval: how would this even reflect as a bad shock event if we get denied? We have a system that is about all kinds of shocks. How would we cover all other types of negative shocks then, etc. There's also the issue of introducing bias by labels as well.

*Meeting Discussion Points*

- We need to figure out how to synthesize the classifier rather then relying on data. It could be kind of annecdotal but as long as it logically makes sense it should be okay.

**On the overall validation**
- use real world statistics like what we did in AIES. We set up the model with realitic human analoguos behaviors, feed it real world data, use reasonable classifiers and real world shock profiles. To make the claim stronger we can compare our results with real world statitics as well.