Over the weekend, Andrew Cuomo for the first time detailed a tentative plan to slowly move towards reopening New York.
Construction and manufacturing businesses operating in areas that were not particularly hard-hit by the coronavirus may begin to reopen next month, he said Sunday.
It marked a potentially pivotal moment in the epidemic. New York is, effectively, the global epicenter of the virus. That Cuomo – whose approach to daily briefings has generally been applauded both for substance and tone – is prepared to move the state down the long road towards normalization, says a lot about the improvement in key COVID-19 metrics.
As some market participants are aware, JPMorgan’s Marko Kolanovic has tracked the evolution of the virus for the better part of two months. His models have, for the most part, been accurate in predicting inflection points as well as case and mortality rates.
I’ve covered those projections in real time. Late last week, in “The Financial Media Ignored Marko Kolanovic On New York’s Coronavirus Outbreak. It Was Society’s Loss“, I summarized the evolution of Kolanovic’s forecasts and lamented the comparatively sparse coverage those forecasts received in the financial media.
On Tuesday, Kolanovic is out with a new piece providing additional details and updates.
“As we are learning in this crisis, inaccurate scientific forecasts, politicization of research, and a sensationalist approach can have significant impacts”, he writes, noting that “based on what is believed to be the worst-case forecasts, the policy response can range from targeted to indiscriminate when applied to the lockdown of geographic regions, protection of certain demographic segments, or assessing the risks of a specific economic activity being shut down”.
In the second linked post above, I looked at Kolanovic’s forecasts against some of the projections which were prominent late last month, noting that a simple, side-by-side comparison suggested Marko’s were, in fact, more accurate. Here is a table from his latest note which summarizes his forecasts, and compares them to the prevailing and accepted models at the time.
Even if you wanted to quibble with what studies or forecasts should be included in the “prevailing opinion at the time” column, the more important point is that Kolanovic’s projections were accurate. And they come with “receipts”, so to speak. They were delivered in timestamped notes, as indicated in the “our prediction at the time” column.
Below are two key points from Tuesday’s note. They serve as quick reminders regarding how the projections were derived (these are verbatim from Marko):
- An important factor in our forecast was relying on real-time data from smart thermometers (i.e., big data). This helped us determine the beginning, growth rate, peak, and decline in regional infection rates. In an exponential process (pandemic), having real time data is critical. Modeling the early-stage growth based on hospital data (which lag by 7-10 days) and use of wrong mortality rates is the primary reason for misleading consensus forecasts.
- We estimated mortality by developing a theoretical framework, and then applied it to cross-sectional country data. Our model incorporated differences in how countries responded to the COVID-19 crisis, and adjusted for the impact of age distribution on mortality. Once we backed out mortality, we could estimate virus prevalence rates and level of herd immunity.
Again, this modeling process was summarized and discussed in these pages (see the archive here) as it evolved. As far as I’m aware, no other notable media outlet (mainstream or otherwise) chronicled these projections in real time. That’s unfortunate, given their accuracy as documented in the figure above.
After going back over the process for modeling the COVID-19 mortality rate (as summarized here), Kolanovic takes things further.
“These observations are important as they allow us to model theoretical curves that can be fit on observed data in a robust way”, he says, adding that “in particular, mortality can be modeled with the help of 2 parameters (rational function)”.
The figure below depicts mortality as a percentage of population for the same virus, and four different “countries”. Again, this is theoretical – as Marko writes, “countries” means “different efficiencies in epidemic response”. Our “countries” are here taken to have identical demographics in order to simplify things.
Again, there are two parameters, one for testing and virus control effectiveness, and the other for the mortality rate with 100% of people tested. Obviously, in countries where all is demographically equal, the true mortality rate for the exact same virus will generally be the same (that’s a general statement – this is a virus we’re talking about, so there will always be differences across locales). In this instance, it is the effectiveness and efficiency of the response that determines the shape of the curve.
Kolanovic also models mortality curves for different viruses in the same country. In other words, he takes a look at different mortality rates with the response parameter held constant. “The presence of different COVID-19 virus ‘strains’ would complicate the analysis and one would need to fit curves”, he says. This would be applicable to studying countries where demographic profiles vary, as in such cases, mortality rates will generally be different, effectively making the same virus look like different viruses.
The latter part of Marko’s Tuesday note is probably best presented without paraphrasing or attempts at editorializing, as it’s possible that something might be lost in translation. Given that, I’m going to employ a block quote here (I assume Kolanovic wouldn’t mind, given it’s done out of respect for preserving the integrity of the analysis). To wit, from Marko:
Analyzing mortality and the % of the population tested on a log scale, one can isolate mortality in a cross sectional regression. This is shown in the sequence of figures below. Figure 5 shows regression of % mortality vs % population tested on the left. The middle chart takes log of mortality, and the right chart takes log of both mortality and % of population tested. The log transformation of variables gives the mortality level as the regression intercept (Log(100%)=0), and the error is estimated from standard linear regression theory. Such estimation is relatively robust to outliers, changes to observed data in time, and addition of new data points. It also allows adjustments of individual data points (e.g., adjust for the sample age) and to quickly understand how each data point such as a small country, ship, etc., impacts the ‘true’ mortality rate estimate. It also allows us to run scenario analyses (e.g., how much will the mortality estimate change with each new fatality on a ship, or, for example, how many additional fatalities in a country can one have to still be consistent with previous mortality estimate, etc.).
Clearly, all of this matters a great deal for the economy and thereby for markets.
What I think it’s important to point out in this particular case is that this needn’t have been a “with the benefit of hindsight”-type of scenario for market participants. The projections summarized in the first table above could have been communicated to the investing public in a way that preserved the value of the original research for those with access to it, while still conveying the general thrust in summary fashion, in case doing so might have helped everyone make more informed decisions. After all, that is quite literally the job description of the mainstream financial media – to summarize complex information in a way that is informative to market participants.
Or at least that should be the job of the mainstream financial media. Unfortunately, that is often superseded by other considerations including politics and, especially, revenue.
To address some reader feedback I’ve received recently, none of the above is to suggest that public policy should be based on sell-side research notes. Nobody is arguing that. Rather, what I’m suggesting (and what Marko is trying to communicate) is simply that in acute situations like pandemics, anyone with a plausible claim on expertise in some area that might be of use (in this case mathematical modeling) should have their ideas amplified enough to ensure those ideas are considered in conjunction with all the other suggestions and models being circulated.
Kolanovic is diplomatic about it. “Access to data, analyses, and unhindered exchange of research ideas should be of high importance”, he reiterates. “It should go without saying, research should not be biased by political views”.