It is surprising indeed to feel like we are the in the 90s or earlier as I assumed back then the constraints regarding many of the most exciting problems would have probably been not enough computational resources and that the current physical models are oversimplified, but indeed it does feel like we are still in the comparative dark ages in this field. However, like you, I am hopeful for progress given the instruments now being deployed to measure the earth's movement.

So for example, if we see the coronavirus is infecting people in the United States, the known known is that it affects older people more and people with certain preexisting conditions. We also might have some sparse data and some rudimentary understanding of the mechanisms that lead to the spread of the virus and the way it leads to hospitalizations and deaths. From this, we might calculate a risk (or the odds) of death and hospitalizations. However, this risk may have a large uncertainty because we know that we do not know a whole number of things like does it spread through inanimate objects or how capable is the government in withstanding public anger when initiating lockdowns or how many people will actually wear masks. This might give us an initial range of uncertainty but then we also have the unknown unknown items such as we have no idea if another natural disaster occurs or if another country invades another country or how and when will the virus mutate. These are all unknown unknowns that further exacerbate the uncertainty. If there is good reason to worry that tail-risks can have dramatic impact on the outcome of the pandemic on # of deaths, then we might model with fat tails to represent the unknown unknown and we might keep a large range to model the high uncertainty reflected in the known unknowns.

Hi Zev, I am not so sure I understand what you mean by "you have understood machine learning." I assume you are referring to the fact that Machine learning is an exception in that it is understandable but at the same time highly complex. I am not so sure I would agree with this because while yes machine learning is indeed highly complex, I am not so sure the common "black-box" methods are understandable. Why did my model assign 0.8 weightings to one variable over the other? I cannot answer such a question. In fact, I believe returning to Occam's Razor and testing the efficacy of several simple models through Machine Learning's "white-box" methodology might make it the exception that you speak of.

Hi Vincent, that is a fantastic argument and something I have been thinking about too. I think in the case of the stock market, it might be relevant to distinguish crashes occurring from external shocks (natural disasters (earthquake)/pandemics and/or supply shocks (Suez Canal/Oil Crisis of the 1970s)) and crashes occurring endogenously (08 financial crisis). One can argue that the external shocks are random events but perhaps in a globalized world with rising nationalism and climate change intensifying we have good reason to fatten the tails for such external shocks. However, for financial crisis occurring endogenously because certain companies have taken far too much leverage at inflated asset prices, that we are much better off building a model to identify the risk of a correction leading to a run on the short-term debt markets and thus a subsequent fire sale.

That is an interesting and valid comment. I too was curious at this, but I guess I was confused by how would one measure how uncertain is the "right" amount of uncertainty about a particular model especially ex-ante. These calculations likely occur ex-post that denigrates experts for being too in love with their models thus operating from a bias with downward pressure on the level of uncertainty. I would think, to be honest, that you want experts to be more confident about their model but ideally, that confidence should come because when they are making certain predictions they are also releasing a detailed report that transparently documents how all relevant sources of uncertainty have been accounted for when the final level of uncertainty (variance) for any prediction is articulated. If someone for example says, I believe next month the GDP of the country will be 4% more than it was during the same month last year, but states that with supreme confidence as if fact, then we should likely label such an expert as one that we might be more suspicious of, but if an expert claims the GDP range should be 3.8 to 4.2% with 95% confidence as the variation will occur from these three particular factors, then we can be much more positive regarding this expert.